The Microstructure of Foreign Exchange Markets
A National Bureau of Economic Research Conference Report
BANCA D'ITALIA
The Microstructure of Foreign Exchange Markets
Edited by
Jeffrey A. Frankel, Giampaolo Galli, and Alberto Giovannini
The University of Chicago Press
Chicago and London
JEFFREY A. FRANKEL is a research associate of the National Bureau of Economic Research, where he is also director for International Finance and Macroeconomics. He is also professor of economics at the University of California, Berkeley, and senior fellow at the Institute for International Economics. GIAMPAOLO GALLI is chief economist of the Confederation of Italian Industry in Rome. When this book was written, he was head of the International Section of the Research Department of the Banca d'Italia. ALBERTO GIOVANNINI is a research associate of the National Bureau of Economic Research, a research fellow of the Centre for Economic Policy Research, and senior adviser of Long Term Capital Management, L. P.
The University of Chicago Press, Chicago 60637 The University of Chicago Press, Ltd., London © 1996 by the National Bureau of Economic Research All rights reserved. Published 1996 Printed in the United States of America 05 04 03 02 01 00 99 98 97 96 1 2 3 4 5 ISBN: 0-226-26000-3 (cloth)
Library of Congress Cataloging-in-Publication Data The Microstructure of foreign exchange markets / edited by Jeffrey A. Frankel, Giampaolo Galli, and Alberto Giovannini. p. cm.-(A National Bureau of Economic Research conference report) Papers from a conference sponsored by the Bank of Italy, the National Bureau of Economic Research of Cambridge, USA, and the Centre for Economic Policy Research of London, UK, and held at S.A.DI.BA., the Banca d'ltalia's conference center in Perugia, Italy, on July 1-2, 1994. Includes bibliographical references and index. 1. Foreign exchange-Congresses. I. Frankel, Jeffrey A. II. Galli, G. III. Giovannini, Alberto. IV. Series: Conference report (National Bureau of Economic Research) HG205.M53 1996 95-43757 332.4' 5-dc20 CIP
@ The paper used in this publication meets the minimum requirements of
the American National Standard for Information Sciences-Permanence of Paper for Printed Library Materials, ANSI Z39.48-1984.
National Bureau of Economic Research Officers Paul W. McCracken, chairman John H. Biggs, vice-chairman Martin Feldstein, president and chief executive officer
Geoffrey Carliner, executive director Gerald A. Polansky, treasurer Sam Parker, director offinance and administration
Directors at Large Peter C. Aldrich Elizabeth E. Bailey John H. Biggs Andrew Brimmer Carl F. Christ Don R. Conlan Kathleen B. Cooper Jean A. Crockett
George C. Eads Martin Feldstein George Hatsopoulos Karen N. Horn Lawrence R. Klein Leo Melamed Merton H. Miller Michael H. Moskow
Robert T. Parry Peter G. Peterson Richard N. Rosett Bert Seidman Kathleen P. Utgoff Donald S. Wasserman Marina v. N. Whitman John O. Wilson
Directors by University Appointment Jagdish Bhagwati, Columbia William C. Brainard, Yale Glen G. Cain, Wisconsin Franklin Fisher, Massachusetts Institute of Technology Saul H. Hymans, Michigan Marjorie B. McElroy, Duke Joel Mokyr, Northwestern
James L. Pierce, California, Berkeley Andrew Postlewaite, Pennsylvania Nathan Rosenberg, Stanford Harold T. Shapiro, Princeton Craig Swan, Minnesota Michael Yoshino, Harvard Arnold Zellner, Chicago
Directors by Appointment of Other Organizations Marcel Boyer, Canadian Economics Association Mark Drabenstott, American Agricultural Economics Association Richard A. Easterlin, Economic History Association Gail D. Fosler, The Conference Board A. Ronald Gallant, American Statistical Association Robert S. Hamada, American Finance Association
Charles Lave, American Economic Association Rudolph A. Oswald, American Federation of Labor and Congress of Industrial Organizations Gerald A. Polansky, American Institute of Certified Public Accountants James F. Smith, National Association of Business Economists Josh S. Weston, Committee for Economic Development
Directors Emeriti Moses Abramovitz George T. Conklin, Jr. Thomas D. Flynn
Franklin A. Lindsay Paul W. McCracken Geoffrey H. Moore James J. O'Leary
George B. Roberts Eli Shapiro William S. Vickrey
Since this volume is a record of conference proceedings, it has been exempted from the rules governing critical review of manuscripts by the Board of Directors of the National Bureau (resolution adopted 8 June 1948, as revised 21 November 1949 and 20 April 1968).
Contents
Preface
Introduction
ix 1
Jeffrey A. Frankel, Giampaolo Galli, and Alberto Giovannini
I.
TRADING VOLUME, ASYMMETRIC INFORMATION, THE BID, AND THE ASK
1. Risk and Thrnover in the Foreign Exchange Market
19
Philippe Jorion Comment: Bernard Dumas
2. Bid-Ask Spreads in Foreign Exchange Markets: Implications for Models of Asymmetric Information
41
David A. Hsieh and Allan W. Kleidon Comment: Zhaohui Chen Comment: Antti Suvanto
3. Interdealer Trade and Information Flows in a Decentralized Foreign Exchange Market
73
William Perraudin and Paolo Vitale Comment: Silverio Foresi Comment: Alan Kirman
4. One Day in June 1993: A Study of the Working of the Reuters 2000-2 Electronic Foreign Exchange Trading System Charles Goodhart, Takatoshi Ito, and Richard Payne Comment: Richard K. Lyons vii
107
viii
Contents
5. Foreign Exchange Volume: Sound and Fury Signifying Nothing? Richard K. Lyons Comment: Mark D. Flood Comment: Antonio Mello
II.
183
SPECULATION, EXCHANGE RATE CRISES, AND MACROECONOMIC FUNDAMENTALS
6. Dynamic Hedging and the Interest Rate Defense 209 Peter M. Garber and Michael G. Spencer Comment: Richard K. Lyons 7. Heterogeneous Behavior in Exchange Rate Crises Fabio C. Bagliano, Andrea Beltratti, and Giuseppe Bertola Comment: Lorenzo Bini-Smaghi Comment: Richard K. Lyons 8. Exchange Rate Economics: What's Wrong with the Conventional Macro Approach? Robert P. Flood and Mark P. Taylor Comment: Andrew K. Rose Comment: Lars E. O. Svensson 9. Is There a Safe Passage to EMU? Evidence on Capital Controls and·a Proposal Barry Eichengreen, Andrew K. Rose, and Charles Wyplosz Comment: Jose Viiials
229
261
303
Contributors
333
Author Index
337
Subject Index
341
Preface
The project that produced this volume was originally inspired by recent turbulence in the foreign exchange markets, including large movements in the value of the dollar and, especially, the collapse of the European exchange rate mechanism in the crises of 1992-93. Standard macroeconomic models seemed able at best to explain only some of these major movements and able to explain even fewer of the many lesser short-term movements in exchange rates. An alternative approach, based on the microstructure of the foreign exchange market, seemed worth exploring. The project had three sponsors-the Bank of Italy, the National Bureau of Economic Research, and the Centre for Economic Policy Research-corresponding to the three coeditors of this volume. The conference itself was held at S.A.DI.BA., the Bank of Italy's conference center in Perugia, on 1-2 July 1994. It managed to draw together, either as paper authors or as discussants, many of the economists throughout Europe and the United States who have worked on microstructure-relevant aspects of foreign exchange markets. The conference concluded with a panel discussion chaired by Antonio Fazio, Governor of the Bank of Italy, and featuring Andrew Crockett of the Bank for International Settlements, David Mulford of Credit Suisse First Boston, Ian Plenderleith of the Bank of England, and Fabrizio Saccomanni of the Bank of Italy. From the beginning, the organizers were determined that the volume not be just another collection of the usual macroeconomic sort of papers. A focus on such features as the heterogeneity of participants in the foreign exchange market, trading volume, bid-ask spreads, and intradaily movements-all factors that are usually neglected in the standard approach-was considered essential. This is near-virgin territory for academic research, which alone makes it worth exploring. . That the microstructure approach is already well advanced in the case of equity markets shows that the task can be done. It calls for something more ix
x
Preface
than applying existing models to a new market, however. The structure of the foreign exchange market is inherently different from that of equity or commodity markets. It lacks their centralization and transparency, as many of the contributions here explain. We must acknowledge from the outset of the book that the microstructure approach is a long way from being able to explain major movements in exchange rates. For the time being, we can hope only to have made progress in understanding the microstructure of the foreign exchange market for its own sake. Further aspirations, that the knowledge gained might provide building blocks for models of exchange rate determination, pertain to the future. The editors would like to thank Antonio Fazio, Martin Feldstein, and Richard Portes, of the three sponsoring institutions, for their support, and in particular to thank the Bank of Italy for its hospitality.
Introduction Jeffrey A. Frankel, Giampaolo Galli, and Alberto Giovannini
Why the Need for Microstructure? Exchange rate economics has made progress since the start of the generalized floating regime in 1973. Much research has appeared, explaining important phenomena in the behavior of exchange rates, and using an ever larger set of data to test such explanations. Research in recent years has even refined our theoretical understanding, as well as the empirical analysis, of the dynamics of exchange rates within bands and the determinants of realignments. The key theoretical insight of all recent research on exchange rates is the socalled asset market approach. The foreign exchange market is seen no longer as a market where the flow supply and demand over time determine the equilibrium price-as in the case of perishable agricultural goods-but as a market where price is determined by expectations of income that can be generated by holding assets denominated in a certain foreign currency. The asset market approach was prompted by the observation that much of the fluctuation in foreign exchange rates is difficult to reconcile with the net flows of goods and capital that occur between countries. These fluctuations appear to be associated with political and economic news that, rather than affecting current flows, signals possible future changes in the value of the currency.
Jeffrey A. Frankel is a research associate of the National Bureau of Economic Research, where he is also director for International Finance and Macroeconomics. He is also professor of economics at the University of California, Berkeley, and senior fellow at the Institute for International Economics. Giampaolo Galli is chief economist of the Confederation of Italian Industry in Rome. When this book was written, he was head of the International Section of the Research Department of the Banca d'Italia. Alberto Giovannini is a research associate of the National Bureau of Economic Research, a research fellow of the Centre for Economic Policy Research, and senior adviser of Long Term Capital Management, L. P.
1
2
Jeffrey A. Frankel, Giampaolo Galli, and Alberto Giovannini
The asset market approach to exchange rates has produced a number of models that have proved useful in explaining and quantifying exchange rate movements. The first characteristic of these models is that they are macro models; that is, they are highly aggregated. They attempt to capture and make explicit all determinants of the demand for and supply of foreign exchange, including those that are quite outside the foreign exchange market per se. Given this general macro approach, the focus of the models is on financial asset markets' and therefore the emphasis is on the behavior of agents in these asset markets. There is a tension between the all-encompassing macro approach and the need to highlight the dynamics of asset markets. How is this tension resolved? It is resolved with the adoption of simplifying assumptions on asset markets, not uncommon in the finance literature, that it is useful to spell out. Agents are identical. Information is perfect. Trading is costless. Although models that relax them exist, these assumptions can be thought of as the hallmark of the asset approach to macroeconomic models of exchange rates. 1 The most important implication of these assumptions is the absence of an explanation for trading in assets. Thus, asset prices adjust every period (or every instant) to make agents content with the specified amount of assets in their portfolios. The adjustment of asset prices instantaneously reflects the arrival of new information in the marketplace, which all participants observe and interpret in the same way. Hence, the basic macro model of the exchange rate implies that all information pertaining to the current and future "fundamental" determinants of exchange rates, that is, all information that implies a current and/or future change in the return on assets denominated in different currencies, has an immediate and unambiguous effect on exchange rates. Why study foreign exchange market microstructure? The interest in the working of the foreign exchange markets stems, at least in part, from some of the problems that the asset market macro models have displayed. The first is a prima facie contradiction between the models and reality. As noted, such models imply the absence of trading in assets. By contrast, one of the most important empirical facts about the foreign exchange market is the high volume of transactions that occur daily. 2 This inconsistency raises the question of whether the failure of the standard models to account for the volume of foreign 1. Of course, there are models that relax some of these assumptions. In particular, the perfect information assumption has been modified by authors who have studied the implications of imperfect information on the dynamics of monetary policy disturbances and authors who have studied the implications of asymmetric information or differences in prior beliefs. 2. The most recent triennial surveys of the foreign exchange markets were conducted in April 1995. The Bank of England (1995) announced a 60 percent increase in trading volume over the preceding three years, to the level of $464 billion per day, in London; the Federal Reserve (1995) announced a 46 percent increase in volume in New York over the preceding three years, to the level of $244 billion a day; and the Bank of Japan announced a 34.3 percent increase, to the level of $161.4 billion, in Tokyo. The rate of increase in the major centers suggests that the current worldwide total is now in the vicinity of $1,300 billion a day.
3
Introduction
exchange transactions is a symptom of more serious· problems, which might cause the lack of success at explaining other empirical phenomena on which researchers have concentrated. These empirical phenomena include the behavior of excess returns in the foreign exchange market, the near total inability to predict exchange rates at short horizons, the inability to explain exchange rate movements even ex post, and the volatility of exchange rates. Standard models have been unable to explain these phenomena satisfactorily. In particular, asset pricing formulas implicit in the standard macro models seem, to date, to have fared poorly. For example, even though the existence of ex ante (i.e., forecastable) returns in the foreign exchange markets can in theory be explained as risk premia, the estimated returns in practice do not match what is predicted by asset pricing models based on the covariances among asset returns. Furthermore, models seem to have a difficult time predicting future movements in exchange rates, suggesting that the information contained in the macro variables that are usually included in these models is of limited value. Finally, the volatility of these macro variables is generally smaller than the observed volatility of exchange rates, suggesting that-unless certain variables have especially strong effects on the spot exchange rate, as, for example, in the case of large overshooting in reaction to monetary disturbances-the information affecting exchange rate movements may be in part extraneous to the variables belonging to standard macroeconomic models. Theories of rational speculative bubbles and speculative attacks can in one sense account for the existence of excess volatility. But they are inherently unsatisfying in that they have nothing to say about how or when such bubbles and attacks get started or how they end. It is only natural to ask whether these empirical problems of the standard exchange rate models-problems that stem from the assumptions on asset market equilibrium-might be solved if the structure of foreign exchange markets was to be specified in a more realistic fashion. This suggests a sort of micro-foundations approach to the foreign exchange markets, according to which a more satisfactory description of the foreign exchange market microstructure might help sort out some of the problems displayed by existing macro models. Thus, we have established one reason to study foreign exchange market microstructure: the dissatisfaction with the empirical performance of standard models. While we find this motivation perfectly justified, and while a research project that attempts to fix the empirical problems mentioned above by adopting a richer description of the foreign exchange markets is interesting, we also believe that it. is uncertain of success. The excess volatility problem and the difficulty of explaining ex ante rates of return also characterize other financial markets, such as, for example, the stock market. These markets, however, are organized in a way that differs significantly from the foreign exchange market and-unlike the foreign exchange market-are also subject to sub-
4
Jeffrey A. Frankel, Giampaolo Galli, and Alberto Giovannini
stantial regulation. The fact that similar empirical problems surface in markets having microstructures that differ in some ways makes it less likely that microstructure-based models will help explain away these empirical problems. On the other hand, the markets also have some similarities, which might hold the key to the problem. A second reason to study market microstructure is only loosely related to the first. Like any market, the foreign exchange market is an interesting subject for research that attempts to identify the economic effects of its organization. This is, as opposed to the macroeconomic approach to foreign exchange microstructure, the microeconomic approach. The questions that are addressed with this approach include, for example, transparency, decentralization, the use of brokers (vs. marketmakers, vs. auctioneers), the location of trading, the efficiency of clearing of foreign exchange transactions, the relation between spot and derivative markets, and the importance of systemic risk in the market. Both the ambitious macro approach to the foreign exchange market and the less ambitious but equally interesting micro approach represent good reasons to take stock of the state of research on the microstructure of the foreign exchange market. We believe that this field of research is beginning to grow tremendously, and we hope that a systematic look at its progress represents a useful addition to the literature on the foreign exchange market.
The Chapters in Part I The study of market microstructure has already produced at least one empirical regularity: the high intraday correlation of trading volume and volatility. As noted above, standard macroeconomic exchange rate models have little hope of explaining trading volume. Typically, they assume homogeneity of market participants. If all traders are the same, why should they trade? Of course, the standard models do not attempt to explain volume, considering it of little relevance except to those who make their living trading. But the observed correlation between volume and volatility suggests something of more general interest. Frankel and Froot (1990b), for example, find a high contemporaneous correlation between volume and volatility. They also find some evidence that dispersion of traders' forecasts, as reflected in survey data, Granger-causes both volume and volatility. Given that trading volume seems to be relevant, there are two possible broad interpretations. One is that the market is processing information in an efficient way. Here, efficient is not to be understood as in the narrowest definition of the efficient markets hypothesis, where all traders are homogeneous, all information is instantly and fully reflected in the market price, and there are no profits to be made by trading. Rather, the microstructure perspective presupposes heterogeneity' is often based (more specifically) on asymmetric information, and allows that relatively more skillful or informed traders may succeed at the expense of those who are less skillful or less informed or of customers who must transact because they need to eliminate exposure ("liquidity traders"). The first
5
Introduction
interpretation is simply that the market works to aggregate the individual bits of information available to each trader in a relatively rapid and smooth way. The chapters here shed light on a number of leading models of asymmetric information and the need for liquidity. What constitutes information in the foreign exchange market is less obvious than it is in the equity market, where, for example, individual analysts have information on individual corporations. But, in much of the work in this book, orders from customers (especially nonfinancial corporations) constitute the bits of information to which some traders have access and others do not. Lyons (1995) follows the behavior of a particular marketmaker over the course of five days. 3 Lyons sheds some light on how bits of such information are processed, in the form of a statistically significant effect of orders received by traders on the prices at which they transact. (Earlier high-frequency data, e.g., Goodhart's thirteen weeks of "indicative quotes" obtained from the Reuters screen, did not include actual order flow or transaction prices [see Goodhart and Figliuoli 1991].) The alternative interpretation, which often goes by the name noise trading, is that trading volume can itself generate "excessive volatility" (see, e.g., Tobin 1978; Goodhart 1988; De Long et al. 1990; and Frankel and Froot 1990a). In a well-known study of the stock market, French and Roll (1986) found that, when the market closed for election days or special shutdowns, volatility was not "stored up" to await the reopen of the market, even though the generation of new information continued. Rather, it seemed, volatility depended directly on trading volume, holding information flow constant. Global foreign exchange markets are open twenty-four hours a day, with the result that a repeat of the French-Roll experiment is not easy. One attempt is Ito and Roley (1990). Some of the work in this volume may shed light on the French-Roll hypothesis that trading volume can gratuitously generate volatility. Writing quite soon after the beginning of the floating-rate era, McKinnon (1976) claimed that exchange rates were excessively variable owing to a deficiency of stabilizing speculation. One would think that banks would be the best candidates to take open positions in a currency considered undervalued relative to fundamentals and to hold such positions for however many months it took to correct the "misalignment." Banks are reluctant, however, to take large positions overnight. Others confirm that, even though typical traders incur high exposure during the day, they try hard to close out their positions overnight (Fieleke 1981). Lyons (1995) finds evidence of an inventory effect during the course of the day in addition to the information effect: when the outstanding position is large, traders modify their bid and ask prices so as to discourage further exposure in that direction. In chapter 1 of this volume, Philippe lorion seeks to test one important mi3. More specifically, the data set consists of time-stamped quotes and trades, the marketmaker's indirect (brokered) trades, and the time-stamped prices and quantities for transactions mediated by a broker.
6
Jeffrey A. Frankel, Giampaolo Galli, and Alberto Giovannini
crostructure theory, that of Tauchen and Pitts (1983). The theory says that the correlation between trading volume and volatility should be positive when the source of trading volume is disagreement (heterogeneity of beliefs) and negative when volume is determined by the number of traders, owing to averaging over larger numbers (liquidity). He finds support for the theory in that the varianceis observed to depend positively on volume and negatively on a time trend intended to reflect the steadily growing number of traders. He uses options prices to obtain a measure of the anticipated component of the variance, which has not previously been studied in this context. lorion also looks at the bid-ask spread, the standard measure of transactions costs. He confirms earlier findings (Glassman 1987; Bessembinder 1994) that the spread widens before weekends and holidays, supporting the liquidity effect. He also confirms earlier findings that the bid-ask spread depends positively on the variance but negatively on volume. (He, like Wei [1994], uses the option-implied volatility for this purpose. Glassman [1987], Boothe [1988], Bollerslev and Domowitz [1993], and Bollerslev and Melvin [1994] used GARCH models of the variance rather than the option-implied volatilities.) The presumption here is that information is processed efficiently. At a time when beliefs are particularly heterogenous and therefore trading volume is particularly high, the presumption is that the market is responding to a rapid generation of information. Chapter 2, by Hsieh and Kleidon, casts some doubt on the proposition that information is processed efficiently. The point of departure is a model by Admati and Pfleiderer. It features a crucial distinction between well-informed traders and liquidity traders, some of whom have some discretion as to when they trade and so seek to trade at a time when high volume drives down the cost of transaction (the liquidity effect on the bid-ask spread). Hsieh and Kleidon confirm the correlation of volume and volatility that the Admati-Pfleiderer model is designed to explain: there is a bunching of volume and volatility at both the open and the close in the foreign exchange market. A deeper look, however, uncovers serious problems. First, the bid-ask spread is observed to go up, not down, at the open and the close, contradicting the notion that liquidity traders are deliberately bunching at these times to save on transactions costs. Second, at the close in London, when volume and volatility are high in that market, there is no detectable simultaneous effect in the open New York market. This seems to contradict existing models of asymmetric information, which presuppose a common knowledge of economic structure despite the existence of idiosyncratic information. If volatility is high in London because information relevant to the pound/dollar rate is coming out, then why shouldn't the same effect show up in the pound/dollar rate in the New York market? Hsieh and Kleidon think that the answer lies in models where information is aggregated imperfectly and inventories are important. They take at their word traders who explain that, at morning open, they need to get a "feel" for the market, thus explaining the combination of high trading volume, high vola-
7
Introduction
tility, and high spreads in the morning. Toward evening close, traders are anxious to unload excess inventories, explaining the reappearance of the heightened volume, volatility, and spreads. In chapter 3, Perraudin and Vitale build a theory that can explain why such a high percentage of trading volume takes place among dealers instead of with customers. They emphasize that, because the foreign exchange market is decentralized, order flow cannot be observed by everyone. This setup differs from most of the conventional microstructure literature, which was designed for transparent centralized markets such as the equity markets. Some foreign exchange traders acquire useful new information at intervals, in the form of orders from customers. Perraudin and Vitale model the trading process by means of which this information gets disseminated to the marketplace by dividing the interval into four stages. Those traders without information seek by their trading strategy at an early stage to tease out the information from those who have it. Informed traders in effect sell or rent private information to others. Whereas the prominent Glosten-Milgrom theory of microstructure says that traders should quote a wide spread when dealing with someone whom they believe to be better informed so as not to be "taken in," the Perraudin-Vitale theory explains why traders would want to quote a narrow spread in such a case. They want to get the order so that they can derive information that will be useful at the next stage of trading. One needs to test such theories of asymmetric information. The Jorion and Hsieh-Kleidon chapters gave us somewhat conflicting verdicts: the first more supportive of some theories of optimization subject to asymmetric information and the second less supportive. In chapter 4, Goodhart, Ito, and Payne suggest that a good deal of skepticism is warranted regarding such tests because they are based on "indicative quotes," that is, the bid and ask quotes that are posted to all potential customers. Traders usually quote better prices to each other when they transact. Goodhart and his coauthors use data from a new trading system, the Reuters 2000-2, to compare actual transactions prices with the indicative quotes ("FXFX"). They find that movements in the two are very close, with the result that for some purposes either series can be used. But the behavior of the high-low spread in the Reuters 2000-2 data is quite different from behavior of the FXFX bid and ask quotes. In other words, one should not mistake the publicly posted bid and ask prices for the prices at which foreign exchange traders trade with each other. Fortunately, Richard Lyons's data set constitutes direct observation of trader behavior. It contains the time-stamped transaction prices and quantities at which his trader traded in the New York mark/dollar market. In chapter 5, Lyons confirms the existence of both an effect of the dealer's inventory on transactions prices and an effect of a dummy variable indicating whether the trade was initiated by the buyer or by the seller. These significant effects confirm, respectively, an asymmetric-information channel and an inventorycontrol channel. The main contribution of the Lyons chapter, however, is to test an additional
8
Jeffrey A. Frankel, Giampaolo Galli, and Alberto Giovannini
effect on the transactions price: the effect of the quantity traded. Lyons seeks to shed light on two competing theories of why trading volume is, sometimes very high. What he calls the "event uncertainty" view is that high trading volume indicates that information is being processed rapidly. What he calls the "hot potato" theory is that high trading volume indicates that little information is being processed; rather, liquidity-trader customers are placing orders with their brokers, who then unload their overextended positions on other traders, who continue to pass the exposure like a hot potato. (For this second theory, he cites the Admati-Pfleiderer model of discretionary liquidity traders, tested by Hsieh and Kleidon in chapter 2.) The evidence supports the hot potato view: the quantity traded has a significant effect only when the time between transactions is long. When the time between transactions is short, the quantity traded has no significant effect on the trader's prices, suggesting that the trader views these orders as coming from liquidity traders rather than informed traders.
The Chapters in Part II Part II provides clues as to the potential relevance of microstructure for understanding exchange rate dynamics, especially in times of crisis. A widely reported phenomenon during the European currency crises of 1992 was the use of dynamic hedging strategies that caused increased sales of the currencies under attack following defensive interest rate hikes by central banks (G-l 0 1993). Sterling and, perhaps more intensely, the Swedish krona and the lira were reported to be subject to this source of pressure. The sales were undertaken by the institutions that needed to hedge short positions in put options on these currencies. Standard assumptions of macroeconomic models (in particular, gross substitutability among assets) make it difficult to understand how interest rate hikes could intensify pressure on a currency. The debate regarding the European Monetary System (EMS) crises of 1992 and 1993 has pointed to the possibility that large interest rate hikes aggravate domestic economic conditions (especially where large debt overhangs are a problem), thus inducing markets to expect very low interest rates in the future and weakening the currency. While "wrong" own-price effects are not unknown to macroeconomists, it should be noted that the economics of dynamic hedging is not related to expectations about the future evolution of exchange rates or other asset prices: as is well known, arbitrage considerations alone (plus the Wiener process assumption on the underlying asset) are required for deriving the most commonly used formulas for option prices and delta hedging. The issue is instead closely related to so-called stop-loss and portfolio-insurance strategies, which have been extensively analyzed after the stock market crash of 1987 in the Brady Commission (1988) report and in subsequent professional literature. "Portfolio insurance" (the behavior that results from aggregating the customer, who holds a put option, and the bank, which hedges its short option position) is a positive
9
Introduction
feedback strategy (buy high and sell low): hence it implies unconventional own-price derivatives. The added twist in the case of delta hedging on currencies analyzed in chapter 6 by Garber and Spencer is the extension of unconventional signs to the interest rate effect. The questions that can be asked are very much the same as those that have been asked in the stock market literature: to what extent these strategies are still efficient when the standard distributional assumptions are violated or there is a likely discontinuity in the market; what utility function may justify a demand for insurance that appears to be insensitive to relative prices; and whether mechanical trades are likely to have a major effect on market prices. The first question seems to be particularly relevant in the case of pressures against a pegged exchange rate, when central banks' actions may significantly affect price dynamics. For instance, applying Krugman's (1991) smooth pasting principle to the behavior of the exchange rate within a band yields a formula that implies much smaller sales of the currency under attack than those implied by the standard Garman-Kohlhagen (1983) model. As to the second question, it should be recognized that the same bank whose option desk increases the short position in the weak currency to maintain a hedged option position may take the strategic view that the interest rate hike will reinforce the currency. It may therefore decide to reduce its aggregate short position in the weak currency rather than increasing it. The same may be done by the bank's customer who is holding a long option position. The point is that there may be a division of labor within large organizations but that the aggregate behavior of each institution may well conform to more standard economic assumptions. Some recent microstructure literature suggests reasons why mechanical selling pressure by hedgers may instead be self-reinforcing: essentially, other, poorly informed agents infer from the price drop caused by the hedgers that an unknown negative shock has occurred and join the selling. According to Gennotte and Leland (1990), the price effects of this mechanism might be quite dramatic, even if trades undertaken by hedgers are a small proportion of total turnover. To avoid this problem, Grossman (1988) proposed to enforce publicity of the trades undertaken by certain categories of professional agents so that other agents may be able to distinguish the actions of hedgers from those that may be generated by genuine new information about economic fundamentals. Considerable work is clearly required to assess the practical relevance of positive feedback trades in the foreign exchange markets, but data are virtually nonexistent. Central banks and international financial institutions could take the lead in this area. The analysis of dynamic hedging suggests the possibility that different agents respond differently to the same information. The issue of heterogeneity across classes of operators is taken up explicitly in chapter 7. Bagliano, Beltratti, and Bertola analyze Italian statistics from the balance of payments and
10
Jeffrey A. Frankel, Giampaolo Galli, and Alberto Giovannini
banks' balance sheets and conclude that there is evidence of heterogeneous behavior on the eve of the September 1992 crisis. A few main facts stand out. The net (spot and forward) foreign currency position of the domestic banking system displays very small changes around zero throughout the entire year 1992. Nonbank residents instead remained net exporters of capital until August 1992, a trend that had started with the liberalization of capital movements and probably reflected gradual portfolio diversification. The pace of these outflows did not increase very much in the months preceding the devaluation. On the other hand, nonresidents continued to be net investors in lira-denominated assets. Within the nonbank domestic sector, households seem to have taken positions against the lira, while the evidence on firms is mixed: some seem to have increased their foreign currency liabilities, especially with domestic banks, while others increased their foreign currency assets. As argued above, some sort of heterogeneity, such as dispersion of beliefs, is essential to understand why trading occurs in any asset market. The interesting suggestion of this chapter is that there are sources of heterogeneity other than heterogeneity of beliefs-linked to risk aversion, need for liquidity, and asset preferences-that may be important in determining the direction and intensity of trading by different agents. Formal or informal regulatory constraints, especially on banks, may also be an important factor. Relaxing the representative agent assumption considerably complicates the analysis, but it may in some cases usefully complement the standard tools of macroeconomic analysis. While the relevance of microstructure analysis for exchange rate dynamics is not yet established, it is clear that macroeconomic tools alone provide unsatisfactory explanations, except possibly over long time horizons. This point is stated with particular clarity by Flood and Taylor in chapter 8. The authors carefully review the very large body of evidence on major exchange rate models: purchasing power parity, monetary models, and sticky-price, equilibrium, and portfolio-balance models. Although at one time or another each of these models has found some support in the empIrical literature, the overall picture is that, for industrial countries during normal times (excluding periods of hyperinflation), conventional macro fundamental models are incapable of explaining the greater proportion of the variation in nominal exchange rates. In spite of the very large literature that followed the seminal paper by Meese and Rogoff (1983), beating the random walk still remains the standard metric by which to judge empirical exchange rate models. Improvements over the random walk have been slight and generally not statistically significant. This is true whether actual or predicted values of future explanatory variables are used in the forecasting equations of macro models. Even for the real exchange rate, it is difficult to beat a random walk model on conventional data sets, which is often taken as evidence against purchasing power parity. Permanent real demand and supply shocks may in principle account for the random walk property of real exchange rates; this is the key point emphasized in equilibrium models. There is, however, very little evidence that
11
Introduction
the large movements that have been observed in real exchange rates can be explained by real shocks. Moreover, according to equilibrium models, real exchange rates should be largely invariant to the choice of monetary regimes, a hypothesis that seems to be strongly contradicted by the evidence. In particular, variability of real exchange rates is much larger in floating exchange rate regimes without there generally being greater variability in macroeconomic fundamentals (Frankel and Rose 1995). Nevertheless, some recent evidence suggests that purchasing power parity has some explanatory power over very long time horizons or across very diverse countries. Flood and Taylor add their own piece of evidence to this proposition by pooling cross-sectional and averaged time-series data on twenty-two countries and twenty years. The interpretation of their results is that simple macro fundamentals have explanatory power with respect to nominal exchange rate movements over five years or more. They conclude that further attempts to provide explanations of short-term exchange rate movements based solely on macro fundamentals may not prove successful, although the macro fundamentals are important in setting the parameters within which the exchange rate moves in the short run. The last contribution of this book, chapter 9, takes it largely for granted that macroeconomic analysis is only part of the story: it assumes that there can be self-fulfilling speculative attacks against a pegged exchange rate or, at the very least, that crises can occur even among countries with an extremely high degree of convergence in economic policies and conditions. The belief of Eichengreen, Rose, and Wyplosz in this proposition is so strong as to lead them to state that a European monetary union will not be achieved unless the Maastricht Treaty is amended in one of the following ways. Either it is decided to eliminate the transitory stage (in which exchange rates are required by the treaty to remain within "normal" fluctuations margins), or capital controls are imposed to maintain the required stability in exchange rates during the transition. Their specific proposal is to introduce a non-interest-bearing deposit on bank lending to nonresidents. Bank lending is seen as the indispensable raw material for speculating against a currency. To support their proposal, the authors provide evidence on the effectiveness of capital controls. Interestingly, their statistical tests tend to show that, while capital controls allow countries some greater degree of economic policy autonomy, they do not significantly contribute to reducing the costs of resisting speculative attacks. Reserve loss and interest rate hikes by central banks do not appear to be smaller when exchange controls are in place. This proposition may help reconcile two widely held views: one claiming that the EMS survived in the 1980s because of capital controls (e.g., Rogoff 1985; Wyplosz 1986; and Giovannini 1989) and another arguing that capital controls are usually ineffective because they can be evaded (e.g., Gros and Thygesen 1992; and Dini 1994). The underlying microeconomics-the key concern for this book-are quite clear. Since the legal risks of evading the controls are essentially a fixed
12
Jeffrey A. Frankel, Giampaolo Galli, and Alberto Giovannini
cost, agents, may be willing to forgo the (moderate) benefits of portfolio diversification in normal times, but the incentive to evade becomes very high when a large devaluation is perceived as imminent.
An Agenda for Future Research Our brief survey, together with a more careful reading of the papers and comments in this collection, provides an argument as to the importance of research on the microstructure of foreign exchange markets. The increasing availability of tick-by-tick data in the spot foreign exchange markets and in the various derivative markets (forward, swaps, options, etc.) yields an inexhaustible source of evidence against which theories can be tested. The literature on the microstructure of foreign exchange markets is bound to grow significantly in the near future. We conclude this introductory chapter by discussing the directions in which this literature is likely to embark. This research is newborn. It has a long way to go before it can claim to produce a model of exchange rate determination. After realistic models of dealer behavior are constructed, the next step would have to be letting such dealers interact in the marketplace, in order to derive a central tendency to the torrent of bid and offer quotes and transaction prices in which each individual deals. That central tendency would be what in macroeconomic models we call the market-clearing exchange rate. Then the interaction among dealers would have to be embedded in the larger universe of borrowers, lenders, importers, and exporters who playa role in the foreign exchange market so that economic fundamentals could ultimately be brought back in. This is a very tall order. Even a superficial reading of the papers in this collection should convince one of the difficulty of constructing a general equilibrium model of the exchange rate. A successful combination of microstructure theory and macroeconomic theory appears to be out of reach at this stage. From one standpoint, however, this should not be a total loss. Information on macroeconomic aggregates like monetary aggregates, gross domestic product, the balance of payments, and. the components of aggregate demand arrive with a frequency that is much lower than the frequency of trading in foreign exchange markets and is affected by significant statistical errors. Information about most prices of manufactured goods and services is affected by even larger statistical errors. It is perhaps no surprise that these variables have little to do with foreign exchange markets and the short-run determination of exchange rates. The relation between these variables and exchange rates is also not very reliable in the long run. For these reasons, a theory that forces a tight link between macroeconomic prices and quantities is perhaps not especially needed. By contrast, there are a number of important questions, confined to the working of the foreign exchange market, that we find particularly intriguing. At a minimum, the study of microstructure should provide insight into the implications of the organization of the market-and in particular its decentral-
13
Introduction
ized nature-on the volume of trade, the determination of bid-ask spreads, and perhaps the volatility of prices. In addition, theoretical work could help understand the behavior of foreign exchange dealers and their interaction. These models can help assess the optimality of trading techniques like stop-loss orders (Krugman and Miller 1993), "chartism" or "technical analysis" (Allen and Taylor 1989; Frankel and Froot 1990a, 1990b; Schulmeister 1987,1988; Goodhart 1988; and Goodman 1979), and "support levels" (De Grauwe and Decupere 1992). They could ultimately describe short-run price dynamics and improve our understanding of concepts such as speculative bubbles and speculative attacks. In the international finance literature, these concepts are burdened with a considerable amount of implicit theorizing. Speculative bubbles are often taken to be the deviations of exchange rates from fundamental macro models; speculative attacks are viewed as accelerating movements in exchange rates in anticipation of large changes in underlying macroeconomic variables. Research on market microstructure can give new and more useful meaning to these ideas.
References Allen, Helen, and Mark Taylor. 1989. Chartists, noise and fundamentals: A study of the London foreign exchange market. Working Paper no. 341. London: Centre for Economic Policy Research. Bank of England. 1995. The foreign exchange market in London. London, 19 September. Press notice. Bank of Japan. 1995. Tokyo foreign exchange market turnover survey (April 1995). Tokyo, 19 September. Bessembinder, H. 1994. Bid-ask spreads in the interbank foreign exchange markets. Journal of Financial Economics 35:316-48. Bollerslev, Tim, and Ian Domowitz. 1993. Trading patterns and prices in the interbank foreign exchange market. Journal of Finance 48:1421-43. Bollerslev, Tim, and Michael Melvin. 1994. Bid-ask spreads and volatility in the foreign exchange market: An empirical analysis. Journal ofInternational Economics 36, nos. 3/4 (May): 355-72. Boothe, Paul. 1988. Exchange rate risk and the bid-ask spread: A seven-country comparison. Economic Inquiry 26:485-92. Brady Commission. 1988. Report of the Presidential Task Force on Market Mechanisms. Washington, D.C.: U.S. Government Printing Office, January. De Grauwe, Paul, and Danny Decupere. 1992. Psychological barriers in the foreign exchange market. Discussion Paper no. 621. London: Centre for Economic Policy Research, January. De Long, J. Bradford, Andrei Shleifer, Lawrence Summers, and Robert Waldmann. 1990. Noise trader risk in financial markets. Journal of Political Economy 98, no. 4:703-38. Dini, Lamberto. 1994. Turbulence in the foreign exchange markets: Old and new lessons. In Monetary stability through international cooperation, ed. A. Bakker, H. Boot, O. Sleijpen, and W. Vanthoor. Dordrecht: Kluwer Academic Publishers.
14
Jeffrey A. Frankel, Giampaolo Galli, and Alberto Giovannini
Federal Reserve Bank of New York. 1995. April 1995 central bank survey of foreign exchange market activity. New York, 19 September. Fieleke, Norman. 1981. Foreign-currency positioning by U.S. firms: Some new evidence. Review of Economics and Statistics 63, no. 1 (February): 35-43. Frankel, Jeffrey, and Kenneth A. Froot. 1990a. Chartists, fundamentalists, and the demand for dollars. In Private behavior and government policy in interdependent economies, ed. Anthony Courakis and Mark Taylor. Oxford: Clarendon. - - - . 1990b. Exchange rate forecasting techniques, survey data, and implications for the foreign exchange market. Working Paper no. 3470. Cambridge, Mass.: National Bureau of Economic Research. Frankel, Jeffrey, and Andrew Rose. 1995. Empirical research on nominal exchange rates. In Handbook of international economics, ed. Gene Grossman and Kenneth Rogoff. Amsterdam: North.. Holland. French, K., and R. Roll. 1986. Stock return variances: The arrival of information and the reaction of traders. Journal of Financial Economics 17:5-26. G-I0. 1993. International capital movements and foreign exchange markets: A report by the group of deputies. Washington, D.C. Garman, Marc, and Steven Kohlhagen. 1983. Foreign currency option values. Journal ofInternational Money and Finance 2 (December): 23-37. Gennotte, Gerard, and Hayne Leland. 1990. Market liquidity, hedging and crashes. American Economic Review 80 (December): 999-1021. Giovannini, Alberto. 1989. How do fixed exchange rate regimes work? Evidence from the gold standard, Bretton Woods and the EMS. In Blueprints for exchange rate management, ed. Marcus Miller, Barry Eichengreen, and Richard Portes. New York: Academic. Glassman, Debra. 1987. Exchange rate risk and transactions costs: Evidence from bidask spreads. Journal ofInternational Money and Finance 6:479-90. Goodhart, Charles A. E. 1988. The foreign exchange market: A random walk with a dragging anchor. Economica 55:437-60. Goodhart, Charles A. E., and L. Figliuoli. 1991. Every minute counts in financial markets. Journal ofInternational Money and Finance 10:23-52. Goodman, S. 1979. Foreign exchange forecasting techniques: Implications for business and policy. Journal of Finance 34:415-27. Gros, Daniel, and Niels Thygesen. 1992. European monetary integration from the European Monetary System to the European Monetary Union. London: Macmillan. Grossman, Sanford J. 1988. An analysis of the implications for stock and futures price volatility of program trading and dynamic hedging strategies. Journal ofBusiness 61 (July): 275-98. Ito, Takatoshi, and V. Vance Roley. 1990. Intraday yen/dollar exchange rate movements: News or noise? Journal of International Financial Markets, Institutions and Money, vol. 1, no. 1. Krugman, Paul R. 1991. Target zones and exchange rate dynamics. Quarterly Journal ofEconomics 106, no. 3 (August): 669-82. Krugman, Paul, and Marcus Miller. 1993. Why have a target zone? Carnegie-Rochester Conference Series on Public Policy 38:279-314. Lyons, Richard. 1995. Tests of microstructural hypotheses in the foreign exchange market. Journal of Financial Economics 39:321-51. McKinnon, Ronald. 1976. Floating exchange rates, 1973-74: The emperor's new clothes. Carnegie-Rochester Conference Series on Public Policy 3:79-114. Meese, Richard A., and Kenneth Rogoff. 1983. Empirical exchange rate models of the seventies: Do they fit out of sample? Journal of International Economics 14, no. 2 (February): 3-24.
15
Introduction
Rogoff, Kenneth. 1985. Can exchange rate predictability be achieved without monetary convergence? Evidence from the EMS. European Economic Review 28:93-115. Schulmeister, Stephen. 1987. An essay on exchange rate dynamics. Research Unit Labor Market and Employment Discussion Paper no. 87-8. Berlin: Wissenschaftzentrum Berlin fur Sozialforschung. - - - . 1988. Currency speculation and dollar fluctuations. Banca Nazionale del Lavoro Quarterly Review 167:343-65. Tauchen, G., and M. Pitts. 1983. The price variability-volume relationship in speculative markets. Econometrica 51:485-505. Tobin, James. 1978. A proposal for international monetary reform. Eastern Economic Journal 3 (July/October): 3-4. Wei, Shang-Jin. 1994. Anticipations of foreign exchange volatility and bid-ask spreads. Working Paper no. 4737. Cambridge, Mass.: National Bureau of Economic Research, May. Wyplosz, Charles. 1986. Capital controls and balance of payments crises. Journal of International Money and Finance 5: 167-79.
I
Trading Volume, Asymmetric
Information, the Bid, and the Ask
1
Risk and Turnover in the Foreign Exchange Market Philippe lorion
The foreign exchange market is the largest and fastest-growing financial market in the world. Yet the microstructure of the foreign exchange market is only now receiving serious attention. As described in table 1.1, daily turnover in the foreign exchange market was $880 billion as of April 1992. To put these numbers in perspective, consider the following data: as of 1992, daily U.S. GNP was $22 billion; daily worldwide exports amounted to $13 billion; the stock of central bank reserves totaled $1,035 billion, barely more than one day's worth of trading. The volume of trading can also be compared to that of the busiest stock exchange, the New York Stock Exchange (NYSE), about $5 billion daily,! or to that of the busiest bond market, the U.S. Treasury market, about $143 billion daily (Federal Reserve Monthly Review [April 1992]). Since the advent of flexible exchange rates in the early 1970s, the foreign exchange market has been growing at a record rate. Figure 1.1 compares the volume of world exports to the volume of trading in deutsche mark (DM) currency futures, both expressed on a daily basis. I use futures volume because futures markets provide the only reliable source of daily volume information even if they account for only a small fraction of the foreign exchange market. The figure shows that, since the early 1970s, trading in deutsche mark futures has increased much faster than the volume of world trade. This reflects the overall growth in the foreign exchange market, where turnover has increased from $110 billion in 1983 to $880 billion in 1992. Because transaction volume is many times greater than the volume of trade flows, it cannot be ascribed to the servicing of international trade. To illustrate Philippe lOTIon is professor of finance at the Graduate School of Management of the University of California, Irvine. Thanks are due to participants in the NBER conference for useful comments. Partial financial support was provided by the Institute for Quantitative Research in Finance. 1. Average volume is 250 million shares, with an average price per share of about $20.00.
19
20
Philippe Jorion
Table 1.1
Daily Thrnover in the Foreign Exchange Market (billions of dollars) April 83
Market London (8:00 A.M.-16:00 P.M., GMT) New York (14:00 p.M.-22:00 P.M., GMT) Tokyo (23:00 p.M.-7:00 A.M., GMT) Singapore Zurich Hong Kong Germany Paris Canada Total a
34
110
April 86
April 89
April 92
90 59 48
187 129 115 55 57 49
9
26 15
300 192 126 76 68 61 57 36 22
206
640
880
aVolume for all countries may not add up to total owing to omissions, gaps in reporting, and double counting. GMT = Greenwich Mean Time.
Exports
Futures
3
10
World Exports 2
5
OM Futures o~-----_-.--....------
73 74 75 76
Fig.1.1
n
o
78 79 80 81 82 83 84 85 86 87 88 89 90 91 92
Comparison of daily volume-billions of dollars
this point, table 1.2 describes the changing patterns of activity in the New York foreign exchange market. Over time, activity in the Canadian dollar has dwindled to about 5 percent of the market; given that Canada is the largest trading partner of the United States, trade cannot be the prime determinant of turnover in a currency. It is also interesting to note that the share of the Dutch gulden has fallen sharply after 1980; this is due to the pegging of the gulden to the mark, which, after March 1979, allowed traders to cross-hedge efficiently and more cheaply with the mark. These two examples suggest that volatility and turnover are correlated: low turnover is associated with the low volatility of the Canadian dollar or of a cross-rate.
21
Risk and Turnover in the Foreign Exchange Market
Table 1.2
Currency German mark Japanese yen British pound Swiss franc Canadian dollar French franc Dutch gulden Belgian franc Italian lira Other Total (%) Total ($billion)
Breakdown of Foreign Exchange Market Thrnover by Currency (percentage terms, New York market) 1969
1977
1980
1983
1986
1989
17.0 2.0 45.0 7.0 21.0
27.3 4.3 17.0 13.8 19.2 6.3 5.7 1.5 1.1 2.8
31.7 10.2 22.8 10.1 12.3 6.8 1.9 1.0 .9 2.2
32.5 22.0 16.6 12.2 7.5 4.4 1.6 .4 .8 2.1
34.2 23.0 18.6 9.7 5.2 3.6 1.4
33.0 25.0 15.0 12.0
4.4
15.0
100
100 5
100 23
100 34
100 58
100 129
Previous academic literature has viewed the positive correlation between volume and volatility as reflecting joint dependence on a common directing variable or event. This common "mixing" variable represents the random number of daily equilibria, due to new information arriving to the market. According to this class of models, known as the mixture of distribution hypothesis (MDH), unexpected risk and unexpected volume are positively correlated through their dependence on an information-flow variable. In addition, Tauchen and Pitts (1983) show that expected turnover may change over time and increases with the number of active traders, with the rate of information flows, and with the amount of trader disagreement. This is consistent with the idea that, since trading reflects capital transactions, turnover must be driven by heterogeneous expectations combined with volatility. In previous work, the positive correlation between risk and turnover was derived from ex post measures. Given the substantial amount of time variation in risk and turnover, however, it is crucial to distinguish between expected and unexpected volatility. This paper measures expected volatility from options on deutsche mark currency futures traded on the Chicago Mercantile Exchange (CME) over the period 1985-92. For a given market price, inverting the appropriate pricing model yields an implied standard deviation (ISO). It is widely believed that ISOs are the market's best estimate of future volatility. After all, if it were not the case, one could devise a trading strategy that could generate profits by trading in mispriced options. This study also investigates bid-ask spreads in spot markets. The literature on spreads identifies inventory costs as one of the main components of spreads. Higher volatility means, ceteris paribus, that dealers face the risk that the exchange rate will move unfavorably while the position is held. Although this risk might be diversifiable in theory, in practice active currency dealers effec-
22
Philippe Jorion
tively focus on one currency only and therefore worry about idiosyncratic risk. As a result, when volatility increases, so should the spread, which reflects the compensation that dealers expect for taking on currency risk. Again, to test this hypothesis, it is crucial to distinguish between expected and unexpected volatility. ISDs should provide better volatility forecasts than time-series models. This paper is organized as follows. The literature on the turnover-risk relation, on the spread-risk relation, and on measuring risk from options is reviewed in section 1.1. Section 1.2 describes the data. The measurement of expectations for volume and risk from time-series data is presented in section 1.3. Section 1.4 discusses how implied volatilities are derived from 'option prices. Empirical results are presented in section 1.5. Finally, section 1.6 contains some concluding observations.
1.1 1.1.1
Literature Review Turnover and Risk
The domestic microstructure literature has long been concerned with the relation between turnover and risk. This relation is important for several reasons. First, it provides insight into the structure of financial markets by relating new information arrival to market prices. Also, it has implications for the design of new futures contracts; a positive relation suggests that a new futures contract can succeed only when there is "sufficient" price uncertainty with the underlying asset, which cannot be effectively cross-hedged with other contracts. Finally, the price-volume relation has a direct bearing on the empirical distribution of speculative prices. The mixture of distribution hypothesis (MDH), first advanced by Clark (1973), assumes that price variability and volume are both driven by an unobserved common directing variable. Indeed, numerous studies have reported a strong contemporaneous correlation between volume and volatility.2 Cornell (1981) provides considerable empirical evidence on how pervasive the relation is for eighteen futures contracts. Grammatikos and Saunders (1986) analyze foreign currency futures contracts and find that detrended volume is positively related to variability. At the same time, there are secular increases in volume, without corresponding increases in volatility. These observations have been brought together in a seminal paper by Tauchen and Pitts (1983). The authors present a model where the volatilityvolume relation can take two forms: (1) as the number of traders grows, market prices, which can be considered as an average of traders' reservation prices, become less volatile because averaging involves more observations; (2) with a fixed number of traders, higher trading volume reveals higher disagreement 2. Karpoff (1987) provides a survey of the evidence in the futures and equity markets.
23
Risk and Turnover in the Foreign Exchange Market
among traders and is thus associated with higher price variability. This link is stronger when new information 1flows to the market at a higher rate. Formally, market prices P and volume V are modeled as
(1)
IlP
= (Jl -W-Zl'
V=
J.L2 1 +
(J2
-W-Z
2,
where ZI and Z2 are independent N(O, 1) variables, and 1represents the random number of daily equilibria, due to new information arriving to the market. In the above, the variance term ai depends both on the variance of a "common" noise component a~, agreed on by all traders, and on the variance of the "disagreement" component, ~2 scaled by the number of active traders N:ai + ~P/N. Volatility of prices then increases with the rate of information flow I, increases with the common noise a o' increases with trader disagreement~, and decreases with the number of active traders N. As for the volume parameters, these can be written as J.L2 -;- ~N and a~ -;- \lPN. Turnover then increases with the rate of information flow I, with trader disagreement ~, and with the number of active traders N. Because both Ilp2 and V depend on the mixing variable I, their covariance is positive and equal to aiJ.L2 Var (I). At the transaction level, however, V and IlP are independent. These relations can be summarized as
Var(IlP) (2)
= (a~
+
~2/N) • E(l),
E(V) -;- ~N· E(/),
COV(llp2,
V) -;-
(a~
+ ~2/N)
~N· Var(l).
However appealing, this model has the severe limitation that the mixing variable is unobservable. In addition, the unknown parameters (fo' ~, and N most likely change over time, especially when long horizons are considered. Testing the model involves making specific assumptions for the distribution of unobserved variables. Assuming a lognormal distribution for I and a logistic model for the number of traders, Tauchen and Pitts (1983) estimate the model for Treasury bill futures. They find that the model matches general trends in the data reasonably well. 3 The main empirical confirmation of the model is the fact that, as predicted by the theory, variance and volume are positively correlated. Additional evidence can be found from controlled experiments. Batten and Bhar (1993), for instance, explore the V - Ilp2 relation for yen futures across the International Money Market (IMM), during U.S. trading hours, and the Singapore International Monetary Exchange (SIMEX), during Asian trading hours. They find that the volume-volatility correlation is similar across the IMM and the SIMEX 3. Another approach is by Richardson and Smith (1994), who conduct GMM (generalized method of moments) tests of the model by focusing on moments and cross-products of J1.P and V.
24
Philippe Jorion
markets. Given that the volume of trading is much larger on the IMM, they conclude that information emanating from Japan must have a large effect on trading. Another piece of evidence is by Frankel and Froot (1990), who consider the relation between the dispersion of survey forecast, volatility, and volume of trading. They find that dispersion, proxying for the parameter tV, Grangercauses both volume and volatility, which provides some support for the MDH. In this context, implied volatilities may prove more informative than timeseries models since forecasts of Var(tiP) include forecasts of the common noise component, (J"0' of the disagreement parameter tV, of the number of traders N, and of the expected information flow E(l). Simple time-series models are less likely to be able to capture variation in these parameters. 1.1.2
Bid-Ask Spreads
Microstructure theory implies that bid-ask spreads reflect three different types of costs: (1) order-processing costs; (2) asymmetric-information costs; and (3) inventory-carrying costs. Order-processing costs cover the cost of providing liquidity services and are probably small given the size of transactions in the foreign exchange market and the efficiency with which transactions are consummated. Asymmetric-information costs are relevant in the stock market, where corporate officers have access to inside information and analysts actively research firm prospects; given that there is little inside information to trade on in the foreign exchange market, this component is probably small for the foreign exchange market. 4 Finally, inventory-carrying costs are due to the cost of maintaining open positions in currencies and can be related to forecasts of price risk, interest rate costs, and trading activity. When price volatility increases, risk-averse traders increase the spread in order to offset the increased risk of losses. Glassman (1987) reports that spreads increase with recent volatility. Bollerslev and Melvin (1994) and Bessembinder (1994) have also looked at the role of uncertainty in determining bid-ask spread. They find that spreads are positively correlated with GARCH expected volatility. An interesting question is whether volatility forecasts implied in option prices provide a better measure of risk. Regarding the second component of inventory-carrying costs, interest rate costs, Bessembinder (1994) reports that using term structure information as a proxy for the cost of investing capital in short-term investments has little effect on the spread. Therefore, this component will be ignored here. Finally, the third component of inventory-carrying costs involves trading activity. As shown in Glassman (1987) and Bessembinder (1994), there is evidence that, when markets are less active (as before the weekend or a holiday), 4. Lyons (1995), however, showed that marketmakers change prices in response to the perceived informativeness of the quantity transacted. Lyons argues that this finding "calls for a broader conception of what constitutes private information." Perhaps private information consists of information about order flows or price limits.
25
Risk and Turnover in the Foreign Exchange Market
spreads tend to increase. I will thus include variables representing weekend or holiday. Trading activity is also measured by trading volume. Previous authors have shown that spreads are positively correlated with trading volume. Empirically, however, trading volume is highly autocorrelated, implying that movements in volume can be forecast. In addition, expected and unexpected volume can have a different effect on bid-ask spreads. Cornell (1978) argues that spreads should be a decreasing function of volume because of economies of scale leading to more efficient processing of trades and because of higher competition among marketmakers. Therefore, expected trading volume should be negatively related to spread. Easley and O'Hara (1992) formally develop a model implying such a relation. Unexpected trading volume, however, reflects contemporaneous volatility through the mixture of distribution hypothesis and should be positively related to bid-ask spreads. 1.1.3
Implied Volatility
There are only a few studies using the information content of implied standard deviation (ISD) in the foreign exchange market. This is due to the fact that option trading started only in 1982 on the Philadelphia Stock Exchange and in 1984 on the Chicago Mercantile Exchange. It is only now, after ten years, that there may be sufficient data to perform time-series tests with any statistical power. 5 Scott and Tucker (1989) relate the ISD to future realized volatility and report some predictive ability in ISDs measured from Philadelphia Stock Exchange (PHLX) currency options, but their methodology does not allow formal tests of hypotheses. 6 Wei and Frankel (1991) and Jorion (1995) test the predictive power of ISDs by matching ISD with the realized volatility over the remaining days of the option contract. They find that ISDs appear to be biased predictors of future volatility but also outperform time-series models. Even though ISDs should be construed as a volatility forecast for the remaining life of the option, this paper considers only the information content of ISDs for the next trading day. Presumably, better results could be obtained by focusing on short-term options or measuring an instantaneous value of the volatility by extrapolating the term structure of volatility to a very short horizon.?
5. Lyons (1988) used option ISDs over 1983-85 to test whether expected returns on currencies are related to ex ante volatility and found that ISDs can explain some of the movement in expected returns, although he did not test the model restrictions. 6. Scott and Tucker (1989) present one OLS regression with five currencies, three maturities, and thirteen different dates. Because of correlations across observations, the usual OLS standard errors are severely biased, thereby invalidating hypothesis tests. 7. The problem with short-term options is that their "vega" decreases sharply as the option approaches maturity, which implies that ISDs will be measured less accurately, especially if a large fraction of the time value is blurred by bid-ask spreads.
26
Philippe Jorion
1.2 Data and Preliminary Evidence The futures and option data are taken from the Chicago Mercantile Exchange's closing quotes for deutsche mark (DM) currency futures and options on futures over January 1985-February 1992. 8 This represents more than seven years of daily data, or 1,811 observations. I chose deutsche mark futures given that this is the most active currency futures contract. The volume of trading is taken as the total volume of daily trades in deutsche mark contracts. 9 Although the level of futures trading volume is much less than that of the over-the-counter market, it serves as a proxy for the total interbank trading volume. In markets where both spot and futures trading volume can be observed, the two are highly correlated. Data for the bid-ask spreads comes from DRI, up to December 1988, after which the data are collected from Datastream. It should be noted, however, that these quotations are much less reliable than the futures data. Futures data are carefully scrutinized by the exchange because they are used for daily settlement and therefore less likely to suffer from clerical measurement errors. In contrast, institutions reporting bid-ask quotes have no incentive to check the numbers provided; in some instances, there were obvious errors in the data, which have been corrected. Also, the bid-ask spreads reported are only indicative quotes and do not necessarily represent actual trades; banks tend to quote "wide spreads" in order to make sure that all customer transactions fall into the reported spread. Implied volatilities were obtained from contracts with the usual MarchJune-September-December cycle. On the first day of the expiration month, which is the time around which most rollovers into the next contract occur, the option series switches into the next quarterly contract. 10 Daily returns are measured as the logarithm of the futures prices ratio for the un,derlying futures contract. This generates a time series of continuous one-day returns and implied volatility. Although the implied volatility is strictly associated with the volatility over the remaining life of the contract, it presumably also contains substantial information for the next day volatility. Table 1.3 presents preliminary regressions with volume and volatility. Standard errors are heteroskedastic consistent, using White's (1980) procedure. The top panel reports results from regressing log volume on a time trend. The relation is strong and significant. Trading activity increases with time, reflecting
8. Options on futures started to trade in January 1984, but volume was relatively light in that year. In addition, there were price limits on futures, which were removed on 22 February 1985. 9. The face value of one contract is DM 125,000. Volume is thus measured in deutsche marks, although turnover could also be measured in dollars. 10. Some error might be imparted in implied volatilities if options trade with a bid-ask spread or if option hedging entails costs. Leland (1985) shows how costs tend to increase the observed ISD. Given the very low costs of transacting in the foreign exchange markets, however, the bias is very small.
27
Risk and Turnover in the Foreign Exchange Market Unconditional Regressions with Volume and Variance
Table 1.3
Regressors Volume
Model
Constant
Time
Volume
9.903 (234.44) .707 (7.53) -8.626 (-8.21) -10.892 (-8.74)
.00036* (9.72) -.00010 (-1.08)
Variance Variance Variance
-.00052* (-5.50)
.186 .002 .904* (8.62) 1.171* (9.17)
.096 .132
Note: Regressions of log volume and variance on a time trend and log volume. Volume is the
number of contracts traded daily; variance is measured as the squared log return on the nearby futures contract. The period is January 1985-February 1992 (1 ,811 observations). Asymptotic t-statistics are in parentheses. Standard errors are heteroskedastic consistent. *Significantly different from zero at the 5 percent level.
the increasing number of traders. The second panel finds a negative but weak correlation between variance and the time trend. In the third panel, variance is found to be strongly contemporaneously correlated with volume; these results are in line with most of the volume-volatility literature. Finally, the fourth panel shows that risk is positively correlated with volume and at the same time negatively correlated with the time trend. This is generally consistent with the Tauchen-Pitts model, where the disagreement component of risk decreases because of averaging over an increasing number of traders. These results, however, should be explored further by distinguishing between expected and unexpected volatility.
1.3 Measuring Expectations 1.3.1
Time-Series Model for Volatility
Expected volatility is measured using a simple but robust time-series model, the GARCH(I,I) mode1. 11 The GARCH model, developed by Engle (1982) and extended by Bollerslev (1986), posits that the variance of returns follows a deterministic process, driven by the latest squared innovation and by the previous conditional variance:
(3)
Rt
=
f.1
+
r t,
r t - N(O, h t ),
ht
= a o + a1r;-1 +
~ht-l'
where R t is the nominal return, rt is the de-meaned return, and h t is its conditional variance, measured at time t. To ensure invertibility, the sum of parame11. For evidence on the GARCH(I,1) model applied to exchange rates, see, e.g., Hsieh (1989).
28
Philippe Jorion
Table 1.4
Modeling Volatility R t = f.l
Model
IJv
Normal
.0304 (1.64) .0299 (1.65)
GARCH
+ r t,
rt
-
N(O, h),
(Xo
ht =
Uo
~
(XI
.619* (30.04) .027* (4.50)
+ ulr~_1 +
Ilh t - 1t
Log-Lik. 6,187.23
.0785* (4.53)
.8802* (72.26)
6,242.09
109.71 [.000]
tWhere Rt is defined as the return on currency futures, expressed in percentages, and ht is the conditional variance of the innovations. The period is January 1985-February 1992. Asymptotic t-statistics are in parentheses; p-values are in square brackets. The X2 statistic tests the hypothesis of significance of added GARCH process. *Significantly different from zero at the 5 percent level. Table 1.5
Modeling Volume Stationarity: 4 log (Vr) = a + ht ARMA: log (V) = a + ht + E t,
Model
Constant
Time
Stationarity
4.3781 (23.65) 9.8942 (133.54)
.00017 (11.50) .00037* (5.21)
-.442* (-23.72) 1.306* (31.32)
ARMA
+ ~1 log (Vt - 1) + "t; E = ~IEt-l + ~2Et-2 + 6 1"t-l + "/ t
81
R2 .2186
-.335* (9.68)
.852* (27.13)
.4625
tTime-series model for log (V), where V is the number of contracts traded daily. The period is January 1985-February 1992. Asymptotic t-statistics are in parentheses. *Significantly different from zero at the 5 percent level.
ters (Ci l + (3) must be less than unity; when this is the case, the unconditional, long-run variance is given by Ci o/(l - Ci l - (3). Estimates of the GARCH(l,l) process are presented in table 1.4. In line with previous research, I find that the GARCH model is highly significant, with ax 2(2) statistic exceeding 100. This is much higher than the 1 percent upper fractile of the chi square, which is 9.2. There is no question, therefore, that realized volatility does change over time. The process is persistent but also stationary, with values of (Ci l + (3) around 0.96. This number implies that a shock to the variance has a half-life oflog(0.5)/log(0.96), which is about seventeen days. The conditional variance generated by this model will be taken as the time-series forecast of risk. Note that the GARCH model will be given the benefit of the doubt, by using "ex post" parameter values estimated over 198592, whereas ISDs have access only to past information. 1.3.2
Time-Series Model for Volume
To model expected volume, one must first assess whether volume is stationary. If not, first differences should be taken. To test for trend stationarity, I regress the daily change in volume on a trend and the lagged volume:
29
Risk and Turnover in the Foreign Exchange Market
a log(V
(4)
t)
== a + bt +
Estimates of the regression are presented in table 1.5. The t-statistic on
vt
==
log(Vt ) == a
+ bt + 8 t ,
==
8t
+
An ARMA(2, 1) process appears to provide a parsimonious fit since upperorder terms are not significant. The time-series model allows us to decompose the volume into an expected component, ErCVt+l)' and an error process ut • Estimates of the ARMA process are presented in table 1.5. The ARMA coefficients are highly significant, as is the time trend coefficient. There was a marked upward trend in the number of future contracts traded over 1985-92, implying an annual growth of 9 percent. When measured in dollars, the volume of trading has grown at an annual average rate of f 9 percent over this period.
1.4
Computing Implied Volatilities
Implied volatilities are derived from the Black (1976) model for European options on futures:
(6)
c
= [FN(d]) - KN(dz)]e-
d2 == d 1
-
rT
,
d]
_ log(F/K) -
(J~
(J"~ + -2-'
(J"~,
where F is the futures rate, K is the strike price, ,. is the time to option expiration, r is the risk-free rate (taken as the Eurodollar rate), and (J" is the volatility. Note that the futures contract might expire later than the option contract, in which case F is related to the spot through a cost-of-carry relation involving the time to expiration of the futures contract. For a given option price, inverting the pricing model yields an implied standard deviation. Because Beckers (1981) showed that using at-the-money options was preferable to various other weighting schemes, only at-the-money calls and puts are considered here. In addition, these are the most actively traded and therefore the least likely to suffer from nonsimultaneity problems. On any given day, one computes the ISD as the arithmetic average of that obtained from the two closest at-the-money call and put options. These options have the highest "vega," or price sensitivity to volatility, and therefore should provide the most accurate estimates of volatility. Averaging over one call and one put lessens the effect of bid-ask spreads and of possible nonsynchronicity between futures and option prices.
30
Philippe Jorion
Since CME options are of the American type, using a European model introduces a small upward bias in the estimated volatility. This bias is generally small for short-maturity options. 12 For instance, with typical parameter values, using a European model overestimates a 12 percent true volatility by reporting a value of about 12.02 percent. 13 The difference, however, is less than half of typical bid-ask spreads when quoted in terms of volatility and thus barely economically significant. Another potential misspecification is that the Black-Scholes model is inconsistent with stochastic volatilities. If volatility changes in a deterministic fashion, ISD can be construed as an average volatility over the remaining life of the option. But, if volatility is stochastic, there is more than one source of risk in options, and the arbitrage argument behind the Black-Scholes option pricing model fails. Recent papers by Hull and White (1987), Scott (1987), and Wiggins (1987) have examined the pricing of options on assets with stochastic volatility. The general approach to pricing options in these papers is to treat the volatility as a random state variable. In order to derive tractable results, the innovations in volatility and returns are generally assumed to be uncorrelated; prices are then calculated by Monte-Carlo simulation. Scott (1988) and .Chesney and Scott (1989), for instance, present a careful empirical analysis of the random variance model (implemented on a Cray supercomputer) and find that the random variance model actually provides a worse fit to market prices than the BlackScholes model using ISDs.14 For U.S. stock options, differences are on the order only of $0.02, much lower than typical bid-ask spreads of $0.05-$0.25. Duan (1995) extends the risk-neutral valuation to the case where logarithmic returns follow a GARCH process. Under some combination of preferences and distribution assumptions, he derives a GARCH option-pricing model, but the magnitude of the bias, computed by simulations, is very small, at most $0.10$0.15 for at-the-money options on a $100 underlying asset. Because options with stochastic volatility are priced using Monte-Carlo
12. The bias depends on the difference between U.S. and foreign interest rates. When U.S. rates are higher than foreign rates, the American premium on spot currency options is close to zero for calls and positive for puts. Jorion and Stoughton (1989) compare market prices of American PHLX (Philadelphia Stock Exchange) and European CBOE (Chicago Board Options Exchange) options and find that differences are minor, essentially undistinguishable from bid-ask spreads. Adams and Wyatt (1987) and Shastri and Tandon (1986) use numerical procedures to show that biases in measured implied volatilities are generally minor for short-term at-the-money options. 13. With a futures prices of $0.50, a strike price of 50, a U.S. interest rate of 6 percent, 50 calendar days to expiration, and a true volatility of 12 percent, the values of an American and a European call are 0.8799 and 0.8786, respectively. Inverting the American call value using a European model yields an apparent volatility of 12.02 percent. With the same parameters but 95 days to expiration, the estimated volatility is 12.04 percent. With 5 days to expiration, it is 12.00 percent. 14. Melino and Turnbull (1990) compare option prices derived from Black-Scholes and a stochastic volatility model, using parameters derived from the time-series process, and find that the stochastic volatility model provides a better fit to options than the standard model using historical volatility. They do not, however, consider a Black-Scholes model with implied volatility.
31
Risk and Turnover in the Foreign Exchange Market
Table 1.6
Comparison of Volatility Regressions + b 1u;,ISD + b/lt+1 + b 3 E,(v)
R;+1 = a
+ C[Vt+1
- E,(v)]
+
Et+1
t
Slopes on:
a -.117 (.095)
ISD
GARCH
E(v)
v - E(v)
.0464
1.192* (.182)
.0243
.724* (.126)
.164 (.071)
R2
.0304
.598* (.079)
.113 (.063)
.0465
-.123 (.094)
1.150* (.243)
.051 (.159)
-.659 (1.051)
1.153* (.243)
.037 (.160)
.053 (.104)
.741 * (.119)
.038 (.098)
1.540* (.122)
.1737
.178 (.098)
1.532* (.122)
.1493
.055 (.097)
1.500* (.116)
.1873
-.237 (.989) -1.201 (.999) -.628 (.985)
.906* (.220)
.206 (.148)
.0467
tVariance over next day is related to forecast variance from option implied standard deviation (ISD), (J~SD, GARCH(l,1) forecast, h,+l' expected log volume from ARMA time-series model, E,(v), and unexpected log volume over next day, [V'+l - E,(v)]. The period is January 1985February 1992. Heteroskedastic-consistent standard errors are in parentheses. *Significantly different from zero at the 5 percent level.
methods, no published research has ever recovered the implied (instantaneous) standard deviation from a stochastic volatility model. Recently, however, Heston (1993) has developed a closed-form solution that efficiently computes option values under stochastic volatility. To implement this model, the researcher requires knowledge of additional parameters, including those describing the time-series process for the volatility, as well as the price of volatility risk. In summary, although stochastic volatility models are theoretically more appealing than the standard Black-Scholes approach, they have severe shortcomings. Besides computational costs, the estimation of many additional parameters introduces elements of uncertainty. In the debate between purists and empiricists, my view is that the Black-Scholes approach, a simple and robust model, provides a sufficient approximation to ISDs.
1.5 Empirical Results The mixture of distribution hypothesis postulates a positive relation between volume and volatility for a given number of traders. To capture this relation, I
32
Philippe Jorion
estimate a regression of the squared return on expected variance and an innovation component: (7)
The advantage of this approach is that slow changes in (10' tV, and N may be captured by the rational forecast ElR~+l). In the above regression, we expect the coefficient c to be positive. Lamoureux and Lastrapes (1990) apply a GARCH model to a sample of twenty stocks and find that GARCH effects disappear once volume is included as an exogenous variable. They interpret this evidence as support for the hypothesis that GARCH effects are a manifestation of the time dependence in the rate of information arrival to the market. This, however, assumes that the best available forecasts of volatility are generated by a GARCH model. In fact, better forecasts may be available from the option markets. The issue is whether the correlation between volatility and volume remains in the presence of implied volatilities. If not, the usefulness of the mixing model would be in serious doubt. To test the information content of various forecasts, table 1.6 reports regres-
Comparison of Bid-Ask Spread Regressions + b l O';,ISD + b2ht+1 + b3 E t- l (v) + c[vt
Table 1.7
St = a
-
Et-l(v)]
+ dDt + E/
Slopes on: a
ISD
.040 (.004)
.1055* (.0084)
GARCH
v - E(v)
Fri./Hol.
R2 .1728
.061 (.004)
.0705* (.0082)
.085 .003
.1095 .0020 .0036
.038 (.005)
.0914* (.0093)
.0170 (.0086)
.152 (.042)
.0897* (.0093)
.0212* (.0095)
.084 (.037) .149 (.042)
E(v)
.0886* (.0093)
.0220* (.0095)
.0001 .1761
-.0113* (.0043)
.1792
.0021 (.0036)
.0078 (.0062)
-.0114* (.0043)
.0016 (.0058)
.0019 .0108* (.0032)
.1850
tBid-ask spread measured in deutsche marks is related to forecast variance from option implied standard deviation (ISD), SD , GARCH(1,l) forecast, ht+i' expected log volume from ARMA time-series model, Et_l(v), unexpected log volume [v t- Et-l(v)], and Friday-holiday dummy variable D t • The period is January 1985-February 1992. Heteroskedastic-consistent standard errors are in parentheses. *Significantly different from zero at the 5 percent level.
cr:
33
Risk and Turnover in the Foreign Exchange Market
sions of the one-day squared return against several predetermined variables: (8)
where (J"~SD is the option IDS, ht + 1 is the GARCH forecast using information up to time t, and Et(v) is the expected volume, also measured at time t. All predetermined variables-the implied variance, the GARCH forecast, and the expected volume-are positively, and significantly, related to future risk. More interestingly, when pitting all three forecasts against each other, only the implied variance appears significant. Note that these results are particularly impressive since the GARCH model was given the benefit of the doubt, using "ex post" parameter values estimated over 1985-92. In contrast, ISDs have access only to past information. The table also shows that ISDs are nearly unbiased forecasts of the next day's variance, with the slope coefficients generally close to unity. ISDs, in theory the best forecast of volatility over the remaining life of the option, are also proving to be useful short-term forecasts. Focusing on volume, regressions of risk on expected and unexpected volume indicate that the strongest association appears between risk and unexpected volume, as predicted by the Tauchen-Pitts model. The last regression in the table uses the three predetermined variables as well as the unexpected volume variable. The positive relation between risk and unexpected volume is still strong, as predicted by the information-flow model. However, in contrast with the Lamoureux-Lastrapes results, measures of ex ante risk are still significant. Even when volume measures are included, the GARCH forecast is stilrsig-
Table 1.8
Using Spreads to Forecast Volatility + b 10';,ISD + b/tt+l + b 3 E,(v) + b 4S, +
R;+l = a
Et+l
t
Slopes on:
a
ISD
GARCH
E(v)
R2
1.623* (.702)
.0055
.685* (.133)
.560 (.696)
.0249
-.411 (.577)
.0468
-.398 (.578)
.0469
.449 (.067) .130 (.030)
Spread
-.107 (.101)
1.187* (.223)
.058 (.163)
-.610 (1.056)
1.189* (.222)
.045 (.166)
.050 (.104)
tVariance over next day is related to forecast variance from option implied standard deviation (ISD), (J'~SD, GARCH(l,l) forecast, ht+l' expected log volume from ARMA time-series model, Et(v), and bid-ask spread, St' The period is January 1985-February 1992. Heteroskedastic-consistent standard errors are in parentheses. *Significantly different from zero at the 5 percent level.
34
Philippe Jorion
nificant. This suggests that expected volatility captures some of the time variation in the information-flow variable. Next, table 1.7 reports various regressions of the bid-ask spread St against the same variables as in table 1.6 above. The most general setup is (9)
+ b 1u},ISD + b2ht+1 + b 3E t- 1(Vt) + c[Vt + dDt + 0t'
St == a
E t- 1(Vt)]
where variables are defined as above, and D t is a dummy variable set to one on a Friday or before a holiday. The first two regressions show that the spread is significantly positively related to measures of risk, separately taken as the implied variance and the GARCH variance; the spread is not related to expected volume. When comparing GARCH and implied volatilities, we again find that there is little information content in GARCH forecasts besides that in implied volatility. Finally, the bottom of the table reports the results using all regressors, the Friday/holiday indicator, three predetermined variables, and unexpected volume. Confirming previous research, spreads increase on a Friday or before a holiday. Spreads also increase with implied and GARCH variances but decrease with expected volume, as predicted. These results confirm that bid-ask spreads reflect inventory-carrying costs that primarily depend on price uncertainty and trading activity. Finally, table 1.8 investigates whether the spread contains information above and beyond that in other risk forecasts. The full regression is (10)
R~+1
== a +
bl(J'~,ISD
+ b2ht+1 + b 3E lv t+1) + b4St + 0t+l.
The first panel, using the spread as the only regressor, shows that the spread is a significant leading indicator of volatility.. However, in the full regression reported at the bottom of the table, the coefficients b2 , b3 , and b4 are all insignificantly different from zero. This confirms that neither GARCH forecasts, nor expected volume, nor spreads, have any information content beyond that in ISDs. Options appear to embody all economically relevant information for future risk.
1.6
Conclusions
Many elements of the microstructure of the foreign exchange market depend critically on perceived risk. Bid-ask spreads should increase with inventorycarrying costs, which depend on risk forecasts. Volume is positively correlated with volatility through the mixture of distribution hypothesis. The premise of this paper was that risk measures contained from option prices, ISDs, provide superior forecasts for exchange rate volatility. Indeed, the paper reports that ISDs are markedly superior to the current state of the art in time-series volatility forecasting; GARCH models appear to contain no information besides that in ISDs. Neither do expected volume or bid-ask spreads.
35
Risk and Turnover in the Foreign Exchange Market
Further, ISDs also dominate all other risk measures for the purpose of explaining bid-ask spreads. Studies of the stock market, in contrast, find that there is not much information in ISDs. Canina and Figlewski (1993) analyze S&P100 index options and find that ISDs have little predictive power for future volatility and appear to be even worse than simple historical measures. Lamoureux and Lastrapes (1993) focus on individual stock options and find that historical time series contain predictive information over and above that of implied volatilities. My results are in sharp contrast to those of the stock option literature and may be indicative of measurement problems in the stock option market. if the arbitrage between options and the underlying stocks is costly, then there may be deviations between options and underlying stock prices. Alternatively, nonsynchronicity in the stock index value may induce measurement errors in implied volatilities. Because of the depth and liquidity of CME futures and options, traded side by side in the same market, implied volatilities are less likely to suffer from the measurement problems that affect stock options and provide better measures of forecast volatility. The superiority of ISDs is reassuring because it indicates that option traders form better expectations of risk over the next day than statistical models, even when the latter are based on "ex post" parameter values. To some extent, these results were expected since time-series models are unable to account for events such as regular announcements of macroeconomics indicators, meeting of G-7 finance ministers, and so on. Because the timing of these events is known by the foreign exchange market, we would expect options to provide better forecasts than naive time-series models. Using ISDs, the paper confirms the positive relation between unexpected risk and unexpected volume predicted by the mixture of distribution hypothesis. In contrast with results in the stock market, however, we find that expected variance does not disappear when volume is included in the variance equation. The paper also finds that spreads are positively correlated with expected risk. Overall, the information content of ISDs suggests that an important aspect of anticipated risk is ignored when focusing solely on time-series models of volatility.
References Adams, P., and S. Wyatt. 1987. On the pricing of European and American foreign currency call options. Journal of International Money and Finance 6:315-38. Batten, J., and R. Bhar. 1993. Volume and price volatility in yen futures markets: Within and across three different exchanges. Working paper. Sydney: Center for Japanese Economic Studies, Macquarie University. Beckers, S. 1981. Standard deviations implied in option prices as "predictors of future stock price variability." Journal of Banking and Finance 5: 363-81.
36
Philippe Jorion
Bessembinder, H. 1994. Bid-ask spreads in the interbank foreign exchange markets. Journal of Financial Economics 35:316-48. Black, F. 1976. The pricing of commodity contracts. Journal of Financial Economics 3:167-79. Bollerslev, T. 1986. Generalized autoregressive conditional heteroskedasticity. Journal of Econometrics 31 :307-27. Bollerslev, T., and M. Melvin. 1994. Bid-ask spreads and volatility in the foreign exchange market: An empirical analysis. Journal of International Economics 36:355-72. Canina, L., and S. Figlewski. 1993. The informational content of implied volatility. Review of Financial Studies 6:659-81. Chesney, M., and L. Scott. 1989. Pricing European currency options: A comparison of the modified Black-Scholes model and a random variance model. Journal of Financial and Quantitative Analysis 24:267-84. Clark, P. 1973. A subordinated stochastic process model with finite variance for speculative prices. Econometrica 41: 135-55. Cornell, B. 1978. Determinants of the bid-ask spread on forward foreign exchange contracts under floating exchange rates. Journal of International Business Studies 9:33-41. Cornell, B. 1981. The relationship between volume and price variability in futures markets. Journal of Futures Markets 1:303-16. Dickey, D., andA. Fuller. 1979. Distribution of estimators for autoregressive time series with a unit root. Journal of the American Statistical Association 74:427-31. Duan, J.-C. 1995. The GARCH option pricing model. Journal of Mathematical Finance 1:13-32. Easley, D., and M. O'Hara. 1992. Adverse selection and large trade volume: The implications for market efficiency. Journal of Financial and Quantitative Analysis 27: 185-208. Engle, R. 1982. Autoregressive conditional heteroskedasticity with estimates of the variance of United Kingdom inflation. Econometrica 50:987-1007. Frankel, 1., and K. Froot. 1990. Exchange rate forecasting techniques, survey data, and implications for the foreign exchange market. Working Paper no. 3470. Cambridge, Mass.: National Bureau of Economic Research. Glassman, D. 1987. Exchange rate risk and transaction costs: Evidence from bid-ask spreads. Journal of International Money and Finance 6:479-90. Grammatikos, T., and A. Saunders. 1986. Futures price variability: A test of maturity and volume effects. Journal of Business 59:319-29. Heston, S. 1993. A closed-form solution for options with stochastic volatility with applications to bond and currency options. Review of Financial Studies 6:327-43. Hsieh, D. 1989. Modeling heteroskedasticity in daily foreign exchange rates. Journal of Business and Economic Statistics 7:307-17. Hull, J., and A. White. 1987. The pricing of options on assets with stochastic volatility. Journal of Finance 42:281-300. Karpoff, J. 1987. The relation between price changes and trading volume: A survey. Journal of Financial and Quantitative Analysis 22: 109-26. Jorion, P. 1995. Predicting volatility in the foreign exchange market. Journal of Finance 50:507-28. Jorion, P., and N. Stoughton. 1989. An empirical investigation of the early exercise of foreign currency options. Journal of Futures Markets 9:365-75. Lamoureux, C., and W. Lastrapes. 1990. Heteroskedasticity in stock return data: Volume versus GARCH effect. Journal of Finance 45:221-29. - - - . 1993. Forecasting stock-return variance: Toward an understanding of stochas-
37
Risk and Turnover in the Foreign Exchange Market
tic implied volatilities. Review of Financial Studies 6:293-326. Leland, H. 1985. Option pricing and replication with transaction costs. Journal of Finance 40: 1283-1301. Lyons, R. 1988. Tests of the foreign exchange risk premium using the expected second moments implied by option pricing. Journal of International Money and Finance 7:91-108. - - - . 1995. Tests of microstructural hypotheses in the foreign exchange market. Journal of Financial Economics 39:321-51. Melino, A., and S. Turnbull. 1990. Pricing foreign currency options with stochastic volatility. Journal of Econometrics 45:239-65. Richardson, M., and T. Smith. 1994. A direct test of the mixture of distribution hypothesis: Measuring the daily flow of information. Journal of Financial and Quantitative Analysis 29: 101-16. Scott, E., and A. Tucker. 1989. Predicting currency return volatility. Journal ofBanking and Finance 13:839-51. Scott, L. 1987. Option pricing when the variance changes randomly: Theory, estimation and an application. Journal of Financial and Quantitative Analysis 22: 419-38. Scott, L. 1988. Random variance option pricing: Empirical tests of the model and deltasigma hedging. University of Illinois. Mimeo. Shastri, K., and K. Tandon. 1986. On the use of European models to price American options on foreign currency. Journal of Futures Markets 6:93-108. Tauchen, G., and M. Pitts. 1983. The price variability-volume relationship on speculative markets. Econometrica 51 :485-505. Wei, S., and J. Frankel. 1991. Are option-implied forecasts of exchange rate volatility excessively variable? Working Paper no. 3910. Cambridge, Mass.: National Bureau of Economic Research. White, H. 1980. A heteroskedastic-consistent covariance matrix estimator and a direct test for heteroskedasticity. Econometrica 48:817-38. Wiggins, J. 1987. Options values under stochastic volatility: Theory and empirical estimates. Journal of Financial Economics 19:351-72.
Comment
Bernard Dumas
Philippe 10rion's paper contains empirical tests of three hypotheses relating bid-ask spreads in the futures market for currencies, expected and unexpected volume, and price volatility. For reasons given in the paper, it is claimed that expected volume would be negatively related to spreads and unexpected vol.. ume positively related to them. A mixing variable, I, representing the rate at which information arrives on the foreign exchange market, further leads 10rion to hypothesize a positive relation between unexpected volume and price volatility.
Bernard Dumas is on the faculty of the Hautes Etudes Commerciales School of Management in France. He is a research professor at Duke University and a research associate of the National Bureau of Economic Research, the Centre for Economic Policy Research, and Delta.
38
Philippe Jorion
I have four comments on this paper, three of which pertain to the way in which variables are measured, and one of which pertains to the manner in which the mixing variable is specified.
The Measurement of Volume Volume in this study is the volume of trade in the futures market. Jorion points out that the futures market in foreign exchange is a very small part of the total world foreign exchange market. More damaging may be the observation that the futures market is not "representative" of the overall market because it is a centralized, organized market while the bulk of the market is an interbank, decentralized market in which traders cannot observe order flows and volume of trade. The way in which people channel their orders to the futures as opposed to the interbank market would presumably depend on the institutional features of the two types of markets. I The decision to trade on one market rather than the other is a missing variable in the theory being tested here. That missing variable may obfuscate the test of the relation between volume and price volatility. In a very indirect attempt to estimate the severity of this problem, one could make use of the fact that futures markets in foreign exchange are not the only centralized exchange on which volume is directly observable. There exist centralized foreign exchange options markets. Jorion could measure the correlation between volume in the futures and the options markets. That measurement would, of course, leave unobserved the degree to which trades shift between organized and decentralized trading places.
The Role of the Mixing Variable in the Model Specification This study exhibits an apparent contradiction concerning the way in which the mixing variable, I, is specified. This mixing variable is either a random variable or a random process. To keep the discussion simple, let us imagine that the derivations leading to equation set (2) of the paper remain valid under either formulation. The empirical analysis, however, needs to be adapted depending on the assumed specification. If the mixing variable is a random variable, then, according to the third equation of equation set (2), there is indeed a constant positive covariance between unexpected volume and squared price increments, as claimed. But, under the same assumption, the first two equations in equation set (2) make it plain that (expected) volatility and expected volume are constant over time. It is then incoherent to proceed to estimate movements of these quantities, as Jorion does. If the mixing variable follows a stochastic process, expected volume and 1. By way of analogy, see the recent work of Easley, O'Hara, and Srinivas (1994) on the choice made by informed traders to trade in the options market or in the underlying cash market.
39
Risk and Turnover in the Foreign Exchange Market
volatility are allowed to change over time, but the relation between unexpected volume and squared price increments is not stable over time (although it always has the same sign). It is not clear then that this relation can be captured simply by measuring the cross-moment (sample covariance or sample correlation) between the observed values of these two variables. This would have to be shown. The Measurement of Volatility in the Presence of a Mixing Variable In this study, (expected) volatility is measured in two separate ways. One is the Black-Scholes implied standard deviation. The other measurement is based on the estimation of a GARCH process. Both the Black-Scholes model and the GARCH model are specified in real time, when in fact the presence of a mixing variable in the model being tested would require the use of a random time deformation. Several authors have adapted the theory of option pricing to random time scales. 2 It would have been preferable to estimate real-time implied volatility on the basis of a random-time option-pricing model since, under the null hypothesis, time does flow randomly in the foreign exchange market. GARCH models have also been extended to random times by, for example, Stock (1988). In a recent study, Ghysels and Jasiak (1994) fit a random-time GARCH model to the daily time series of the S&P500 index from 1950 to 1987. The fit of the GARCH model identifies clear accelerations of time on the stock exchange. Ghysels and Jasiak also show that the estimated volatility under time deformation follows a much smoother path than in the absence of time deformation. Following Frenkel and Levich (1977) and many others, we have every reason to believe that the foreign exchange market also goes through tranquil (slow time) and turbulent (fast time) periods. That aspect of the behavior of the market is neglected by Jorion when he measures volatility, even though the theory being tested specifically incorporates a mixing variable. The Measurement of Short-Lived Volatility Changes from Medium-Term Options In this study, Jorion endeavors to measure volatility changes on a day-to-day basis. However, when the Black-Scholes implied standard deviation is used for the purpose, the options that serve as a basis for the measurement are mediumterm options (several weeks to maturity). No overnight options are available to allow the measurement of daily volatility. This difficulty is pointed out in the paper. How serious is it? It all depends on whether volatility changes are typically short lived or long lived. If they are long lived, the problem is not as serious as it is if they are short lived. 2. In that extension, the pure arbitrage foundation of the Black-Scholes theory is lost.
40
Philippe Jorion
The success of GARCH models in fitting financial time series is a testimony to the degree of persistence of volatility. However, there exists also evidence that volatility changes are short lived following the arrival of a piece of news. Ederington and Lee (1993), for instance, study the effect of scheduled macroeconomic news on the stock market. They find that the volatility is only slightly elevated for a few hours after the announcement. Donders and Vorst (1994) study the effect of firm-specific, scheduled news releases on implied stockprice volatility. They find that volatility rises steadily for a few days prior to the event date and then drops back to a normal level almost immediately. Such evidence calls into question the method used in the present study to measure short-term volatility. Conclusion
Having not done similar work myself, I am not in a position to ascertain whether the apparent shortcomings that I have identified are capable of overturning the results of Philippe lorion's study. His main conclusion-that neither GARCH modeling nor the information provided by spreads of volume is capable of improving on the Black-Scholes implied standard deviation as a measurement of expected volatility-is a strong one and one that will no doubt generate a lot of interest and controversy. References Donders, M. W. M., and T. C. F. Vorst. 1994. The impact of firm-specific news on implied volatilities. Working paper. Erasmus University, Rotterdam. Easley, D., M. O'Hara, and P. S. Srinivas. 1994. Option volume and stock prices: Evidence on where informed traders trade. Working paper. Cornell University. Ederington, L. H., and J. H. Lee. 1993. How markets process information: News releases and volatility. Journal of Finance 48: 1161-91. Frenkel, J. A., and R. M. Levich. 1977. Transactions costs and interest arbitrage: Tranquil vs. turbulent periods. Journal of Political Economy 85: 1209-26. Ghysels, E., and J. Jasiak. 1994. Stochastic volatility and time deformation: An application to trading volume and leverage effects. Discussion paper. Centre de Recherche et Developpement Economique, Universite de Montreal. Stock, J. H. 1988. Estimating continuous time processes subject to time deformation. Journal of the American Statistical Association 83 :77-84.
2
Bid-Ask Spreads in Foreign Exchange Markets: Implications for Models of Asymmetric Information David A. Hsieh and Allan W. Kleidon
The term market microstructure was coined in 1976 by Mark Garman to define "moment-to-moment trading activities in asset markets." With the stated goal of providing insight and testable implications regarding the transaction-totransaction nature of realistic exchange processes, Garman examines dealership and auction models of marketmakers where the stream of market orders from a collection of market agents is depicted as a stochastic Poisson process. The resulting insights concern bid-ask spreads (based on standard microeconomic analysis), inventories of marketmakers, and the effects of some market power on the part of marketmakers. A more recent microstructure literature is based on information asymmetries among economic agents. In a recent literature review, Admati (1991) defines the area of market microstructure as "the literature on asset markets with asymmetric information and especially on trading mechanisms" (p. 347); Garman is noted as an example of "earlier market microstructure literature" in which "information issues were not typically modelled" (p. 355, n. 11). Two reasons are given for the popularity of asymmetric information models: policy implications of different trading processes, as exemplified by the 1987 crash, and empirical results such as various patterns in trading volume, variances, and returns over the trading day. The belief is that better insights into both issues will be achieved by examining information asymmetries. Moreover, observed empiri-
David A. Hsieh is professor of finance and economics at the Fuqua School of Business of Duke University. Allan W. Kleidon is vice president at Cornerstone Research and a consulting professor of law (in finance) at the School of Law of Stanford University. Support for this research has been provided by Cornerstone Research. The authors are grateful for helpful comments from Jeffrey Frankel, Paul Pfleiderer, Antti Suvanto, currency traders at Citibank and at Wells Fargo-Nikko Investment Advisors (Jeffrey Hord and Vikas Srivastava), two anonymous referees, seminar participants at the Haas School of Business, University of California, Berkeley, and participants at the conference on the microstructure of foreign exchange markets.
41
42
David A. Hsieh and Allan W. Kleidon
cal results are believed to "call for theoretical explanations beyond what can be obtained by traditional models (in which informational asymmetries are not present)" (p. 348). In some ways, foreign exchange data have institutional features that are ideal for testing these now "standard" asymmetric information models. The market is very liquid and is linked around the world by computerized information systems, and the commodity is essentially the same in all markets. Standard information models have been applied to foreign exchange data in Bollerslev and Domowitz (1993), particularly to the behavior of bid-ask spreads and volume around the open and close of trading in regional markets. Bollerslev and Domowitz conclude that data for smaller banks that "operate mainly within regional markets with well-defined market openings and closings" (p. 1422) show the relation between trading activity and spreads that is implied by standard asymmetric information models. In this paper, we further examine how well standard asymmetric information models can explain the behavior of volatility, bid-ask spreads, and volume around the open and close of trading in foreign exchange markets.! We conclude that the standard information models are unable to explain these data. Our analysis differs from most previous studies in that we examine the implications of standard information models for the behavior of data across markets that are open simultaneously, rather than looking at markets in isolation. In particular, a feature of standard information models is that high volatility is associated with trades by privately informed traders whose trading activity incorporates their information into prices and quotes. Within this class of model, if new information results in high volatility of quotes for a trader located in London, then the quotes for a trader who is physically located in New York but who observes the London quotes will also show high volatility. We exploit the fact that foreign exchange transactions occur virtually around the clock, with overlap between the trading day for traders in London and New York. Consequently, the open of trade in New York and the close of trade in London correspond to times when the other market has been trading for some time. We find that the high volatility that shows up at the open in New York and the close in London appears to be unrelated to the concurrent volatility in the other market, even though both sets of quotes appear on exactly the same trading screens at exactly the same time. This volatility cannot be due to new information reaching one market but not the other, within the standard information models. Either these markets that are ostensibly closely linked are segmented in important ways not recognized in standard models, or some phenomenon other than the incorporation 1. Trading patterns at the open and close of trade have been extensively studied within asymmetric-information models, primarily with respect to the New York Stock Exchange (NYSE), but also in the context of cross-country listings and foreign exchange (see, e.g., Admati and Pfleiderer 1988; Subrahmanyam 1991; Freedman 1989; Barclay, Litzenberger, and Warner 1990; and Bollerslev and Domowitz 1993). For an excellent survey of the literature, see Admati (1991).
43
Bid-Ask Spreads in Foreign Exchange Markets
of private information must be responsible for the behavior of quotes. Given the high degree of integration of the international foreign exchange market, we conclude that the observed periodicity of volatility is not due to the incorporation of private information as envisioned by standard asymmetric information models. One way of stating the problem is that if no new information is reaching the international foreign exchange market-which is implied by the absence of unusual volatility in quotes generated by traders in one physicallocation-then quotes generated by traders in another market show excess volatility relative to that implied by standard information models. This is not a new phenomenon. For example, the crash of October 1987 is an example of a large change in stock prices that does not appear to have been caused by new information reaching the market as a whole. However, a recent class of model has been developed that explains such phenomena as the crash in terms of imperfect information aggregation and learning by market participants rather than new information reaching the market as a whole. These models differ from standard asymmetric information models by relaxing the assumption that each trader has perfect knowledge about the structure of the market, that is, about the preferences and beliefs of all traders in the market. 2 While this is a different type of model from standard asymmetric information models, it is possible that at least some of the observed behavior in foreign exchange markets may be attributable to this type of information asymmetry. We examine this possibility below, after closely examining the standard asymmetric information models. It appears that some form of learning about the market structure is important at the start of trading, which results in wide and volatile quotes when traders first enter the market. This process may be as informal as the posting of wide quotes with little expectation of trading during the initial period of learning. At the close of trading, standard information models are again unable to explain the foreign exchange data. We conclude that inventory management by marketmakers in the closing market appears to be the most likely explanation. The paper proceeds as follows. Section 2.1 discusses the institutional features of our data and presents our empirical results. Section 2.2 examines the standard asymmetric information models of intraday price and volume of Admati and Pfleiderer (1988) and Subrahmanyam (1989, 1991) and the current application of these models to foreign exchange data in Bollerslev and Domowitz (1993).3 Section 2.3 examines alternate explanations of the results in
2. These issues are addressed in Gennotte and Leland (1990), Jacklin, Kleidon, and Pfleiderer (1992), Kleidon (1992, 1995), and Romer (1993). 3. Foster and Viswanathan (1990) present a related asymmetric information model that focuses on interday rather than intraday variations in price and volume. Although it can be argued that foreign exchange trading is not tied to any particular "day," our current focus is on behavior between the open and the close of usual trading hours for traders within the geographic markets of London and New York. Foster and Viswanathan examine the behavior of trading across days, particularly the effects of weekends on Monday trading.
44
David A. Hsieh and Allan W. Kleidon
section 2.1, namely, recent learning models based on asymmetric information and other market microstructure models that place more emphasis on the roles of market power and inventory management of marketmakers. Section 2.4 contains concluding remarks.
2.1
Data and Results
This section examines the behavior of spreads and quote volatilities across different regional foreign exchange markets at the same point in time. We demonstrate that a key implication of standard asymmetric information models is rejected in foreign exchange data, namely, that periods of high variance correspond to periods of high concentration of informed trading. When one regional market has high variance (i.e., open and close of the regional market with concurrent high bid-ask spreads), other markets simultaneously have low variance (and low bid-ask spreads), even though the traders from different markets are connected by computer terminals with all quotes appearing simultaneously on all terminals. In section 2.2, we interpret these results as showing that whatever the explanations for these phenomena-and we suggest possibilities in section 2.3 below-they are not consistent with current standard models of asymmetric information. 2.1.1
Data
The foreign exchange market can be roughly divided into two groups.4 The first group comprises marketmakers or the interbank market, which accounts for most foreign exchange trading. 5 Marketmakers deal with each other through a very active computerized market that trades virtually around the clock, either directly or through interdealer brokers. The second group comprises the retail market or customers who approach a local broker or bank and are offered retail foreign exchange quotes by that retail bank. The interbank foreign exchange market, from which our data are obtained, comprises a network of major trading banks throughout the world that are linked interactively via computer screens (either Reuters or Telerate systems). We use data from the Reuters indications system which transmits computerized quotes among interbank dealers. When a trading bank individually updates its quotes, the new quotes directly appear on the screens of all traders around the world. Actual trades are consummated via telephone,6 and price and 4. For an excellent description of the markets, see Bumhanl (1989). See also Goodhart (1990), Goodhart and Figliuoli (1991), Lyons (1992, 1993, 1995), and Bollerslev and Domowitz (1993). 5. Lyons (1993, 2), citing the New York Federal Reserve Bank, states that over 80 percent of trading volume is between marketmakers. 6. Lyons (1995) examines data for a single marketmaker from the Reuters Dealing 2000 System that allows screen trading, although direct telephone communication was necessary for all traders when our data were collected. Since our approach requires cross-market comparisons, the new data set is insufficient for our purposes. It will be useful to replicate our study using data from the new Reuters system, if sufficient data ultimately become available.
45
Bid-Ask Spreads in Foreign Exchange Markets
volume for direct interdealer trades are not publicly revealed. Some information about brokered interdealer trades, namely, price, quantity, and whether the trade is at the bid or the ask, is publicly disseminated to dealers via an intercom system. Major trading banks often perform both interbank and retail roles, with a dedicated foreign exchange desk within the bank linked to the interbank market and with retail customers who are offered quotes that consist of the dealers' interbank quotes plus an additional markup. The deutsche mark/dollar data that we use were originally captured from a Reuters data feed by Charles Goodhart and cover the eighty-two days from 9 April to 30 June 1989. These are the same data used by Bollerslev and Domowitz (1993), who provide valuable descriptions of the characteristics of these data.? For our purposes, we concentrate on two markets-London and New York-but our analysis applies to the other markets documented in Bollerslev and Domowitz. Figures 2.1a and 2.1b document the time periods in which significant trading activity takes place in London and New York, respectively. These figures give the average number of quote arrivals on Reuters' screens, per five-minute intervals, from traders based in London and New York. s Each location has activity beginning around 7:00 A.M. (local time) and lasting until about 6:00 P.M. (local time).9 Figure 2.2 integrates the London and New York data by converting the Eastern Standard Time (EST) New York times to Greenwich Mean Time (GMT) and plotting both figures 2.1a and figure 2.1b together. As noted by Bollerslev and Domowitz (1993, 1426), trading activity (as measured by the number of quote arrivals) in London begins high, declines until New York opens, then increases until the close of London trade. Activity in New York roughly follows that of London but continues strongly after the London close as New York becomes the major open market. We wish to highlight several aspects of these data that make them excellent for studying models of asymmetric information. First, the interbank market is the closest to an ideal twenty-four-hour market of which we know. It is very liquid, especially by comparison with stock markets, in terms of both volume of trade and number of traders; individual traders have continuous access to the market via computer terminal; and some trader is active virtually around the clock (including the markets in the Far East). Second, the commodity is essentially the same in all markets: the deutsche mark and the dollar are the same irrespective of trader location, and settlement issues are trivial by comparison with, say, transactions on the New York Stock Exchange (NYSE) ver7. The data were kindly forwarded to us by Tim Bollerslev and Ian Domowitz. For details of the data capture, error screens, and data characteristics, see Goodhart (1990) and Goodhart and Figliuoli (1991). 8. We exclude quotes from Saturday and Sunday since there is almost no trading on these days (except the last hour on Sunday). 9. There are isolated quotes at times outside this interval, but they are negligible in frequency, and we ignore them in subsequent analysis. Note also that in London there is one period shortly before 6:00 P.M. (GMT) in which there are no quotes; this is reflected in figures 2.4a and 2.6 below by a "zero" spread at that time.
46
David A. Hsieh and Allan W. Kleidon A
500 . - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - ,
400
en Q) (5 :J 300
o
o 'Q)
.c E
200
::J
Z 100
230
500
730
1000
1230
1500
1730
2000
2230
Time (In GMT Hours & Minutes) Fig. 2.ta
Number of quotes per five-minute interval, London
8
500 . - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - ,
400
en Q)
a (5
300
o 'Q)
.c
E
200
::J
Z 100
O~---'----'------....<::...J....----I.--_'_____
230
500
730
1000
1230
_
_ _ L_ _ __L!........::=....._..L____
1500
1730
Time (In EST Hours & Minutes) Fig. 2.tb
Number of quotes per five-minute interval, New York
2000
_ _ L_
2230
_..l
47
Bid-Ask Spreads in Foreign Exchange Markets 500 , . - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - ,
400
[L~~~on I
en Q)
ao
300
'6 ~
Q)
.c
E :::J Z
200
100
230
500
730
1000
1230
1500
1730
2000
2230
Time (In GMT Hours & Minutes) Fig.2.2
Number of quotes per five-minute interval, London and New York
sus those on the International Stock Exchange (ISE) in London (see Burnham 1989, 21). Third, the facts of twenty-four-hour trading and the very size and nature of the foreign exchange markets suggest that standard models of asymmetric information may find it difficult to explain persistent temporal patterns since there may be less systematic private information in these markets compared with, say, a small NYSE-listed stock that has few followers. Lyons (1995) concludes that interpretation of "information" in the foreign exchange market must be broader than that in standard models of equity markets since there are no "insiders" in the foreign exchange market. Our discussion of learning models in section 2.3 below provides such a broader concept of information. 2.1.2
Results
London and New York: Individual Markets
We first document that volatility in the foreign exchange markets follows the same U-shaped pattern from the open to the close of trade as on, say, the NYSE. This is important because it is precisely this result that supports the conclusions of Admati and Pfleiderer (1988) and Subrahmanyam (1989) that there is heavy activity by informed traders at the open and close of trade on the NYSE, which results in the higher variances of returns at those times. Return variances are calculated as follows. The day is first divided into oneminute intervals. At the end of each minute, the last quote (bid/ask) is averaged. If no new quotes occur during that minute, the observation is deleted. Between
48
David A. Hsieh and Allan W. Kleidon
two minutes (if both have quotes), the one-minute rate of return is computed as the discrete rate of change of the average bid/ask between the two minutes. The standard deviation of each half hour (beginning at 7:30 A.M. local time) is computed as the standard deviation of these one-minute returns during the interval. For robustness, we present the medians of these half hourly standard deviations over all days in our sample. Figures 2.3a and 2.3b plot these average (median) standard deviations per half hour interval, from 7:30 A.M. (local time) to 6:00 P.M. for London and New York, respectively. The results are striking. The average variances are much higher at open and close in both markets than during other times of the trading day, confirming the apparent U shape in volatility that has been previously documented in other markets. 10 Figures 2.4a and 2.4b present the average spreads (in pfennig per dollar) by minute over the trading periods indicated by trading activity in figures 2.1a and 2.1b above for London and New York, respectively. These figures confirm the general U-shaped pattern of spreads particularly documented in Bollerslev and Domowitz for smaller regional banks (1993, 1428ff.). London and New York: Integrated Markets
Figure 2.5 shows the standard deviations of both London and New York returns on the same (GMT) time scale. There appears to be no correspondence between the striking volatility patterns across these two markets, which are virtually instantaneously linked in terms of quote information. Not surprisingly given the results in figure 2.5, figure 2.6 shows that changes in the bid-ask spread documented for London and New York separately do not provide any coherence when viewed at the same time across markets. While the average spread is roughly equal in London and New York when both markets are open, there is no apparent effect of the high spreads associated with the open or close of one market on the other market. Table 2.1 presents a test of the difference in spreads between London and New York, by fifteen-minute intervals from noon GMT to 5:30 P.M. GMT. The test assumes that samples are uncorrelated, with the result that the t-statistic for the difference of average spreads in each interval is downward biased if there is in fact positive correlation across the samples (which would be expected if information affected both sets of quotes throughout the day). The results confirm the impressions from figure 2.6 above and demonstrate that the (indicated) spreads in New York are consistently significantly higher than those in London, except for London close, when London spreads are higher 10. Note that, since we delete observations if no quote update occurs, and since we require consecutive observations to calculate a one-minute rate of return, there will be typically fewer observations for any given time interval than one observation per minute times the number of days in the sample. In particular, there are fewer observations in our sample at the open and close of trading than during periods of active trading. Table 2.3 below presents formal tests of the differences in variances.
49
Bid-Ask Spreads in Foreign Exchange Markets
0.20
Q)
~----------------------------,
0.15
~
::l
c:
~ ~
If. 0.10 ~
c:
Q)
e Q)
a..
0.05
400
800
1200
1600
2000
2400
Time (in GMT Hours & Minutes) Fig.2.3a
0.20
Q)
Standard deviation per thirty-minute interval, London
~----------------------------,
0.15
~
::l
c:
~ ~
~
0.10
~
C
Q)
~
Q)
a.
0.05
400
800
1200
1600
2000
Time (in EST Hours & Minutes) Fig.2.3b
Standard deviation per thirty-minute interval, New York
2400
50
David A. Hsieh and Allan W. Kleidon
0.0012
~---------------------------,
0.0010
0.0008
"0
co
~ 0.0006 C-
OO
0.0004
0.0002
230
500
730
1000
1230
1500
1730
2000
2230
Time (In GMT Hours & Minutes) Fig. 2.4a Distribution of the spread over the day, London
0.0012
~----------------------------,
0.0010
0.0008
"0
co
~ 0.0006 C-
oo
0.0004
0.0002
0.0000 "--'-----'------'---..............- -'------"-----'-------'--""-----'----'--------' 230
500
730
1000
1230
1500
1730
Time (In EST Hours & Minutes) Fig. 2.4b
Distribution of the spread over the day, New York
2000
2230
51
Bid-Ask Spreads in Foreign Exchange Markets 0.20 r - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - .
London New York Q)
0.15
+""
::J
c:
~ '-
cf 0.10 +""
c:
Q)
u '-
Q)
Q. 0.05
400
800
1200
1600
2000
2400
Time (in GMT Hours & Minutes) Fig. 2.5 Standard deviation per thirty-minute interval, London and New York
0.0012 r - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - ,
0.0010
0.0008
"0 CO ~ 0.0006 0-
m
0.0004
0.0002
230
500
730
1000
1230
1500
1730
2000
Time (In GMT Hours & Minutes) Fig. 2.6 Distribution of the spread over the day, London and New York
2230
than earlier in the London trading day and on average are higher than those in New York. Table 2.2 tests for the difference between quote midpoints in London and New York, as opposed to the spreads. Although table 2.1 above shows that the indicated spreads in New York are significantly higher than those in London except at London close, table 2.2 shows that this pattern does not carry over to mid-quotes. Although on average the London mid-quotes are greater than the New York mid-quotes, the difference is typically not statistically significant, at least assuming uncorrelated samples. Table 2.3 examines the difference between mid-quote to mid-quote variances for London and New York, again by fifteen-minute intervals beginning at noon GMT. The results from figure 2.5 above are supported in table 2.3. From noon until 3:15 P.M., the variance in New York consistently exceeds that in London with the largest average variances in New York at the start of New York trading (although not all periods are individually significant at conven-
tionallevels based on the conservative assumption of uncorrelated samples). 11 However, the variance in London increases toward the close of London trading and significantly exceeds that in New York in the later part of the London trading period. The (conservative) t-statistics are strongly significant at conventional levels in all periods between 4:00 P.M. and 5:30 P.M. (GMT) (the t-statistics range from 2.64 to 5.53). Even in the last fifteen-minute interval (to 5:45 P.M.) in which there were only twelve observations in London, the t-statistic is 1.73. Thus, these results clearly document a change in variance in one market that is not simultaneously observed in the other market. The cross-market variance results in figure 2.5 and table 2.3 and the crossmarket spread results in figure 2.6 and table 2.1 constitute a challenge for standard asymmetric information models as applied to foreign exchange data. 11. The earliest intervals in New York (and the latest intervals in London) also have few observations (see n. 10 above).
54
David A. Hsieh and Allan W. Kleidon
Table 2.3
Test of Difference in Mid-Quote to Mid-Quote Variance, London and New York (by fifteen-minute interval) London
New York
Time (GMT)
SDa
Nb
SDa
Nb
t-Stat. c
Noon 12:15 12:30 12:45 1:00 1:15 1:30 1:45 2:00 2:15 2:30 2:45 3:00 3:15 3:30 3:45 4:00 4:15 4:30 4:45 5:00 5:15 5:30
.0253 .0203 .0206 .0174 .0241 .0276 .0477 .0310 .0330 .0329 .0293 .0301 .0321 .0333 .0317 .0336 .0372 .0375 .0437 .0451 .0453 .0652 .0853
873 905 836 867 958 964 1,059 1,168 1,091 1,153 1,118 1,105 1,197 1,207 1,184 1,128 811 642 477 343 166 46 12
.0442 .0905 .0445 .0280 .0301 .0327 .0727 .0429 .0397 .0408 .0348 .0315 .0385 .0374 .0309 .0308 .0278 .0284 .0276 .0266 .0261 .0258 .0265
2 30 202 253 367 482 470 549 524 596 620 699 712 783 833 890 1,042 1,065 1,104 1,151 1,183 1,124 1,088
-1.40 -1.27 -2.56 -3.72 -1.56 -1.50 -1.62 -2.11 -1.52 -2.19 -2.41 -.63 -1.57 -1.42 .53 1.35 4.90 4.01 4.46 5.53 5.21 2.64 1.73
aStandard deviation of mid-quote changes for London (SDLON ) and New York (SDNy ) times 100. bNumber of observations. Ct-statistic based on method of moments test for equality of the variances, assuming uncorrelated samples.
2.2
Current Asymmetric Information Models
This section examines the standard asymmetric information literature as applied to the open and close of trade in foreign exchange markets. The general importance of asymmetric information has long been recognized. Bagehot (1971) argues that the marketmaker loses in trades with better-informed traders, with the result that trades with uninformed liquidity traders must make sufficient profit to cover those losses plus costs. This notion is formalized in subsequent work (see Admati 1991), the most relevant for our current purposes being Admati and Pfleiderer (1988) and Subrahmanyam (1989, 1991). Bollerslev and Domowitz (1993) explicitly interpret much of the foreign exchange behavior that we discuss in terms of the models of Admati and Pfleiderer and Subrahmanyam. 12 12. Lyons (1993, 1995) develops microstructure models in the context of the foreign exchange market but does not examine cross-market data as in Bollerslev and Domowitz (1993).
55
Bid-Ask Spreads in Foreign Exchange Markets
An initial problem with this literature is an inability of the theoretical model of Admati and Pfleiderer (1988), and the extension of this model by Subrahmanyam (1989, 1991), simultaneously to account for the observed empirical phenomena of volume, volatility, and bid-ask spreads on the NYSE (the market that these models were originally intended to explain). 13 We then demonstrate that the cross-market foreign exchange results from section 2.1 above are inconsistent with standard asymmetric information models. 2.2.1
Admati and Pfleiderer (1988)
A model of endogenous trading volume is provided by Admati and Pfleiderer (1988), who extend Kyle (1984). They assume three types of agents: informed traders, who will trade only on terms advantageous to them given their superior information; discretionary liquidity traders, who must trade over a given day but who choose when to trade during the day on the basis of trading costs (i.e., they trade in those periods of lowest cost); and nondiscretionary liquidity traders, who must trade at a given time during the day regardless of cost. In this model, trading costs arise solely because of the activity of the informed, whose profits are paid by the uninformed liquidity traders. Given their assumptions, Admati and Pfleiderer show that it is possible to obtain concentrations of volume at arbitrary trading times because in equilibrium these high volume periods attract both informed traders and discretionary liquidity traders. The informed are attracted because there will be more uninformed liquidity traders behind whom they can camouflage their trades. The discretionary liquidity traders are attracted because, in this model, the increased activity of informed traders implies sufficiently increased competition among them that the cost of trading to the uninformed is lowered relative to other periods. Admati and Pfleiderer relate their results to observed empirical behavior, especially to volume and variance at the open and close of a day's trading on the NYSE, and show "that the patterns that have been observed empirically can be explained in terms of the optimizing decisions of these traders" (1988, 4). Their primary motivation is the high volume and concurrent high variance at open and close. Volume is explained by their concentration equilibrium outlined above; high variance follows directly from the increased activity by informed traders at open and close since more (previously private) information is thus incorporated into prices. 2.2.2
Subrahmanyam (1989, 1991)
The key result in Admati and Pfleiderer (1988)-namely, that increased activity by informed traders lowers the costs to the uninformed, who must pay the price of the presence of the informed-is not intuitively obvious. Subrahmanyam (1989,1991) builds on the model of Admati and Pfleiderer and shows
13. Much of this discussion follows Brock and Kleidon (1992, sec. 4.3).
56
David A. Hsieh and Allan W. Kleidon
that their result depends on the assumption of risk-neutral informed traders. If the informed traders are risk averse, then increased activity on their part can increase the trading costs of liquidity traders. Subrahmanyam (1989, 18) cites Foster and Viswanathan (1989) as showing that the adverse selection component of bid-ask spreads is highest at the beginning of the day, which "contrasts with the model of Admati and Pfleiderer [1988], which predicts that spreads should be lowest at the beginning of the day." Brock and Kleidon (1992, sec. 4.1) also provide evidence that appears inconsistent with the key result in Admati and Pfleiderer since spreads follow the same U shape as volume: highest volume is associated with highest, not lowest, costS. 14 Subrahmanyam (1989) interprets this result as consistent with his extension of Admati and Pfleiderer to the case of risk-averse informed traders since then more trading by informed traders results in lower market liquidity and higher costs. To do this, he requires the additional assumption that "more individuals are informed at the beginning of the day than at other times during the day" (p. 17). For this model to be a full explanation of the relation between spreads and volume, he presumably requires that the informed also trade more heavily at close. 2.2.3
Bollerslev and Domowitz (1993)
Bollerslev and Domowitz (1993) examine the same data as in this paper, namely, continuously recorded quotes on the deutsche mark/dollar exchange rate in the interbank foreign exchange market. They document quote arrivals and bid-ask spreads over the trading day, across geographic locations, and across trading participants. The analytic focus is on two main areas: first (which is most relevant for this paper), an evaluation of the predictions of the asymmetric information models of Admati and Pfleiderer (1988) and Subrahmanyam (1989) and, second, time-series modeling of means and conditional variances. Bollerslev and Domowitz conclude that their evidence "is encouraging with respect to the ability to validate and discriminate between theoretical models of trading activity using intraday information on foreign exchange trading" (1993, 1439). They suggest that periodic nondiscretionary liquidity trading around open and close will intensify the results of Admati and Pfleiderer, while round-the-clock trading will weaken them, and conclude that the U-shaped patterns of trading activity from open through close, well documented for the 14. The model of Admati and Pfleiderer (1988) and Subrahmanyam (1989, 1991) assumes sequential batch auctions rather than the continuous auctions associated with bids and asks on the NYSE. However, Admati and Pfleiderer regard their results as applying to the volume behavior on the NYSE, and Subrahmanyam (1989) explicitly equates the costs in Admati and Pfleiderer (which are the same as in his model) with bid-ask spreads. We follow this approach. Further, as Grossman and Miller (1988, 628) point out, transactions costs should be measured by the difference between the price paid now and the price expected to be paid by waiting; but, if the average spread falls after opening and rises at close, one would expect a priori that a given liquidity trader would expect higher transactions costs in such periods.
57
Bid-Ask Spreads in Foreign Exchange Markets
NYSE and addressed by Admati and Pfleiderer (1988), is apparent only in the European markets (p. 1426, n. 8). Bollerslev and Domowitz also find that, for traders who restrict their trading to regional markets within well-defined openings and closings (as opposed to international firms with traders in multiple regional markets), the activity pattern (quote volume) "typically shows a U-shape, as does the distribution of their own spreads over the course of the day" (1993, 1428). Since these traders "operate more like the stock market traders usually modeled in much of the theoretical literature, with behavior influenced by openings and closings" (p. 1429), Bollerslev and Domowitz interpret this evidence as confirming the model of Admati and Pfleiderer (1988), although they note that the "daily patterns of the spread and market activity suggest risk-averse behavior on the part of these traders" (p. 1439). This modification is tied to "Subrahmanyam's (1989) extension of the Admati and Pfleiderer model to include risk-averse behavior [that] predicts that more trading by informed risk-averse participants brings about higher costs" (p. 1426, n. 9). 2.2.4
Foreign Exchange Quote Data and Standard Information Models
We conclude that the asymmetric information models of Admati and Pfleiderer and Subrahmanyam are not consistent with the foreign exchange data on spreads and volatility, for two reasons. First, close examination of Subrahmanyam's extension of the Admati-Pfleiderer model shows that, although he can account for high spreads at times of high informed trading, the cost is that he loses the major result of the Admati-Pfleiderer model, namely, the concentrated trading equilibrium relied on by Admati and Pfleiderer to account for simultaneous high volume and high volatility. Second, the results from section 2.1 above show that volatilities (and spreads) across markets that trade simultaneously do not shew the congruence implied by either the AdmatiPfleiderer or the Subrahmanyam models. On the first point, note that Subrahmanyam's model implies that discretionary liquidity traders who can time their trades will avoid the high-cost, highvolume periods that he links to high levels of informed trading. However, this breaks the concentration of trading relied on in Admati and Pfleiderer's equilibrium. This in turn makes it difficult to explain the observed high volume in terms of discretionary liquidity traders and informed traders, since the former will avoid the high-cost open and close and the number of informed traders must be "sufficiently small" (Subrahmanyam 1989, p. 18) for the result to go through. Further, were the increased volume due to a very large increase in the number of informed traders (sufficient both to offset the departure of discretionary liquidity traders and to account for the total increase in volume), one wonders who takes the other side of their trades, especially since they receive correlated signals in this model. Presumably, the burden falls, once again, on the luckless nondiscretionary liquidity traders, who in these models have zero elasticity of demand and must
58
David A. Hsieh and Allan W. Kleidon
trade at these times regardless of price or cost. That is, the empirical results on spreads and volume-which the Admati-Pfleiderer and Subrahmanyam models attempt to explain-imply that within those models there must be an increase in nondiscretionary liquidity trading at open and close sufficient to offset the departure of discretionary traders in the face of higher transactions costs. Admati and Pfleiderer (1988, 34) conjecture that the orders of nondiscretionary traders may cluster around open and close because of market closure; however, this is not part of their formal model, and they rely on such periodic demand as simply a timing catalyst for their endogenous clustering that requires low trading costs at open and close to attract discretionary traders. The second problem with these asymmetric information models relates to the cross-market results from section 2.1 above. These results show that the observed behavior in spreads and variances cannot be explained within standard information models. It is true that, looking at the two markets individually, the variance results appear similar to those from the NYSE used to motivate the asymmetric information model of Admati and Pfleiderer. At first blush, then, these results may appear to provide confirmation of the conclusion that activity at the open and close of trading in foreign exchange markets is heavily influenced by concentrations of informed traders at those times, resulting in high variances. The results obtained from looking at London and New York separately are highly misleading, however, in terms of evidence concerning any tracks left in the data by privately informed traders. Recall that these quotes appear directly on the Reuters' screens of traders in all locations. Assume that the high variance (and high spread) at close of trade in London is indeed caused by an unusually high concentration of informed traders at that time, which in turn causes rapid changes in quotes and, consequently, high variance. Traders in New York observe directly and simultaneously these London quote revisions that are ostensibly caused by the incorporation of previously private information. Since the commodity is the same whether the quotes are posted in London or New York, it must be the case that the incorporation of new information into the London quotes must virtually simultaneously show up in New York quotes, resulting in simultaneous high variance in New York quotes. The results in table 2.3 and figure 2.5 above clearly refute this implication. Note that the New York opening-which is ostensibly replete with private information, causing the New York variance of returns to rise dramaticallycauses scarcely a ripple on the London market! Similarly, the London market closes with dramatic local effects in terms of variance but with no effect on New York. Similarly, table 2.1 and figure 2.6 above show an apparent lack of integration across these two markets with respect to spreads.
2.3
Evaluation of Results
We regard the results from section 2.2 as striking evidence that, whatever is causing the patterns in variances and spreads at the open and close of trading
59
Bid-Ask Spreads in Foreign Exchange Markets
in the foreign exchange market, it is not the incorporation of previously private information through the heavy trading of informed traders at the open and close of trading in London and New York. This is particularly evident from the comparison of volatility patterns across London and New York in figure 2.5 and table 2.3 above. We consider two main alternative explanations for these results. First, we consider a broader class of information models that has recently been proposed to explain related forms of apparent excess volatility in stock prices, namely, models that relax the assumption in standard information models that traders have perfect knowledge about the preferences and beliefs of other traders in the market. These new models, based on imperfect information aggregation and subsequent learning by market participants, can be linked to experimental results in the behavioral literature and have proved successful in accounting for such difficult phenomena as the crash of October 1987. The second explanation that we consider, which is particularly relevant for behavior at the close of a market, returns to inventory models such as Garman (1976). 2.3.1
Models of Imperfect Information Aggregation: The Open of Trade
Our results at the open of New York and the close of London trading show that no new information is reaching the international foreign exchange market as a whole since there is an absence of unusual volatility in quotes generated by traders in London when New York opens and, conversely, at London close. One way to view these results is that quotes generated by traders in the volatile market show excess volatility relative to that implied by standard information models. Others have noted the difficulty in explaining foreign exchange data in terms of standard information models. For example, Frankel and Froot (1990) emphasize heterogeneity of expectations across traders in foreign exchange markets and suggest three possible implications of the high trading volume in these markets for price movements: (1) greater depth means more efficient processing of fundamental information; (2) there is no relation between trading volume and prices since the market is "already perfectly efficient"; and (3) there may be "excessive volatility" caused by trading based on "noise" rather than "news" (p. 182). Frankel and Rose (1994) explore the third possibility in some detail, with emphasis on the possibility of "endogenous speculative bubbles" in foreign exchange markets. The special nature of our cross-market data implies a high hurdle for potential explanations since such an explanation must account for both the systematic behavior within markets and the lack of congruence across markets around open and close. Thus, for example, if the explanation is to be "noise" rather than "news," it appears that London is not affected by noise when New York opens with high volatility; but at London close, the roles are reversed, with London displaying high volatility but New York now immune from any noise affecting London. This appears to be more complicated than is implied by the noise-trading models cited in Frankel and Rose (1994) or by models of learn-
60
David A. Hsieh and Allan W. Kleidon
ing currently applied to foreign exchange markets such as Lewis (1989a, 1989b). One recent approach that shows promise in this context focuses on imperfect, although rational, information aggregation. 15 There are two extreme views on price formation. One assumes that all private information is instantly aggregated and revealed to market participants through prices. The other assumes that price changes are capricious and irrational. One attempt to find a middle ground assumes that prices are formed as some average of these two types of behavior, an approach similar to the noise-trading models cited in Frankel and Rose (1994, 37). The approach of imperfect information aggregation also supplies a middle ground between the extremes, but in a very different way. It is well known from behavioral laboratory experiments that, under certain conditions, asset prices can be readily generated that display systematic deviations from those predicted by fully revealing rational expectations models. These deviations display both apparent excess volatility and the characteristics often associated with speculative bubbles. The conditions needed to generate these phenomena in experimental laboratories are revealing: market participants must lack common knowledge about other traders' preferences or beliefs, or there must be insufficient traded instruments to theoretically allow traders to invert from prices to infer information. Significantly, if traders have common knowledge about preferences and beliefs, then the consistent result is that prices quickly converge to the fully revealing rational expectations equilibrium if there are sufficient traded assets given the sources of information uncertainty and if the traders have trading experience in the market. 16 While standard information models typically assume common information among market participants about traders' preferences and beliefs, we regard this assumption as unrealistic in many settings, including the foreign exchange market. When traders first begin trading at the open of their local market, a commonly described issue is that they need to get the "feel" of the market at that time. I7 The most important elements. of this "feel" are the participants in the market at that point in time and what their trading behavior has been in the immediately prior trading period. Clearly, current prices are not sufficient statistics for these items. Traders attempt to obtain this information by contacting traders who have been trading for some time; for example, New York traders will have contacts in London (not necessarily in the same bank) whom they will call to obtain a sense of the current market. Our interpretation of this phenomenon is consistent with the importance of knowledge about the preferences and beliefs of other traders to allow information aggregation across traders. In the more limited view of an
15. For a detailed discussion of this approach, see Kleidon (1995). This general approach has been examined in the context of foreign exchange markets by Cmja (1993). 16. For details, see Kleidon (1995, sec. 2). 17. This is not restricted to the foreign exchange market. This phenomenon has been consistently described to us in conversations with both equity and foreign exchange traders.
61
Bid-Ask Spreads in Foreign Exchange Markets
individual trader, the question is the interpretation of current trading behavior in the market. For example, is a request for a bid on a certain sized trade likely to be followed by the same trader for another? If so, then that will affect the terms offered for the first trade. Significantly, traders report that, until they have got the feel of the market, they are uncertain of their "view" and hence, for example, of whether they will be going long or short one or another currency in their early trading during the day. This translates into initial high spreads, with rapidly changing quotes as traders develop their view for the current trading period. This, of course, appears to correspond to the evidence that we presented in section 2.1 above at the open of trade. The market that has been trading for some time (London, at New York open) does not show any "excess volatility" relative to surrounding times, and the new traders (in New York) rapidly get the feel of the market, that is, assimilate important information about the current market structure that is not contained in the current price. While there is little documented evidence about the mechanisms that traders use to get the feel of the market, conversations with traders indicate that they include perusal of overnight information from various sources beyond the trading history of the particular currency being traded, as well as more formal arrangements with counterparts at overseas banks who agree to provide specific information about the market to the "novice" who is just starting to trade. One possibility consistent with our data (and not addressed by standard information models) is that the new dealers set wide spreads while they gather this information and then narrow the spreads once they are ready to trade. However formal or informal the mechanism for obtaining information about the market structure, the level of spreads and the volatility of quotes settle down once traders get the feel of the market, and the markets trade in an integrated fashion until the next major disruption-the close of trading on one of the markets. 2.3.2
Inventory-Based Models: Close of Trade
A large proportion of foreign exchange trading, including most of that done by smaller, regional banks, is "day trading," in which traders start and end the day with flat positions. While this does not cause any particular problems at the start of the day since positions can be accumulated during the trading day that has just begun, the inventory problem for day traders becomes increasingly acute as the close of trading approaches. In the limit, a trader who must close out a position by the end of the day has increasingly inelastic demand to trade and will be more willing to accept a relatively poor price to accomplish the trade. 18 18. Brock and Kleidon (1992) exploit differences in trading demand at open and close to predict higher spreads at these times if a marketmaker such as a monopolist specialist has the ability to price discriminate at these times of inelastic demand to trade. This idea may have some application in foreign exchange markets, particularly with respect to quotes by regional banks to customers who have peak foreign exchange trading demands at open and close.
62
David A. Hsieh and Allan W. Kleidon
The effects of inventory on prices and quotes are well discussed in market microstructure literature beginning with Garman (1976) (see also Amihud and Mendelson 1980; Ho and Stoll 1981, 1983; O'Hara and Oldfield 1986; and Son 1991). Recent empirical work in equity dealer markets in London (see Hansch, Naik, and Viswanathan 1993) and the United States (see Chan, Christie, and Schultz 1995) documents strong inventory effects, and Lyons (1995) documents inventory effects in the foreign exchange market. Discussions with
tflldets iti the foreign exchange luarket eonfirttl that high quote volatility and spreads at close are linked by traders to the activities of day traders who are attempting to close out their positions. 19
2.4
Conclusions
In this paper, we have explored the implications of foreign exchange markets for alternate models of intraday price and volume behavior. There are empirical difficulties in reconciling current asymmetric information models with stock price data in individual markets, but it is possible for some form of these models to be consistent with NYSE prices and volume, assuming that liquidity traders have strong demand to trade at open and close for reasons that are not motivated within the information models. Our choice of the foreign exchange market is motivated by the ability to test whether any form of the current asymmetric information models can explain prices and volumes in this market. The advantages of foreign exchange data are that the market extends around the clock and that traders from any location have equal access via computer terminal to the posted quotes of all traders from all locations. If new information is the cause for revisions in quotes for traders in, say, London, then those quote revisions are immediately indicated to all other traders. Moreover, since there is an overlap between the trading days of traders in London and New York, we can observe whether the quote behavior is integrated in the fashion implied by the asymmetrical information models. The results are inconsistent with these models. If London and New York are examined individually, they display the same patterns of spread and volatility from "open" to "close" as does the NYSE, which may indicate similar forces in each market. However, when New York foreign exchange traders begin their day, with attendant high volatility and spreads, London has been trading for hours and is still trading. There is no effect on London quotes, in either volatility or spreads, of the striking New York quote behavior. Similarly, although London "closes" its day with high variances and high spreads, this does not cause a ripple on the quotes of New York traders. This is inconsistent with standard sources of asymmetric information being the fundamental cause of this behavior. 19. A related feature of the New York close is the decrease in depth of the market as traders stop trading, with a consequent increase in spreads.
63
Bid-Ask Spreads in Foreign Exchange Markets
Our primary conclusion is that the current models of asymmetric information appear inadequate to explain our results. This raises questions concerning their application in other markets as well. Indeed, our findings are not limited to quotes in the foreign exchange market. Kleidon and Werner (in press) demonstrate that similar results apply to U.K. stocks that are cross-listed on London and New York exchanges and that consequently trade simultaneously in a similar fashion to the foreign exchange trading that we document here. It appears that we must turn to explanations other than current information models if we are to account for the behavior of prices and volumes in both stock and foreign exchange markets. We suggest two possible explanations for the empirical results documented in section 2.1. First, we place importance at the open of trade on learning by traders about the structure of the market, particularly concerning the identity and behavior of other traders. This can be linked both to behavioral experimental results and to the results in recent literature that examines the implications for information aggregation of a lack of common knowledge among traders concerning the preferences and beliefs of market participants. During the period when traders pursue various mechanisms to learn the feel of the market, spreads may be set sufficiently wide to avoid trading at an informational disadvantage. Second, at the close of trade, much of the activity in the closing market documented in section 2.1 may be due to the inventory-related activities of traders at market close, especially for the large group of day traders who must close their positions by the end of trading. While these explanations seem plausible-and indeed are confirmed by foreign currency traders-further investigation is warranted.
References Admati, A. R. 1991. The informational role of prices: A review essay. Journal ofMonetary Economics 28: 347-61. Admati, A. R., and P. Pfleiderer. 1988. A theory of intraday patterns: Volume and price variability. Review of Financial Studies 1:3-40. Amihud, Y., and H. Mendelson. 1980. Dealership market: Market-making with inventory. Journal of Financial Economics 8:31-53. Bagehot, W. (pseud.). 1971. The only game in town. FinancialAnalysts Journal 27: 1214,22. Barclay, M. J., R. H. Litzenberger, and J. B. Warner. 1990. Private information, trading volume, and stock-return variances. Review of Financial Studies 3:233-53. Bollerslev, T., and 1. Domowitz. 1993. Trading patterns and prices in the interbank foreign exchange market. Journal of Finance 48: 1421-43. Brock, W. A., and A. W. Kleidon. 1992. Periodic market closure and trading volume: A model of intraday bids and asks. Journal of Economic Dynamics and Control 16:451-89.
64
David A. Hsieh and Allan W. Kleidon
Burnham, J. B. 1989. Marketmakers, speculators and customers: The U.S. foreign exchange market. Working Paper no. 130. St. Louis: Center for the Study of American Business, Washington University. Chan, K. C., W. G. Christie, and P. H. Schultz. 1995. Market structure and the intraday pattern of bid-ask spreads for Nasdaq securities. Journal of Business 68:35-60. Crnja, Z. 1993. Environmental uncertainty in the foreign exchange market. Honors thesis, Department of Economics, Stanford University. Foster, F. D., and S. Viswanathan. 1989. Variations in volumes, spreads and variances. Research Paper no. 88-108. Fuqua School of Business, Duke University. --.- . 1990. A theory of the interday variations in volume, variance, and trading costs in securities markets. Review of Financial Studies 3:593-624. Frankel, J. A., and K. A. Froot. 1990. The rationality of the foreign exchange rate: Chartists, fundamentalists, and trading in the foreign exchange market. American Economic Review 80: 181-85. Frankel, 1. A., and A. K. Rose. 1994. An empirical characterization of nominal exchange rates. In The handbook of international economics, ed. G. Grossman and K. Rogoff. Amsterdam: North-Holland. Freedman, R. 1989. A theory of the impact of international cross-listing. Graduate School of Business, Stanford University. Typescript. Garman, M. B. 1976. Market microstructure. Journal of Financial Economics 3:257-75. Gennotte, G., and H. Leland. 1990. Market liquidity, hedging and crashes. American Economic Review 80:999-1021. Goodhart, C. 1990. "News" and the foreign exchange market. Discussion Paper no. 71. Financial Markets Group, London School of Economics. Goodhart, C., and L. Figliuoli. 1991. Every minute counts in financial markets. Journal of International Money and Finance 10:23-52. Grossman, S. 1., and M. H. Miller. 1988. Liquidity and market structure. Journal of Finance 43:614-33. Hansch, 0., N. Naik, and S. Viswanathan. 1993. Trading profits, inventory control and market share in a competitive dealership market. Fuqua School of Business, Duke University. Typescript. Ho, T., and H. R. Stoll. 1981. Optimal dealer pricing under transactions and return uncertainty. Journal of Financial Economics 9:47-73. -~-. 1983. The dynamics of dealer markets under competition. Journal of Finance 38: 1053-74. Jacklin, C., A. W. Kleidon, and P. Pfleiderer. 1992. Underestimation of portfolio insurance and the crash of October 1987. Review of Financial Studies 5:35-63. Kleidon, A. W. 1992. Market and environmental uncertainty. In The new Palgrave dictionary of money and finance, vol. 2, ed. P. Newman, M. Milgate, and 1. Eatwell. London: Macmillan. - - - . 1995. Stock market crashes. In Finance handbook, ed. K. Jarrow, V. Maksimovic, and W. T. Ziemba. Amsterdam: North-Holland. Kleidon, A. W., and 1. W. Werner. In press. U.K. and U.S. trading of British cross-listed stocks: An intraday analysis of market integration. Review of Financial Studies. Kyle, A. S. 1984. Market structure, information, futures markets, and price formation. In International agricultural trade: Advanced readings in price formation, market structure, and price instability, ed. G. G. Storey, A. Schmitz, and A. H. Sarris. Boulder, Colo.: Westview. Lewis, K. K. 1989a. Can learning affect exchange-rate behavior? The case of the dollar in the early 1980s. Journal ofMonetary Economics 23:79-100. - - - . 1989b. Changing beliefs and systematic rational forecast errors with evidence from foreign exchange. American Economic Review 79:621-36.
65
Bid-Ask Spreads in Foreign Exchange Markets
Lyons, R. K. 1992. Private beliefs and information externalities in the foreign exchange market. Columbia University, March. Typescript. - - - . 1993. Optimal transparency in a dealership market with an application to foreign exchange. University of California, Berkeley, September. Typescript. - - - . 1995. Tests of microstructural hypotheses in the foreign exchange market. Journal of Financial Economics 39:321-51. O'Hara, M., and G. S. Oldfield. 1986. The microeconomics of market making. Journal of Financial and Quantitative Analysis 21:361-76. Romer, D. 1993. Rational asset-price movements without news. American Economic Review 83: 1112-30. Son, G. 1991. Dealer inventory position and intraday patterns of price volatility, bid! ask spreads and trading volume. Department of Finance, University of Washington. Typescript. Subrahmanyam, A. 1989. Risk aversion, market liquidity, and price efficiency. Anderson Graduate School of Management, University of California, Los Angeles. Typescript. - - - . 1991. Risk aversion, market liquidity, and price efficiency. Review ofFinancial Studies 4:417-41.
Comment
Zhaohui Chen
In this paper, Hsieh and Kleidon begin with a precise description of the state of the "market microstructure literature." They remind the readers of the original definition of market microstructure by Mark Garman (1976) and summarize the current dominating paradigm in the literature-the asymmetric information models in the line of Admati (1991), Admati and Pfleiderer (1988), and others. This introduction is useful to exchange rate researchers attempting to exploit microstructure models. It serves as a warning that, given its narrow focus, we should not expect this literature to give us answers to all the macroeconomic puzzles related to foreign exchange rates. On the bright side, this paper has made a welcome attempt to expand the current literature with alternative paradigms. A broader intellectual base may well be what we need to understand the behavior of the foreign exchange market. The empirical evidence amassed in the paper clearly indicates that the opening and closing of New York and London markets seem to be local events. The associated trading activity and volatility do not transmit from one location to the other, even though the two markets are open to trade at the same time. This evidence could be interpreted in two different ways. First, the asymmetric information model may be wrong, in the sense that private information associated with high volatility in one market is not picked up by the other market. Given the simultaneity of information processing on the Reuters screen in the two markets, and given that currencies are the most standardized vehicle of trade, this interpretation does not stand well. This leads to the second interpreZhaohui Chen is the Jean Monnet Lecturer in the Economics of the European Union at the London School of Economics and a research affiliate of the Centre for Economic Policy Research.
66
David A. Hsieh and Allan W. Kleidon
tation: the unusual market activities around the opening and closing of the markets are not driven by private information. Under this interpretation, the evidence should not be seen as a rejection of the asymmetric information model since the events used are not information driven and therefore cannot be used to test the information transmission story behind the Admati and pfleiderer-type models. However, the evidence and the institutional analysis in the paper unambiguously show that the conventional asymmetric model is not applicable to the problems of this paper. This is an important message since it proves, by counterexamples, that the currently dominating paradigm in the microstructure literature is inadequate to cover a possibly broad array of issues in the foreign exchange market. The second half of the paper explores different paradigms, including the traditional inventory models and the new models of learning. At this stage, most of the learning models are based on experimental results. Nevertheless, it would be extremely interesting if the authors, or future researchers, could apply 'their econometric tools to test the applicability of such theories and to provide evidence that can help validate different assumptions behind the new theories. Following the discussion of the learning models in the paper, I would like to raise two questions. One is what the traders try to learn in the process; the other is whether there is a welfare gain from learning. Regarding the first question, the paper mentioned specific parameters such as preferences and beliefs and the ambiguous notion of the "feel" of the market. One possible explanation is that traders need to learn each other's trading rule, or asset valuation scheme, and even the theories behind them, as they do not want to be the outliers in the market. They can do that by testing each other's reaction to various quotes, hence the large number of quotes and high volatility from the screen data around the opening of the market. Regarding the second question, my hypothesis is that a pricing rule formed through consultation with other market participants is more likely to be correct (in absolute terms) than a simple average of other people's quotes (the whole is larger than the sum of the parts). This point is illustrated plainly (or mysteriously) in the popular management experimental game known as "Desert Survival Game," where a group of people are asked to form a survival strategy after an imaginary plane crash in a desert. The group decision following a full consultation among the group members is almost always better than the simple average of individual opinions before the consultation. Finally, the use of the Reuters quote data in this paper is open to question in light of the recent study by Goodhart, Ito, and Payne (chap. 4 in this volume), in which they find no linkage between these quotes and real trading activities as captured by a more up-to-date data screen. Given the dismissal of the private information story, the problem is less damaging to this paper because, as the later part of the paper suggests, there may be different reasons leading to the
67
Bid-Ask Spreads in Foreign Exchange Markets
use of quotes, such as learning and testing ("feeling") the market, that are not necessarily linked to actual trading activities. References Admati, A. R. 1991. The information role of prices: A review essay. Journal of Monetary Economics 28:347-61. Admati, A. R., and P. Pfleiderer. 1988. A theory of intraday patterns: Volumes and price volatility. Review of Financial Studies 1:3-40. Garman, M. B. 1976. Market microstructure. Journal of Financial Economics 3:257-75.
Comment
Antti Suvanto
The authors make the empirical observation that volatility of the deutsche mark/dollar exchange rate and the spread are high at both the open and the close of the trading day. This U-shaped pattern is evident for both New York and London. Because of time-zone differences, London opens when New York is still closed. When New York opens, London still has a few hours to go. The second observation made in the paper is that the higher volatility and wider spread in New York at the time of opening are in no way reflected in the volatility and spread in the London market. Similarly, the higher volatility and wider spread at the time London closes do not affect midday volatility and spread in New York. These observations seem to contradict predictions of recent microstructure theory based on asymmetric information between informed insiders and liquidity traders. This theory has been used to explain the similar U-shaped intraday pattern of the volatility and the spread in organized securities markets, such as the New York Stock Exchange. The authors argue that, because the foreign exchange market is globally integrated and the object of trade the most homogeneous of all, the participation of informed traders at the beginning of the trading day in New York should be visible in London as well. Therefore, the spread should widen and the volatility increase at the same time in both markets. The same should be true if informed trading in London is concentrated toward the end of the trading day. Because this is not the case, the asymmetric information argument is not applicable to the foreign exchange market. The authors suggest two types of explanations for these observations. The first is based on the heterogeneity of information across marketmakers. The
Antti Suvanto is head of the Economics Department of the Bank of Finland.
68
David A. Hsieh and Allan W. Kleidon
second is based on the inventory (position) adjustment behavior of "day traders." I find these suggestive explanations plausible, and I agree with the authors' final conclusion that further investigation along these lines is warranted.
Position Constraint Let us take the inventory argument first. The authors point out that a large proportion of foreign exchange trading is "day trading," by which they mean that the banks open and close the day with closed positions. I am not sure whether it is accurate to refer to smaller regional banks alone in this context because all banks, including branches of the big international banks, are "day traders." They stop trading at some point in the day, whether their overnight position is open or not. Otherwise, there would not be distinct but overlapping trading days in London and New York. The point made by the authors is, however, relevant. The fact that important news may be disclosed during the time when the bank is not trading and the fact that trading continues in other time zones imply that the dealers must pay attention to overnight positions. As the trading day winds down, dealers must either close a position or aim at a desired fixed open position. Whatever the position target for the end of the day, the problem of achieving this target becomes increasingly acute toward the end of the trading day, as the authors correctIy point out. I have shown elsewhere that this is true in a partial equilibrium model of the pricing behavior of an isolated profit-maximizing marketmaker facing stochastic but price-sensitive order flow from uninformed customers and aiming at a closed position or at any other fixed position target for the end of the day (Suvanto 1993, chap. 2). Applying dynamic programming techniques, the resulting sequential pricing rule implies weak efficiency of the mid-rate and low-volatility of prices (relative to the volatility of order flow) during the most of the day, with volatility increasing toward the end of the day. The spread is chosen so as to maximize the expected trading income at each short trading period given the price sensitivity of customer orders. The reluctance of the dealer to make frequent and drastic changes in the quoted two-way price in the absence of new information derives from the fact that such changes are generally revenue reducing, that is, selling at a low price a currency just purchased at a high price. Note that in this model the binding two-way price is set before the direction and size of the trade are known. The counterpart to small price adjustments is the dealer's willingness to accept comparatively large open intraday positions. In fact, the ex ante position variance increases at the beginning of the day and starts to decline toward the end of the day. Although the environmental assumptions of this (monopolist) model are far from those of the highly competitive and closely integrated foreign exchange
69
Bid-Ask Spreads in Foreign Exchange Markets
market of today, the results can be generalized to a more realistic setting. Opening the model to competition and interdealer transactions makes the quotation of an individual dealer dependent on the price quotes of others and narrows the spread. From the position-adjustment point of view, the life of an individual dealer is easier because he can undo an unwanted position by calling another marketmaker or he can stay on the "right side of the market" and attract orders from other dealers by making a very small price adjustment. This is true for the most of the day, but the mechanism does not work in the same way just before the close. Dealers who are already satisfied with their positions do not want new orders. Widening the spread decreases the likelihood of new orders. There may be unwanted open positions that do not find buyers until the spread is sufficiently large to make the price attractive to somebody, or sufficiently unattractive to the seller, until either of them is willing to carry an open overnight position. Allowing for overlapping time zones eliminates this problem because dealers in a closing time zone can sell unwanted open positions to buyers in open time zones. Dealers in the latter are happy to buy these positions because they are on the "right side of the market" and have plenty of time to undo the position during the remainder of the day. Heterogeneous Information Let us now turn to the heterogeneity of information. Information may be heterogeneous even if all participants have the same information but interpr~t its significance for the exchange rate somewhat differently. Economists vety often have the same information. Yet they may permanently disagree on the interpretation. Could one reasonably expect dealers always to have a uniform interpretation of the remainder of the day? The authors point out correctly that dealers need to get the "feel" of the market before engaging in active trading. What does this feel mean? Assume that at the time of the London opening all dealers have the same information on the macroeconomic and political news that has been reported since the previous close and that they have the same expectations about news that is forthcoming during the day. Assume further that each dealer knows with probability one that all nonbank customers that may appear during the day are price-sensitive liquidity traders. These rather extreme assumptions should eliminate all information asymmetries at the beginning of the day, but this is not the case. Depending on the situation, each dealer acts in two different roles in relation to his interbank counterparties. Sometimes he is on the "right side" of the market and acts as a marketmaker in relation to another dealer who is requesting a quote. On other occasions, he is on the "wrong side" of the market, requesting a quote from someone else. Each time dealer A quotes a two-way price to another dealer, B, he does not know exactly what prices are quoted by his rivals. In particular, he does not
70
David A. Hsieh and Allan W. Kleidon
know whether at the very same moment his dealer-customer B is receiving from dealer C a price that is inconsistent with his own quote in the sense that his ask price is below the bid price quoted by dealer C (or vice versa). If this is the case, there is an arbitrage opportunity for dealer B, who would never leave such an opportunity unexploited. I have argued elsewhere that the possibility of inconsistent quotes cannot be eliminated entirely as long as two-way prices are quoted independently without full knowledge of all other quotes at the same moment (Suvanto 1993, chap. 3). This brings potential insiders into the picture because some participants may have observed a pair of inconsistent prices. This explains the positive bid-ask spread in the same way as the possibility of informed traders explains the positive spread in the stock market. Fundamental valuation efficiency plays no role in this context. A dealer who has a given interpretation of the significance of publicly available information on the exchange rate during the remainder of the day may take a position based on this view by asking a price from another dealer and transacting accordingly, but he does not make a large adjustment to his own price immediately if he is not absolutely sure that all other marketmakers have already made a similar price adjustment. If he did, he might receive unwanted orders because his quote might be inconsistent with prices quoted by others. Rather, he is likely to widen the spread in order to reduce the likelihood of such an occurrence. It is plausible to assume that dealers opening their screens in the morning have different outlooks for the remainder of the day even in the absence of any private inside information. Dealers are cautious not to start the day with a price far out of line with other quotes, which explains the wide spread. But with a wide spread there is little or no trading and hence little or no trading profit. Therefore, the dealers are eager to start trading as soon as possible, which explains the rapid convergence of the spreads to the level that makes active trading possible. Once established, the "feel" of the market is maintained because the prices and the spreads can be constantly tested in active trading. The short period during which the "feel" of the market is established in an opening time zone need not affect prices quoted in a time zone where the markets are soon to close, provided there is sufficient overlapping. Dealers in the opening zone may get attractive quotes from the soon-to-close zone, where some dealers may get rid of their remaining positions at an attractive price. In addition, sufficient overlapping provides dealers in the soon-to-close zone sufficient time to wait until the "feel" is established in the opening zone.
Additional Remarks The paper marks an opening to a new area of research, that is, the intermarket connections across overlapping time zones. It would be interesting to have comparable data on the overlapping Asian and European markets. Data on the
71
Bid-Ask Spreads in Foreign Exchange Markets
spread and volatility at closing time in Continental Europe might also be illuminating. The data are based on indicative quotations from the Reuters' screen (FXFX), which do not give the actual transaction prices or trade volume. As the authors point out, it would be useful to check the results using alternative and more accurate sources of information. As the data reveal, the changeover from the opening phase, when dealers are getting the "feel" of the market, to active trading takes place in a relatively short span of time. It is more or less over by the time corporate treasurers and other nondealer customers step in and begin to request prices. Similarly, most of the nonbank customers have probably already disappeared when, shortly before closing, the spread widens and the volatility increases. This raises the question of the role of nonbank customers in the foreign exchange market. Despite the high proportion of interdealer trade, nondealer customers may be needed after all to maintain the depth of the market. Otherwise, how would dealers in the aggregate earn their trading income over the long run? Reference Suvanto, A. 1993. Foreign exchange dealing: Essays on the microstructure of the foreign exchange market. Helsinki: Research Institute of the Finnish Economy (ETLA).
3
Interdealer Trade and Information Flows in a Decentralized Foreign Exchange Market William Perraudin and Paolo Vitale
How trading arrangements within the foreign exchange market might affect the behavior of prices and allocative efficiency is largely unexplored territory. In a recent survey article, Flood (1991) stresses how little is known about basic features of the foreign exchange market. For example, why is the magnitude of interdealer trading (80 percent of total volume) so great? Why is half of that trading intermediated by brokers? Of the few studies of foreign exchange market institutional arrangements, Grossman and Zhou (1991) study the effect of stop-loss trading rules typically used by foreign exchange dealers on optimal portfolio strategies of individual traders. Krugman and Miller (1993) adduce various consequences of such rules for the behavior of the foreign exchange market as a whole, arguing that they may provoke market crashes, which could in turn provide a justification for commonly observed intervention behavior by central banks like target zones. Bossaerts and Hillion (1991) examine the implications of dealer behavior for unbiasedness tests in foreign exchange markets. Finally, an interesting series of papers by Lyons (1992, 1993, 1995) directly analyzes the microstructure of the foreign exchange market. Lyons concentrates particularly on the role of brokers, viewing them as a means by which order flow information is aggregated and then disseminated among dealers. William Perraudin is the Woolwich Professor of Financial Economics at Birkbeck College, London, a research associate of the Department of Applied Economics in Cambridge University, and a research fellow of the Centre for Economic Policy Research. Paolo Vitale is a Ph.D. student in economics at King's College, Cambridge, and a research economist at the Bank of England. The authors thank Silviero Foresi, Alan Kirman, Maureen O'Hara, Kathy Rudd, and seminar participants at the central bank of Sweden and the Perugia conference for their comments. Perraudin's research was supported by grants from the Economic Social Research Council (under grant ROOO 234641) and the Newton Trust, while some of the work was completed while he was visiting the Stockholm School of Economics. Vitale's research was supported by a Luciano Iona scholarship of the San Paolo Bank, Italy.
73
74
William Perraudin and Paolo Vitale
While this is certainly a fruitful line of research, it is not clear in such models why individual dealers do not have an incentive to deal directly rather than through brokers. Also, conversations with traders suggested to us that brokers are primarily important because of the efficient access that they provide to large numbers of other market participants. For these reasons, the analysis of foreign exchange trading that we attempt in this paper has a somewhat different focus from that of Lyons. We concentrate on the consequences for efficiency and exchange rate behavior of the market's decentralized nature, that is, the fact that dealers are ignorant of the order flow of other marketmakers. Interbank trading is modeled as a means by which marketmakers "sell" each other information about their transactions with outside customers. We show that, under these assumptions, decentralized market arrangements are privately efficient for the group of marketmakers. The reason that a decentralized market works well in this case is that it allows marketmakers to capture, through interdealer trade, the informational rents associated with receiving outside orders and hence gives them an incentive to adjust their spreads optimally to maximize those rents. If marketmakers are able to transact with only a fraction of other dealers between customer orders, a centralized market in which order flow information is freely and instantaneously available may be preferable to decentralized arrangements. Although incentives to adjust bid-ask spreads efficiently are diluted in a centralized market, at least dealers can observe customer orders and rationally update their subjective probabilities. An important aspect of the relative efficiency of different market arrangements is their robustness to extreme informational asymmetries. Glosten and Milgrom (1985) discuss the market crashes that may occur when dealers suspect that large numbers of informed agents are present. During such crashes, volume dries up as spreads widen and the informativeness of prices is lost. One of our more interesting findings is that such crashes happen much less in decentralized than in centralized markets such as those studied by Glosten and Milgrom (1985). The reason is that dealers have an incentive to maintain at least some turnover in order to elicit information that they can us in future trading. As well as supplying results, our analysis sheds light on the implications of decentralized markets for the time-series behavior of exchange rates. Most notably, we find, first, that the usual martingale properties of prices are absent, in that, in the decentralized market, the bid and ask on average increase and decline, respectively, as trades reveal information. Second, the unconditional variance of changes in bid and ask quotes is greater in the decentralized than in the centralized market. We also prove a series of comparative static results, demonstrating that, as one might expect, in the decentralized market spreads widen as the proportion of informed traders increases, while the bid price is increasing in the probability of higher exchange rate values.
75
Interdealer Trade and Information Flows
Previous theoretical work on decentralized markets has been quite limited, although recently several authors have begun to address the issue. Biais (1993) studies the effects of deviations from transparency in an inventory model of microstructure. Our approach, concentrating as it does on informational asymmetries, may be viewed as complimentary to Biais's work. An interesting recent paper by Neuberger, Naik, and Viswananthan (1993) examines the effect of trade publication delays on price formation in the London Stock Exchange. Although their modeling approach is quite different from ours,l they emphasize as we do the informativeness of customer trades. The structure of our paper is as follows. In section 3.1, we set out the model, studying first a static model of dealing with informed and uninformed outside customers and then showing how this fits into a more complicated dynamic framework with two periods of customer trading separated by a period of interbank transactions. Section 3.2 describes the results that we obtain with the model. These include results on efficiency, statistics of bid and ask prices, comparative statics for dealer prices and informational rents, and a result on interdealer market volume. An appendix provides a complete account of the proofs of the various lemmas and propositions.
3.1 3.1.1
The Model Basic Assumptions
Suppose that there are n identical dealers and four periods, denoted 0, 1, 2, 3. In period 0, each dealer selects a bid and ask and then trades with a customer if one presents himself. In period 1, dealers may trade among themselves if they so wish. In period 2, dealers trade again with customers if there are any. In period 3, all uncertainty is resolved. Assume for simplicity that all agents are risk neutral and that the interest rate is zero. 2 The structure of trading and of the stochastic order flow from customers is as follows (for a summary, see figure 3.1). The value of the exchange rate in period 3 is the realization of a random variable, z. Let the unconditional distribution of z be binomial in that z takes the values 1 and -1 with probabilities q and (1 - q). In period 0, a single customer arrives and is allocated with probability lin to a given dealer. With probabilities a and 1 - a, respectively, the trader is either informed or a liquidity trader. Informed traders transact if and only if their observation of the true value, z, exceeds the ask quoted by the dealer, SA
1. Their model has a batch trading structure with a single price rather than a bid-ask spread, and risk pooling plays an important role. 2. The latter is simply a normalization as we could value assets relative to the value of a safe bond.
76
William Perraudin and Paolo Vitale
1 n
No Customer
Customer
for Dealer i
for Dealer i
I-a
Informed
Uninformed
Customer
Customer
.!.=..§A
1_ q
2
q
No
Buy
Sell
Trade
t=t+l Up-date q. If t < 2 inter-bank trade and goto START. Else stop.
Fig. 3.1
t=t+l
If t
< 2 go to
START. Else stop.
Stochastic order flow structure
(in which case they buy), or falls below the bid price, s B (in which case they sell). We suppose that liquidity traders buy, sell, or do not trade with probabilities (1 - sA)/2, (sB + 1)12, and (SA - SB)/2. Note that in our model the price sensitivity of liquidity trader orders receives more stress than it does in much of the market microstructure literature. The dealers in our formulation set their spread in order to exploit their monopoly power over part of the uninformed order flow. In stressing this aspect, our study resembles the important early paper by Copeland and Galai (1983). In period 1, dealers may trade among themselves in such a way as to convey information about the customer order received in period O. In period 2, each dealer again receives a customer order with probability lin that once more
77
Interdealer Trade and Information Flows
originates with an informed or an uninformed trader with probabilities a and 1 - a. Finally, we adopt the following definition:
Definition 1. A centralized market is one in which information on dealers' order flow is freely and instantly available to other dealers. A decentralized market is one in which this is not the case. The above structure of orders implies that our model is applicable to an extremely short period of time. To be specific, we analyze interdealer transactions that can occur in the moments between two substantial and possibly informative customer trades. Perhaps the best way to think of the model is as a description of the situation faced by dealers immediately following some event affecting the foreign exchange market. The Static Problem
3.1.2
Before considering the more complicated dynamic problem of the dealers described above, let us study the static problem of a dealer who has a single opportunity to trade with informed and uninformed customers. Again assume that informed traders arrive with probability a, and suppose that q is the dealer's current conditional probability for the event {z == I}. The intuition for some of the more important points that emerge from our analysis may be understood even within this simple framework. It is also important to understand the structure of the static problem since this is what dealers face in period 2 even when they have a fully dynamic problem to solve in period O. The dealer's static value function I1(sA' sB) may be written as (1)
where (2)
lliSA)
==
(l - a) (l
== (1 - a) (3)
II B( sB) == (1 - a)
== (1 - a)
~ SA) (SA
- E(z»
+
aE{(sA - z)l{z = I}}
(1 - s ) 2 A (SA - (2q - 1))
(1
+s
(1
+s
) 2 B (E(z) - sB)
2
) B
+
+
aq(sA - 1),
aE{ (z - sB) 1{z == - 1} }
((2q - 1) - SB) - a(1 - q)(1
+
SB)'
where we use the fact that E(zlq) == 2q - 1. The static value function is quadratic in the quotes, SA and sB. Maximizing this function with respect to SA and SB yields the maximizing arguments
(4)
. SA* == mIn
{q - - , 1} ,s I-a
* B
== max
{I q, - I} . - --I-a
78
William Perraudin and Paolo Vitale
o
I
I
I I I I I
: SAO
/
IT(SA)
I I
o
A
: :
I
""
I
I I I I
SAO
o
IT(SB)
~
o
I
SBO
-1
o
c
B
Fig. 3.2 Static value maximization
The min and max operators in equation (4) appear because the maximum of the unconstrained value function may lie outside the interval [-1, 1]. In this case, the situation is as depicted in figure 3.2b, c. This feature of the model will generate subcases for the various propositions that we develop below. In the absence of asymmetric information, that is, when a == 0, the optimal static quotes, s~ and s;, equal q and q - 1 respectively. In this case, the dealer's calculation is motivated solely by the desire to exploit optimally the downward-sloping demand curve for liquidity trader orders that he faces. For a > 0, the absolute magnitude of the optimal quotes for a static dealer increases as he is now obliged to protect himself against informed trades by widening his spread. Substituting for s~ == SA(q) and s; == SB(q) in (2) and (3), we obtain that, for a ::; 1/2, I1(SA(q), SB(q)) ==
(5)
((1 - a) - q)2/(2 - 2a)
(2q(q - 1)
+ a? +
{ (q - a)2/(2 - 2a)
(1 - 0.)2)/(2 - 20.)
if 0 ::; q < a, if a. ~ q ~ (1 - (1), if (1 - a) < q ::; 1.
One may note that 11 is quadratic and continuous in q and that it has an interior minimum at q == 1/2. The form of 11, which will be important for our results below, is shown in figure 3.3a. 3.1.3
Filtering
Suppose now that the dealer trades in more than one period. In this case, he may be able to use the information that he has gained from period 0 trades. The information is potentially valuable, first, because, in his own period 2 trading, it may permit him to quote a bid-ask spread that yields higher expected profits. Second, he may be able to "sell" the information to other dealers in that they may be willing to trade with him at advantageous terms in the interbank market in order to learn about his order flow. Suppose that dealers use Bayes's rule to update their probability assess-
Interdealer Trade and Information Flows
79
I
.L - - _.
-
;.,... ~
o
: Il(sA)
~.---'. -:--I
,,-
I1(q)
,,-
I
,,-
I I
;'
,.
,.
I
;'
I
,.
I I
I I
I I
o
SAO
1
B
A
Fig. 3.3
o
I-a
Static profits and binomial probabilities
ments. In this case, the conditional probability of the even {z == I} following a buy order at 0, qb' will be (6)
qb == Prob[z == Ilbuy] Prob[z == 1 and buy] _ aq + (1 - a)q(I - sAo)/2 Prob[buy] aq + (1 - a)(1 - sAo)/2
The updated probability for {z == I} following a sell, qs' may be similarly derived as (7)
qs -= Pro b[Z -- 11 se11] --
(1 - a)q(I + sBo)/2 . a(I - q) + (1 - a)(1 + sBo)/2
The important point to note about the updated probabilities is that they depend on the bid and ask quotes in the first period, S BO and SAO. This dependency means that the dealer's choice of period 0 quotes will be influenced by the effect of order flow information on period 2 profits. Since order flow information may be valuable to other dealers, the possibility of "selling" this information through trading in the interbank market will also affect the dealer's optimal choice of period 0 quotes. In standard, dynamic microstructure models such as those of Glosten and Milgrom (1985) or Easley and O'Hara (1987), this particular link between trading in different periods is broken by the fact that order flow information is assumed to be public knowledge. 3.1.4
Information Rents
Consider a dealer who has received no orders in period O. If he trades again with outside customers in period 2 without receiving any information about
80
William Perraudin and Paolo Vitale
other dealers' trades, his expected profits will be II(q). On the other hand, if he can buy information from another dealer who has received, for example, a buy order, his expected profits will be II(qb). The total increase in his expected profits when he learns of a buy order, II(qb) - II(q), may be decomposed into a news effect and a "feedback rent." Let II[sA(q), sB(q)lqb] be the expected profits that the dealer obtains under updated probabilities, qb' but under the assumption that he sets SA and SB as if the probability of {z == I} were q. Now, one may write
(8)
II(qb) - II(q)
==
(ll(qb) - II(sA(q), sB(q)lqb))
+
(II(sA(q), sB(q)lqb) - II(q)).
The first bracketed term on the right-hand side is the expected value to the dealer of being able to adjust his quotes in response to the information. We refer to this extra value as the feedback rent. 3 This represents the increase in expected profits that the dealer can achieve by using the information to select his period 2 bid and ask quotes more efficiently. The second bracketed term in (8) is the pure news value of the information, that is, the change in the dealer's expected profits in a case in which he were, for some reason, unable to adjust his period 2 quotes. A similar decomposition can, of course, be performed for the change in expected profits due to information on a sale. As mentioned above, our formulation of the interdealer market will entail informed dealers passing information to each other through their period 1 trades. The surplus over which they may be expected to bargain will then be the feedback rent, II(qb) - II(sA(q), sB(q)lqb). Note that, even if II(qb) - II(q) is negative, dealers will be willing to pay each other for the information so long as the feedback rent is positive. This is analogous to the willingness of someone to pay for news that he is going to die so as long as the knowledge will allow him to take actions that prolong his life at least a little. Now, in our case, the feedback rent, II(qb) - II(sA(q), sB(q)lqb)' is nonnegative, as one may see from the fact that II(qb) == maxqE[O,l]{II(sA(q), sB(q)lqb)}. Hence, dealers will always be willing to pay for information. We shall suppose the following:
Assumption 1. Feedback rent associated with information on a customer trade is captured by the dealer who performs the trade. This assumption has the merit of substantially simplifying the analysis. Although it represents a polar case, we think it unlikely that our results would be substantially affected if a more even division of feedback rents were allowed for.
3. As in any dynamic programming problem under uncertainty, the dealer will do better if he can employ "feedback" controls that adjust according to information received.
Interdealer Trade and Information Flows
81
3.1.5
The Dynamic Model
Suppose that a dealer who receives an order in period 0 is able to transact with a fraction, k, of other marketmakers in period 1. We think it reasonable to assume that k is closer to unity than zero. A substantial marketmaking operation with a sufficiently large dealing personnel can arrange simultaneous trades with fifteen to twenty other dealers. The number of international banks that trade in substantial size does not greatly exceed this figure. Let P(A) == [exq + (1 - ex)(l - sAo)/2] and P(B) == [ex(l - q) + (1 - ex) (1 + S Bo)/2] denote the probabilities, respectively, of receiving an order at the ask or the bid in period O. The period 0 value function for the dynamically optimizing dealer is (9)
VO(SAO'
+ P(A) +
1
P(B) [ ~II2(sBo) +
sso) =
1 [ ~ II 2 (sAO)
+
k( n - 1 ) ] n rent(sBO)
1n - 1
+ - - - {Prob(A)II(sA(q), n
~ {IIo(SAo) + IIo(sso)
n
+
k(n -
n
1)
rent(sAO)
]
} + (1 - P(A) - P(B» 1 ~ II 2(sA2' SB2)
sB(q)lqb)
+
Prob(B)II(sA(q) , sB(q)lq)
(1 - P(A) - P(B»II(sA2' SB2)}'
where SA2 and S B2 are the optimal static, uninformed ask and bid quotes given in equation (4). When dealers can trade with all other dealers in period 1, the dynamic value function simplifies to (10)
1
Vo(SAO' SBO) = - {IIo(SAO) + IIo(sBO) + P(A)II 2(sAO) n + (1 - P(A) - P(B»II 2(sA2' SB2)}'
+
P(B)II 2(sBO)
Since all dealers are the same, by the symmetry of the problem, each dealer's value function equals lIn times the expected value of the total market order flow. When k = 1, this is the above, simple, bracketed expression. It is possible to obtain analytic solutions for the ask and bid quotes that maximize this value function since the first-order conditions tum out to be cubic functions of the prices. 4 In fact, the complexity of the resulting expressions means that they are not of much practical use. However, as we show in the results section below, one can learn a considerable amount by analyzing the first-order conditions and examining properties of the period 2 problem than by solving directly for bid and ask prices. 4. Although complicated, closed-form solutions to cubic equations are available (see Abramovitz and Stegun 1964).
82
William Perraudin and Paolo Vitale
3.2
Results
3.2.1
Information and Expected Profits
In this section, we establish two propositions on the value of information in our model. We start with the following:
Proposition 1. The following three statements are equivalent: 1. News of a buy order increases total expected profits in period 2. 2. News ofa buy order decreases dealers' estimates of the conditional variance of the exchange rate. 3.
(11)
q
2::
1/2,or
ex >
(1/2 - q)(l - q) 1/2 - q
+ q2
.
A similar result holds for news of a sell order.
To understand what drives this result, examine figure 3.3 above. Panel b of the figure shows the total increase in expected profits associated with a buy order, II(qb) - II(q). One may easily demonstrate that qb 2:: q and qs :5 q. In the case depicted, the initial unconditional probability q equals qo' while qb > qo > 1/2. The fact that II(q) is quadratic and has a minimum at 1/2 immediately implies that II(qb) - II(q) > 0; that is, information of a buy order implies higher expected profits. On the other hand, since in the diagram TI(q) < TI(q), it follows that news of a sale lowers expected profits. Why does more information in the latter case lead to lower expected profits? The reason is as follows. One may show that the variance of the binomially distributed random variable, z, equals 4q(1 - q) and that this has a maximum at q = 1/2. Thus, any information that implies a filtered, updated probability, qu' that lies closer to 1/2 than the original unconditional q also implies a increase in variance. 5 But higher variance lowers the dealer's expected profits for the reason that the profit function (see eqq. [2] and [3]) is made up of kinked functions of the underlying payoff, z. In this respect, the dealer's profits resemble a short position in call and put options. Such claims are concave in the random payoff, so, by Jensen's inequality, adding uncertainty lowers expected value. The pure news value of information in our model takes the simple form (12)
pure news value
==
(TI(qb) - II(q» - (TI(qb) 2q - 1 - II(sA(q), sB(q)lqb» = ~ (qb - q).
This immediately implies the following result: 5. Readers more familiar with normal filtering problems may find this slightly surprising as updating, in that case, reduces uncertainty. In the present context, for certain values of q, updating actually increases the conditional variance.
83
Interdealer Trade and Information Flows
Proposition 2. When buy and sell orders are equally probable, the pure news value of information is zero, and the feedback rent associated with information equals the change in expected profits. In other words, when q == 1/2, the pure news value is zero. To see the intuition behind this finding, suppose that the dealer cannot adjust his quotes in response to the information. Recall that the only three possible events are a buy order, a sell order, and no trade. If there is no trade, the dealer still has II(q). When q == 1/2, the problem is completely symmetric, and, hence, information of either a bid or an ask order must change expected profits by the same amount. But, if all news has the same effect on expected value, it must be that the effect is zero. Hence, the only possible increase in value from the news must come from the dealer's ability to adjust his quotes conditional on information. That is, the increase in total value equals the feedback rent. 3.2.2
Bid-Ask Spreads
In this section, we show that the ask of a dynamically optimizing dealer is greater than that of a static dealer. Since, when n is large, dealers in a centralized market behave as though they are static, profit maximizers, this statement may be regarded as a statement about the behavior of dealer prices in centralized versus decentralized markets. We state the result formally as follows:
Proposition 3. The optimal period 0 ask ofa dynamically optimizing dealer in the decentralized market exceeds that ofa static solution. A corresponding result holds for decentralized market bids that exceed static bids in absolute magnitude.
By increasing the bid-ask spread from its static level, marketmakers sacrifice short-term expected profits. On the other hand, however, they improve the quality of the information that they derive from period 0 trades since uninformed trades are discouraged from transacting. Using the improved information, dealers can earn higher expected profits in subsequent trading.
Corollary 1. As n ~
00,
the bid-ask spread is wider in a decentralized than
in a centralized market.
3.2.3
Efficiency
In this section, we compare the efficiency of centralized and decentralized markets. It should be apparent that the two market organizations each have advantages and disadvantages. In a centralized market, dealers observe all period 0 customer trades so that they can always update their subjective probability assessments and correspondingly adjust their period 2 bid and ask. On the other hand, in a centralized market, the incentives of dealers to adjust their period 0 bid and ask so as to elicit an efficient amount of information are diluted. In the limit, as the number of dealers n ~ 00, individual dealers in a centralized market will set SAO and sBO in a way that totally ignores the informa-
84
William Perraudin and Paolo Vitale
tional rents associated with period 0 trading. In other words, they will act as though they are static profit maximizers. In a decentralized market, the opportunity to sell information improves incentives to elicit information by optimal adjustment of the bid-ask spread. In the limit, when k = 1 and dealers can transact with alJ other marketmakers, the entire market feedback rent associated with a period 0 customer order is captured by the dealer who receives the order. In this case, marketmakers will optimally adjust their spreads to elicit information, and hence the decentralized market will be private efficient from the dealers' collective point of view. However, if dealers can transact with only a fraction of other marketmakers, the advantage of better incentives will be reduced as it will then not be possible to capture the feedback rent associated with information. In addition, the decentralized market will suffer from the fact that dealers who do not receive interdealer trades will be unable to update their probabilities in response to period 0 order flow. To analyze this trade-off formally, we start with the following lemma: Lemma 1. The unconditional expectation ofperiod 2 profits is greater when dealers are able to update probabilities on the basis ofobservation ofperiod oorder flow. Lemma 1 demonstrates that the ability to adjust quotes in response to information on customer trades in period 0 increases the unconditional expectation of period 2 profits. Using this lemma, one may prove the following proposition: Proposition 4. If dealers are able to transact with all other marketmakers in the interval between customer trades, that is, if k = 1, a decentralized market is fully efficient. If k is small, however, total expected dealer profits are higher in a centralized than a decentralized market. In the remainder of the paper, we shall assume for simplicity that k = 1. 3.2.4
Market Crashes
A point stressed by Glosten and Milgrom in their classic 1985 paper on dealer behavior was that markets with too many informed traders may collapse as dealers will be unable to make positive profits given the adverse selection problems that they face. Such market crashes will involve collapses in volume as bid-ask spreads increase until the market closes. Note that it is possible for the market to close on one or both sides of the bid-ask spread. Crashes are costly because they undermine the informativeness of prices. In our discussion of the static model in section 3.1, we have already implicitly considered such market collapses by discussing cases in which optimal ask and bid prices equal, respectively, plus and minus unity. Let us suppose, as seems reasonable, the following: Assumption 2. If SA = 1, informed traders never buy, while, informed traders never sell.
if sa =
1,
85
Interdealer Trade and Information Flows
Of course, in these cases, informed customers will be indifferent between trading and not trading, but any slight friction would be enough to make them strongly prefer not to trade. The expressions for ask and bid prices in the static model, SA = min{ql(1 ex), I} and SB = max{ -(1 - q)/(1 - a), -I}, immediately suggest under what conditions markets with static dealers will collapse. For any given q, if a is large enough, SA = 1, and S B = - 1. On the other hand, for any given a, if q ~ 1, eventually, SA = 1, while, if q ~ 0, eventually, SB = 1. Notice from the discussion in the last paragraph that there are two reasons why the static market may collapse, either (i) too many informed traders (a large for q around 1/2) or (ii) too little uncertainty about tp.e value of the exchange rate (q close to zero or unity). These two reasons lead to qualitatively different outcomes in that in the first both sides of the market close while in the second only one side of the market crashes. Figure 3.4a illustrates the way in which in the static model, for a given a, different assumptions on q may generate crashes. In the case illustrated (in which a > 1/2), for a < q < (1 - ex), the market crashes on both sides, and expected profits are zero. For q > a, the market crashes only on the ask side, while, for q < (1 - a), it crashes only on the bid. One of the most interesting implications of our model is the following: Proposition 5. The decentralized market never collapses in period O. The interest of this result is that it suggests that decentralized markets are significantly more robust to the asymmetric information problems that provoke collapses in the static model. Recall that, as n ~ 00, dealers in centralized markets behave like static profit maximizers, so once again this sheds light on the differences between centralized and decentralized market arrangements.
I
.L - -_.
,.
o
_
;."...~
- ,. -
: ll(SA)
-:-.~
~I
SAO
1
Il(q) / /
o
A
o
I-a
B
Fig. 3.4 Market crashes and information
.I
86
William Perraudin and Paolo Vitale
The intuition behind proposition 5 is that in a decentralized market dealers have an incentive to provide a small but sufficient incentive for informed traders to transact and hence reveal their information. Given our assumptions about the price elasticity of orders by uninformed traders, if SAO == 1 - e or sBO == -1 + e for small, positive e, dealers can obtain very good information in the event of a buy or sell order, respectively. As long as there is some rent to be extracted from this information in the form of higher period 2 profits, dealers will always have an incentive to open the market by adjusting their quotes enough to elicit trades from informed customers. The above proposition is illustrated by figure 3.4b, which shows expected ask-side profits in the static and dynamic models as a function of the period 0 quote. The static expected profits equal zero at SAO == 1, and unity is clearly the maximizing argument. The dynamic expected profits, which appear as a dotted line in the figure, are positive and increasing for ask quotes in an open interval below 1 and then, in fact, drop to zero at 1. The fact that they are positive for SAO == 1 - e for small, positive e is what gives the above result. 3.2.5
Martingale Properties and Volatility
A feature of standard market microstructure models with competitive marketmakers (see, e.g., Glosten and Milgrom 1985; and Easley and O'Hara 1987) is that bid and ask prices are martingales with respect to the information available to dealers. In this section, we shall see that, in our decentralized market model, this is no longer the case and that, in fact, bid and ask prices exhibit mean reversion as information is revealed. Assumption 3. Suppose that interior solutions exist for the static model in period 0, that is, that <X < q < 1 - <X. First, consider the unconditional expectation of the difference between ask prices in periods 0 and 2. Proposition 6. In the static model, (13)
E[sA2 - sAoIS ]
== 0,
E[sB2 - sBoIS]
== 0,
while, in the dynamic model, (14) This proposition shows that weak-form market efficiency does not hold in our dynamic decentralized market while, in a centralized market with a large number of dealers, bids and asks will be martingales with respect to marketmakers' information. The basic feature of the model that permits deviations from martingale behavior is the market power that we assume for dealers. O'Hara (1994) comments on the fact that monopolistic elements can give such deviations.
Interdealer Trade and Information Flows
87
Decentralized and centralized markets also differ in the amount of volatility that they imply. One may show that the unconditional variance of changes in bid and ask is greater in the decentralized case.
Proposition 7. The unconditional variance of quote changes is greater in the dynamic than in the static case, that is, (15) (16)
3.2.6
< Var(sA2 - sAoI D ), Var(sB2 - sBoIS) < Var(sB2 - sBoI D ).
Var(sA2 - sAoI S )
Comparative Statics
Proposition 3 simply states that the bid-ask spread is larger in a decentralized than in a centralized market, without solving for the values taken by the bid and ask quotes. Although exact solutions for these quotes can be found, they are so complex that little more can be deduced. One may still, however, analyze the first-order condition of the maximization problem to learn more about these solutions.
Proposition 8. Let s:o be an internal optimum for the ask price. Then the following results hold: (17)
dS: o 0 as:> 0 for ex < q. o dq> , dex '
As a first result, proposition 8 shows that the ask price always increases when the probability of the event {z = I} rises. Furthermore, an increase in the proportion of informed traders produces a rise in the ask. The reasons for these results are two: first, the need for protection against the informed traders is stronger; second, there is an incentive to increase the size of the bid-ask spread because with more informed traders it is possible to get more information.
Proposition 9. Suppose that 0'. < q < 1 - 0'. < qb and that the quotes chosen by the dealer are internal optima. Then the following results hold: a rent 0 -- < (18)
aq
(19)
,
a rent 0 a-0'. > .
The result on arentlaq is interesting as it indicates that, for given 0'., when there is less uncertainty (q further from 1/2), there is less ex ante information in the order flow, that is, less opportunity to sell information.
3.2.7
Interdealer Market Volume
When a marketmaker has received some information, we assume that he can sell it in the second period to all the remaining marketmakers. He accomplishes this by transacting on favorable terms through the interdealer market.
88
William Perraudin and Paolo Vitale
We suppose that marketmakers will quote other dealers a bid-ask spread that (i) is "regret free" according to the definition of Glosten and Milgron, in that after a transaction marketmakers do not regret having completed it, (ii) transfers feedback rent to the informed marketmaker, (iii) creates no incentives for uninformed marketmakers to pretend to possess information, and (iv) minimizes the quantity transacted. The last property requires some comment. A given rent can be transferred between dealers by various combinations of price and quantity. In this sense, the prices quoted between dealers are indeterminate. However, point iv implies unique interdealer quotes since, if the transaction size is reduced too far and the spread made too generous, eventually uninformed dealers will be able to make profits masquerading as informed. Assuming that the size of trades is reduced to a minimum implies that the incentive constraint implicit in point iii must hold as an equality and hence determines interdealer quotes. Now, suppose that marketmaker 1 has received a buy order at time 0 (the same reasoning applies for a sell order) and wants to sell the information to marketmaker 2. As he is willing to buy the currency and his expected value of the currency is 2qb - 1, marketmaker 2 will quote an ask price in the interval (2q - 1, 2qb - 1), and a transaction will be completed for a quantity Llx that conveys the rent of a buy order to marketmaker 1 (subject to the assumed properties of interdealer trade (points i-iv above). Proposition 10 reports an interesting result concerning this value: Proposition 10. Let us define Ll x * as the minimum quantity transacted among two marketmakers. Suppose that the regularity conditions ofproposition 9 are satisfied. The following result holds: (20)
aLlx* > 0 .
aa
The proposition simply indicates that, given these conditions, the volume of transactions between marketmakers is indicative of the informativeness of the order flow as the minimum quantity transacted among two marketmakers is an increasing function of a. 3.3
Conclusion
This paper has provided a theoretical analysis of a decentralized dealer market. Although our results are relevant to a broad category of markets in which order flow information is not publicly available the primary motivation for our study was the desire to understand price formation and efficiency in the foreign exchange market. Our main findings are the following: 1. Bid ask spreads are wider in the decentralized market. The intuition here is that, by posting wider spreads, dealers can discourage price-sensitive liquid-
89
Interdealer Trade and Information Flows
ity traders and hence improve the informativeness of their order flow. The information embodied in orders can in tum be used to earn higher future profits and can be "sold" to other marketmakers through interbank transactions at advantageous prices. 2. Decentralized markets are privately efficient from the collective point of view of marketmakers when it is possible for dealers to transact with all other dealers in between potentially informative customer trades. This point underlines the potential importance of brokers as a way of facilitating large numbers of simultaneous transactions with other marketmakers. 3. Decentralized markets are much less subject to market crashes than centralized markets. Information on order flow may be used to update subjective estimates of the underlying value of exchange rates. Even in circumstances in which static or centralized markets would crash owing to excessive numbers of informed traders, dealers will have an incentive to preserve some turnover in the decentralized market as they can employ the information in the order flow in subsequent trading. Our model allows only two periods of trading with customers, but we would conjecture that our results on crashes would hold in a multiperiod model, in that dealers would always have an incentive to preserve at least some order flow to gain information. 4. The time-series behavior of exchange rates in our model differs according to whether trading is organized on a centralized or a decentralized basis. When dealers maximize profits in a static fashion (which they will do in a centralized market containing large numbers of marketmakers), bid and ask quotes are martingales with respect to the information available to dealers. In the decentralized market, bid-ask spreads on average shrink as order flow reveals information. It is very interesting to note that this implication of the model is consistent with the findings of Goodhart and Figliuoli (1991). Their study suggests that, prior to jumps in exchange rates, there is an increase in the negative autocorrelation. If we regard jump times as moments at which significant information becomes public knowledge (i.e., corresponding to our period 3), then our model would suggest that, in the immediately preceding period, a small number of agents will know the information and dealers will be adjusting quotes so that the bid-ask spread is contracting on average. 5. Another implication of the model for the statistical properties of exchange rates is that changes in rates will be more variable in the decentralized than in the centralized market. It is perhaps not clear quite what is the quantitative significance of this difference in variance, but, given the widely acknowledged volatility of exchange rates, it is at least reassuring that our model predicts greater variance in decentralized markets.
90
William Perraudin and Paolo Vitale
Appendix Proofs are stated for the ask side of the market throughout. Similar arguments apply to the bid side. Proof of Proposition 1
Proposition 1. The following three statements are equivalent: 1. News of a buy order increases total expected profits in period 2. 2. News ofa buy order decreases dealers' estimates of the conditional variance of the exchange rate. 3. (21)
q
2::
1/2,or
ex
>
(1/2 - q)(1 - q) 1/2 - q
+ q2
.
A similar result holds for news of a sell order.
Proof. The equivalence of the first two statements is obvious; in fact, as II(sA(q), SB(q)) is symmetric around 1/2, we have a gain in the expected profits from a buy order if qb - 1/2 > 1/2 - q; that corresponds to a reduction in the conditional variance of the exchange rate. Moreover, qb - 1/2 > 1/2 - q holds if (22) Then, for SAO
exq2
+ (1
- ex)(1 - SAO)(q - 1/2)
> O.
= q(1 - ex), this condition becomes exq2 > (1 - ex - q)(1/2 - q).
It is immediately obvious that this condition holds for the values of ex and q respecting the condition (21).. This completes the proof. Proof of Proposition 3
Proposition 3. The optimal period 0 ask ofa dynamically optimizing dealer in the decentralized market exceeds that ofthe static solution. A corresponding result holds for decentralized market bids that exceed static bids in absolute magnitude. Proof. Let V~ be that part of the dynamic value function that depends on the ask price, SAO' multiplied by the constant n. V~ equals
Assuming that ex < 1/2, and given the different form of the static value function for different configurations of q and ex, we consider six cases: q < qb < ex;
91
Interdealer Trade and Information Flows
q < a :::; qb < 1 - a; q < a < 1 - a :::; qb; a < q < qb < 1 - a; a < q - a :::; qb; and q > 1 - a; so that qb > 1 - a. In all six cases, we have
=
q - (1 -
a)sAO
We show that in all cases
r
+ r = O.
is positive and therefore
S~O 2::
(25)
< 1
the static solution.
In particular, we show that in the first five cases
S~O > - q - = the static solution,
(26)
1- a
while in the last case the static and the dynamic solutions are both equal to one. Case 1. q < qb < a. The components of r are as follows: (27)
(28)
(29)
a Prob(A) II (
as
)=
2 SAO
AO
Since (30)
aqb = (qb - q)(l - a)
as 2 Prob(A) a Prob(A) 1AO
(31)
2
It then follows that (32)
Case 2. q < a :::; qb < 1 - a. (33)
(34)
r ==
(qb - q)2 . 4
a
92
(35)
William Perraudin and Paolo Vitale
a Prob(A) II (s as 2
AO
)
= (1
- qb)qb _ a
2
+
(1 - ex)2 .
2
AO
4
We end up with
r == 2(qb
(36)
- q)2 - (q - ex)2 .
4
This is positive as qb > q and ex > q. Case 3. q < ex < 1 - a < qb. Using the fact that 1 - ex - - I I2(sA2'
(37)
2
SB2)
=
[(1 - ex) - q]2 4
,
(38)
a Prob(A) II ( ) = _
(39)
(qb - a)2
2 SAO
a~o
•
4
It follows that (40)
r ==
{(qb - a)2
+ 2(qb -- ex)(ex - q) + [(1 -
a) - q]2} .
4
That is positive as qb > ex and ex > q. Case 4. ex < q < qb < 1 - ex. We now have (41)
(42)
(43)
a Prob(A) II (s as 2 AO
) AO
= (1
- qb)qb _ a 2
2
+
(1 - a)2 . 4
Hence
r == (qb -
(44)
Case 5. a < q < 1 - ex (45)
(46)
q)2 .
2 ~ qb.
We have that
1 - ex _ ex 2 + (1 - a)2 (1 - q)q - - I I2(sA2' SB2) - , 2 4 2
93
Interdealer Trade and Information Flows
a Prob(A) IT (
(47)
2 SAO
a~o
)::;::: _ (qb - a)2 . 4
This implies that
r :; {2qb(qb
(48)
- q)
+
[(1 - a) - q]2} .
4
But, since qb > 1 - a > q, it follows that q~ > -(1 - a)[2q - (1 - a)]. So r > 0 as required. Finally, consider the last case: Case 6. q > 1 - ex so that qb > 1 - a. Here we have 1- a
(49)
(q - ex)2
- 2 -IT2 (SA2' SB2) ::;:::
4
'
(50)
a Prob(A) II (
(51)
2 SAO
a~o
)::;::: _ (qb - a)2 4
.
Therefore
r == [2(qb -
(52)
a)(qb - q)
+
(q - ex)2 - (qb - 0'.)2]
4 ::;:::
+
[2q (q - q) b b
q2 - q2] [q _ q]2 b::;::: b
4
4
>
o.
This completes the proof. Proof of Efficiency Results Lemma 1. The unconditional expectation ofperiod 2 profits is greater when dealers are able to update probabilities on the basis ofobservation ofperiod order flow.
o
Proof. Define EIT u as the unconditional expectation of period 2 profits when probabilities are updated at 1 after the observation of a trade should one occur: EII u (53)
::;:::
[max II(SA2' SB2 I qb)] + Prob(sell) [max A2' II (SA2' SB2 I q)]
Prob(buy)
sA2' sB2
s
s
B2
+ [1 - Prob(buy) - Prob(sell)] [maxSA2' SB2 ll(SA2' sB2 I q)]
,
The unconditional expectation of profits without updating is denoted EII. By the iterative property of conditional expectations, this can be expressed in the following way:
94
William Perraudin and Paolo Vitale
Ell
= maxSA2,SB2
[Prob(bU Y) l1(sAz' SBZ I qb)
+ Prob(sell)
(54)
+
ll(sA2' SB2
I q)
[l - Prob(buy) - Prob(sell)] I1( sAZ' SBZ
It is then clear that Ell u > Ell for Prob(buy)
I q)]
.
+ Prob(sell) > 0.0
Proposition 4. If dealers are able to transact with all other marketmakers in the interval between customer trades, that is, if k == 1, a decentralized market is fully efficient. If k is small, however, total expected dealer profits are higher in a centralized than a decentralized market.
Proof. Follows from lemma 1. Proof of Proposition 5
Proposition 5. The decentralized market never collapses in period o. Proof. We can prove this statement in two steps. In the first, we prove that the market cannot be always completely closed in period 2 and at least on one side in period O. In the second, we prove that, if the market is open at least on one side in the second period, it cannot be closed on either side in period o. Step 1. Suppose that the marketmaker's strategy implies that the market is always closed in period 2 and is closed on the sell side in period 0; we show there exists another strategy that dominates it. Suppose that we fix in period 0 SAO - 1 and s BO == -1 + e, for e > 0 and small. Suppose that a dealer will set sA2 == 1/2 and s B2 == - 1 if he receives a sell order and sA2 == 1 and s B2 == -1 otherwise. As the sell order in period 0 for a given dealer will occur with probability a( 1 - q )/n, and as the bid-ask spread will have a negligible effect on period 0 profits, the total expected profits will be (1 - a)(l - sA2)(1 + SA2)/ 2 == 3(1 - a)/8; this implies that the second strategy dominates the first one. A similar argument works for the ask side. Step 2. Suppose that the ask side is closed in period O. Lowering SAO slightly hardly affects period 0 expected profits but means that, in the event of a buy order, the dealer receiving it knows almost surely that z == 1. The profit function with the ask side closed in period 0 is (55)
VO(SAO' SBO) = Vo(l, SBO) =
+ (l With SAO == 1 (56) VO(SAO' SBO)
e,
~ {l1o(SBo) + P(B)l1zC sBo)
- P(B»l1zC sAz' SBZ)} .
we have
=
Vo(l - e, SBO)
=~
[l1 o(SBo)
+ P(A)l1 z(sAO ISAO~l - e)
95
Interdealer Trade and Information Flows
+ P(B)II/sBO ) + (1
- peA) - P(B))II 2(SA2' SB)] .
Hence, the result follows if IT 2(sAO ISAO-1 _-
e
) > IT 2(sA2' SB2). But, as long as q is
different from 1, (57)
II/SAO ISAO~ I _ ) = II/qlq~l) > II 2(q) = II/SA2, SB2) .
Therefore, the ask side of the market will not be always closed in period 2; a similar argument holds for the bid side. This completes the proof. 0 Statistics of Quote Changes Proposition 6. In the static model, (58)
while, in the dynamic model, (59)
Proof. In the dynamic model, we can write the difference in expectations as:
E[SA2 - sAoI D ] (60)
= Prob(sell)
(~ 1-0.
SAO) + Prob(buy)
(~ 1-0.
SAO)
+ Prob(no trade) (-q- - SAO) 1 - 0.
-- -q- - SAO < 0, 1 - 0.
where we use the fact that Prob(sell)(qs - q) = - Prob(buy)(qb - q) and SAO > q/(1 - 0.).0 Proposition 7. The unconditional variance of quote changes is greater in
the dynamic than in the static case, that is, Var(sA2 - sAoI S ) < Var(sA2 - sAoI D ), Var(sB2 - sBoI S )
< Var(sB2 - sBoI D ).
Proof. Consider volatility with the two sets of quote-setting behavior. In both static and dynamic cases, (61)
Var(sA2 - SAO) = Prob(sell) [qs - q]2 1-0. = [
o.Q(1 - q)]2
1- a
+ Prob(buy) [qb - q]2
(1 Prob(buy)
1-0.
1)
+ Prob(sell) ·
96
William Perraudin and Paolo Vitale
The result then follows from the fact that Prob(buy) and Prob(sell) are larger in the static than in the dynamic case.D Proof of Comparative Statics
Proposition 8. Let s~o be an internal optimum for the ask price. Then the following results hold:
as:o >0
(62)
as:o >0 'aa '
aq
for a < q .
Proof. We can use the first-order condition to study the effect of a change in any parameter of the model, J3, on s:o' as the second-order condition guarantees that we still have an internal solution; therefore we consider
a2v'/aJ3as: o
as: o =
(63)
aJ3
a2v'/as:~
Now, as from the second-order condition it follows that a2 v' las~~ < 0, the sign of the derivative of s:o with respect to J3 corresponds to that of the numerator. Hence, consider a2V'laqas: o' which is equal to 1 + ar/aq. To prove that this is always positive, we have to consider five of the six cases discussed in proposition 2 because for q > 1 - a we do not have an internal optimum. It is easy to show that
a(qb - q) aq
(64)
= a(1 -
q - qb)'
a
where a = aq + (1 - a)(l - sAo)/2; for q + qb > 1, as 1 - qb > 0 and a > aq this derivative is negative but larger than -1; while, for q + qb < 1, this derivative is positive. This permits us to show that in all cases ar/aq > -1. In case 1,
ar =
(65)
aq
a(qb - q) qb - q aq 2
Therefore ar/aq > 0 as q + qb < 1. In case 2,
ar _ a(qb -
-aq
(66)
aq
q) ( ) qb-q
a - q +--.
2
Therefore ar/aq > 0 as q + qb < 1 and a > q. In case 3,
(67)
ar aqb aq = {aq (qb -
In case 4,
q) - (qb - ex) - [(1 - ex) - q]}12
> -1.
Interdealer Trade and Information Flows
97
ar _ ( ) a(qb - q) - - qb - q aq aq
(68)
> - 1.
Finally, in case 5, we have (69)
ar =
{(qb - q) a(qb - q)
aq 2
Let us consider a v' = aaas: o (70)
+
q/qb - (1 - a)}/2
aq
> -l.
aq
s:o + aflaa. In this case, we have that
aqb = (1 - s:o)q(1 - q) aa 2Ll 2
> o.
For a < q and qb < 1 - a, we have to discuss only case 4; we can easily prove that s:o + ar/aa is positive. In fact, we have (71)
ar _ aqb - - (qb - q) aa aa
> o.
Proposition 9. Suppose that a < q < 1 - a < qb and that the quotes chosen by the dealer are internal optima. Then the following results hold:
a rent 0 -aq -< ,
a rent> o. aa Proof. We assume that a is such that a < q < qb < 1 - a so that we concentrate on case 4. The rent from a buy order is given by (72)
rent(buy) =
(q _ q)2 b
1- a
•
In case 4, q + qb > 1 so that a(qb - q)/aq is negative. This is sufficient to prove that arent(buy)/aq is negative. Conversely, as aqb/aa > 0, it is immediately obvious that arent(buy)/aa is positive. D Proof of Gearing Effect Results
Proposition 10. Let us define Llx* as the minimum quantity transacted among two marketmakers. Suppose that the regularity conditions ofproposition 9 are satisfied. The following result holds:
aax *
->0. aa
Proof. We assume that a is such that a < q < qb < 1 - a so that we concentrate on case 4. If a buy order has been received, the minimum value of the
98
William Perraudin and Paolo Vitale
transacted quantity is (73)
ax*
= rent(buy)
.
2(qb - q)
We know that the rent of a buy order is given in case 4 by (74)
rent(buy) =
(q _ q)2 b
1- a
•
This implies that (75)
ax*
=
(qb - q) . 2(1 - a)
Therefore, as aqb/aa > 0, it follows that aax*/aa > O. This completes the proof.D
References Abramovitz, M., and I. A. Stegun. 1964. Handbook ofmathematical functions withformulas, graphs, and mathematical tables. Applied Mathematics Studies no. 55. Washington, D.C.: National Bureau of Standards. Biais, B. 1993. Price information and equilibrium liquidity in fragmented and centralized markets. Journal of Finance 48, no. 1 (March): 157-85. Bossaerts, P., and P. Hillion. 1991. Market microstructure effects of government intervention in the foreign exchange market. Review ofFinancial Studies 4, no. 3:513-41. Copeland, T., and D. Galai. 1983. Information effects and the bid-ask spread. Journal of Finance 38:1457-69. Easley, D., and M. O'Hara. 1987. Price, trade size and information in securities markets. Journal of Financial Econolnics 19:69-90. Flood, M. D. 1991. Microstructure theory and the foreign exchange market. Federal Reserve Bank of St. Louis Review 73, no. 6 (NovemberlDecember): 52-70. Glosten, L. R., and P. R. Milgrom. 1985. Bid, ask and transaction prices in a specialist market with heterogeneously informed traders. Journal of Financial Economics 14:71-100. Goodhart, C. A. E., and L. Figliuoli. 1991. Every minute counts in financial markets. Journal of International Money and Finance 10, no. 1:23-52. Grossman, S., and Z. Zhou. 1991. Investment strategies for controlling drawdowns. University of Pennsylvania, Wharton School of Economics. Mimeo. Krugman, P., and M. Miller. 1993. Why have a target zone. Carnegie-Rochester Conference Series on Public Policy 38:279-314. Lyons, R. K. 1992. Private belief and information intermediation in the foreign exchange market. University of California, Berkeley, Business School, November. Mimeo. - - - . 1993. Equilibrium microstructure in the foreign exchange market. University of California, Berkeley, Business School, June. Mimeo. - - - . 1995. Tests of microstructure hypotheses in the foreign exchange market. Journal of Financial Economics 39:321-51.
99
Interdealer Trade and Information Flows
Neuberger, A., N. Naik, and S. Viswananthan. 1993. Disclosure and trading with large market makers: An analysis of the London Stock Exchange. London Business School, January. Mimeo. O'Hara, M. 1994. Market microstructure theory. Cornell University. Mimeo.
Comment
Silverio Foresi
Perraudin and Vitale's paper explores the implications for equilibrium prices of a multidealer market in which dealers cannot see each other's order flow. Their main result is that a decentralized market dealers' market is less prone to market crashes than a centralized market. Given its policy implications, this result is very important and deserves closer scrutiny. Glosten and Milgrom (1985) show that the market closes down if the specialist needs to post too wide a spread to break even when the informational asymmetry is too severe: the dealer prefers not to trade rather than trading at a disadvantage with an informed customer. The failure to trade is an externality on future trades that is not accounted for by the dealer. So Glosten and Milgram (1985) go on to conjecture that a Pareto improvement would result from a dealer who could retain some monopoly power. Perraudin and Vitale propose an interesting mechanism that may give dealers incentives to trade even when they face severe informational asymmetries. If dealers can share information with each other, and, more important, if they can agree to act as a monopolist before trades start, they can extract the surplus from liquidity traders, and, by appropriating this rent, they have an incentive not to let the market break down. The possibility that dealers could learn from other dealers' quotes is quite appealing. (It was probably first presented formally by Garbade, Pomrenze, and Silber [1979], who tested it in the market for U.S. Treasury securities.) The main elements of the model are as follows. There are three classes of traders: informed, uninformed, and marketmakers or dealers. Everybody knows that the value of the underlying is S = 1 with probability q and S = -1 with probability (1 - q). The informed traders know the realization of S before the first round of trades starts. Everybody else learns about it after the last round of trade. The model features three trading periods: in the intermediate trading period, dealers trade among themselves, sharing the information they (may have) received in the first round of trading in order to trade again in the last round of trades. Uninformed trades are (1 - a)/a as numerous as informed traders. Interestingly, their net demands are sensitive to prices. This is crucial because dealers choose prices to maximize their profit from trading with uninformed traders Silverio Foresi is visiting associate professor of finance at Columbia University and assistant professor of finance at the Stem School of Business, New York University.
100
William Perraudin and Paolo Vitale
in a manner similar to a monopolist choosing quantities taking as given the demand function, max sQ(s); s formally, because of the price sensitivity of the net demand, the static profit becomes quadratic in the quoted bid (ask) prices. The price sensitivity of the demand of uninformed traders delivers the concavity of the profit function in the price s and its convexity in the probability of high outcomes q. It is not immediately obvious how to justify the assumption that informed dealers can appropriate the rent by trading with uninformed dealers in period 1. The issue is related to the question, Why aren't quotes in period 1 fully revealing? In standard models with noise that is because noise trade provides camouflage to informed traders. But there are only dealers trading in period 1. What provides camouflage to the "informed dealer" in period I? Since the uninformed dealer cannot see the informed dealer trade, he must learn from other dealers' quotes and possible trades with the informed ones. Assume that dealers do not act strategically. If dealer A, who saw no customer in period 0, calls dealer B in period 1, she learns from the quote whether A saw a sell, a buy, or nothing. But, having seen the quote, she does not need to trade. But why would informed dealers want to post bid-ask prices that reveal any information? There are two effects at play here. On the one hand, if information is better diffused, all dealers may agree to narrow their spread in period 2. This increases liquidity trades and presumably dealers' profits. On the other hand, larger liquidity volumes provide additional camouflage to informed traders. This reduces dealers' profits. It is unclear whether the balance of the two effects is positive. While the rent-sharing rule assumed in the paper is not essential, it is essential to show that some rule is viable. Consider the following story, which may justify the assumption in the paper that 100 percent of the rent is captured by the informed dealers. Only dealers who did not see a trade need to get information; let us assume that they call around to all other dealers and ask to trade, giving their (uninformed) bid-ask spread. If the called dealer is informed, there is a trade, and they get information. If the called dealer is uninformed, there is no trade. At this point, however, there are two types of dealers: the exuninformed, who have learned all the trades, and the ex-informed, who know only their period 0 trade. In a second round of interdealer trade, the exinformed dealers may all call each other to share information or trade with one of the ex-uninformed dealers. While complicated, this story is appealing for two reasons. First, the information is transferred credibly. Second, period 1 volume may far exceed period 0 volume, which agrees with the observation that a large volume of trade is not customer driven. This rent-sharing mechanism also justifies the hypothesis that there is trade in period 1, as assumed in the paper. The interdealers' period 1 trades are otherwise not essential for the main result of the paper. If we did not have
101
Interdealer Trade and Information Flows
period I trades, there would be four types of traders in period 2: uninformed and informed traders and uninformed and informed dealers. Dealers do not know ex ante whether they will be informed or uninformed. However, since they are risk neutral, they will leave the market open in period 0 more often than competitive marketmakers in a centralized market, provided again that they agree to act collusively. The previous discussion shows that the main result of the paper hinges on dealers' collusion. I now look at how the collusion is achieved a bit more closely. The model is designed to mimic a market like the foreign exchange market in which there is no consolidated information on the order flow. Nobody knows whether other trades were executed at bid or ask prices, if at all. The model assumes, however, that customers are served at random. This cuts off the feedback from prices to demand and ensures that there is no incentive to deviate from the dealers' cartel in period O. It is realistic, however, that liquidity traders see the quotes posted by all dealers and prefer to trade at the narrower bid-ask spread. In this case, a dealer has an incentive to post a bidask spread just a bit narrower than her competitors to monopolize information, and in so doing she will break the cartel. The assumption that makes the cartel self-sustaining in the model of Perraudin and Vitale is that dealers cannot attract more customers by offering better prices: the probability of serving a customer is fixed and equal to (lIn). But this assumption is too strong for a market like the foreign exchange market, where there is consolidated information on quotes. Dealers' markets offer services that auction markets cannot offer, such as the certainty of execution of trade, which may be essential to liquidity traders in the foreign exchange market. These are services that make dealers' markets undoubtedly desirable. This paper contains a different argument for the desirability of a decentralized dealers' system: a dealers' market is less prone to market crashes than a centralized one. I have argued, however, that the result does not depend on information sharing. It is essential that dealers agree to collude and act as a monopolist. It is reasonable that in a dealers' market with a small number of players it is easier to collude and agree on rent-sharing rules. But, if we are ready to trade off the market's robustness for liquidity traders' happiness, why not have a single monopolist dealer? A monopolist dealer may be better than a cartel of dealers if there are any costs in monitoring the coalition and sharing the rent.
References Garbade, Kenneth D., Jay L. Pomrenze, and William L. Silber. 1979. On the information content of prices. American Economic Review 69, no. 1:50-59. Glosten, Lawrence R., and Paul R. Milgrom. 1985. Bid, ask and transaction prices in a specialist market with heterogeneously informed traders. Journal of Financial Economics 14:71-100.
102
William Perraudin and Paolo Vitale
Comment
Alan Kirman
This paper is a particularly interesting contribution since it attempts to model certain specific aspects of the microstructure of the foreign exchange market and to explain a particular phenomenon, the large amount of interdealer trading. Many papers on this subject either simply describe the functioning of the market, which is, of course, interesting in itself, or build macro models in which allusion is made to certain features of the microstructure. The latter are thus used to justify rather than to analyze the macroeconomic characteristics. The main line of the early theoretical literature that did analyze the microstructure and that can be identified with Garman (1976) concentrated on how marketmakers would adjust inventories and bid-ask spreads in response to a stream of orders. Later, the problem of asymmetric information became the dominant concern (see, e.g., Hsieh and Kleidon, chap. 2 in this volume), and the current paper fits in this category. In the first part of this comment, I make some specific remarks about the model developed by the authors and in the second suggest other potential modeling strategies to capture the phenomenon in which they are interested. The authors' model can be thought of as one in which bookmakers are faced with a two-horse race. There are a number of experts around who know which horse will win, and the bookmakers know how many of these there are, but not their identity, and have a prior probability as to which horse will win and give odds as a function of this. In the first stage, a bettor arrives and places a bet on one of the horses. This event provides information for the bookmaker, who can adjust his odds, and, since odds are not posted, he can pass this information on to his fellow bookmakers at a price. Thus, a transaction will occur between bookmakers. Bookmakers can now accept further bets from customers, if any are forthcoming. The race is then run, and the whole procedure starts again for the next race. The obvious objection here, and one to which I come back later, is that in the foreign exchange market the "race" is never run, although one could assimilate it to the arrival of some news about which people had prior ideas. Interdealer trading in the model permits the flow of information about customer orders. The fixed horizon keeps the analysis tractable. Leaving this on one side for a moment, two features are striking. First, the position that a dealer holds does not enter into the analysis. It is, of course, often said that many traders, as in this model, do not trade off their positions, but this cannot hold all the time if, as in reality, they are constrained to be in a zero net position at the end of the trading day. One should therefore stick to the authors' interpretation that they are dealing with a very short interval of time, but one that cannot be too near the close of the market. Alan Kirman is professor of economics at the European University Institute, Florence, on leave from the GREQAM (Groupe de Recherche en Economie Quantitative d' Aix Marseille).
103
Interdealer Trade and Information Flows
Second, the probability with which liquidity traders act does not depend on q, which I take to be common knowledge. (In the model, the realization of the
exchange rate is a random variable z that takes on the values 1 and ~ 1 with probabilities q and [1 ~ q].) Indeed, the first customer, however misinformed, could infer the prior q from the bid and ask that he is offered. One would have expected there to be some sensitivity to the probability q in the reaction of liquidity traders. In the model, one dealer receives a trade, and he captures the rent from that trade by doing the smallest transaction size and charging the appropriate price to exact the rent from the purchaser. The relation between the potential profit to be made from receiving a customer and the number of traders is interesting. While it seems reasonable that with a small number of dealers only one would receive a customer in some short period of time, it would seem that as the number of firms rose the number of customers arriving would rise also and therefore that both the probability and the profitability of such an encounter would change. Finally, in the context of this model, agents trade only with each other once an order has been received from an outside customer, but this need not be the case in general. Furthermore, the bids and asks that a new customer faces in period 2, that is, after interdealer trades have taken place, will depend on whether the dealer he meets has just been engaged in such trades. Thus, the sequence of prices will be influenced by this and the properties of the price series further complicated. Incidentally, while the properties of the stochastic price process are of theoretical interest, they are difficult to test precisely because transactions data are not generally available. To tum now to alternative approaches, the most obvious of these is to suggest that dealers holding open positions are aware that they will have to close them by the end of the day in general and will therefore adjust their bids and asks accordingly. This would suggest that an approach based on risk sharing and inventory management (see Lyons 1995) might be appropriate. This is what is suggested by Suvanto (1993) when he says, "Transactions the dealer undertakes in the role of a customer with a market maker are called cover transactions." He also says, "The motive for this kind of transaction, in general, is position adjustment, not trading income as such." Two things feature here: one is the position adjustment because of risk, and the other is adjustment to close the position as the end of day horizon approaches. The horizon problem is thus different from that in Perraudin and Vitale's model and is linked no longer to the arrival of some realization of a random variable but rather to the closing of the market. As I suggested earlier, the "race" in my analogy to their model is never actually run, and for this reason the other sort of horizon seems more plausible. Another problem is that of where information comes from. In reality, a foreign exchange dealer is faced with a continual barrage of information. He sees
104
William Perraudin and Paolo Vitale
screens full of indicative quotes, and he hears the quotes of brokers through loudspeakers as well as observing the electronic broking system. Now, it can be argued that the indicative quotes do not reflect transaction prices clearly, that an actual trade conveys much more information, and that this will change behavior and resultant prices. The difficulty with this is that, having thus obtained theoretical results concerning the characteristics of prices derived from a model of information-generating transactions, it is very difficult to test them since most of the data available correspond to indicative quotes, not to transaction prices. Another observation is that interdealer trading may simply be due to different expectations (see, e.g., Frankel and Froot 1990), which may not be irrational (see, e.g., Kurz 1994) or which may involve agents learning (see Lewis 1989a, 1989b). A last way of looking at exchange rate evolution is as a process in which dealers infer from or are influenced by the actions of others, which leads to "herd behavior" (see, e.g., Banerjee 1992; Kirman 1993; and Sharfstein and Stein 1990) or to "informational cascades" (see Bikhchandani, Hirshleifer, and Welch 1992). Indeed, one can interpret Perraudin and Vitale's contribution as a special case of this type of model, in which one piece of information is passed along sequentially to other dealers. However, in fact what seems to be important is that numbers of dealers are trading with and taking account of the trades of their usual network of partners. How traders react will depend on a combination of their current position and their interpretation of the information contained in a trade. In such a framework the stochastic reactions of the agents mayor may not generate a shift in an exchange rate, but there is not necessarily any fundamental information contained in trades. Thus Perraudin and Vitale view interdealer trading as involving the sale and passage of information contained in orders, while an alternative view developed in Kirman (1995) is that interdealer trading can, of itself, generate exchange rate movements without any exogenous information. In conclusion, the present paper offers an interesting contribution to the literature showing how variations in a particular structure in a model lead to changes in the prices in that model. Whether the aspect that the authors choose-information transmission-is the most important in explaining interdealer trading is an open question, but their contribution provides a way of making a more precise analysis of the question.
References Banerjee, A. 1992. A simple model of herd behavior. Quarterly Journal of Economics 108:797-817. Bikhchandani, S., D. Hirshleifer, and I. Welch. 1992. A theory of fads, fashion, custom, and cultural change as informational cascades. Journal of Political Economy 100, no. 5:992-1026.
105
Interdealer Trade and Information Flows
Frankel, J. A., and K. A. Froot. 1990. The rationality of the foreign exchange rate: Chartists, fundamentalists, and trading in the foreign exchange market. American Economic Review 80:181-85. Garman, M. B. 1976. Market microstructure. Journal of Financial Economics 3:257-75. Kirman, A. P. 1993. Ants, rationality and recruitment. Quarterly Journal of Economics 108 (February): 137-56. Kirman, A. P. 1995. The behaviour of the foreign exchange market. Bank of England Quarterly Bulletin 15 (August): 286-93. Kurz, M. 1994. On the structure and diversity of rational beliefs. Economic Theory 4:877-900. Lewis, K. K. 1989a. Can learning affect exchange-rate behavior? The case of the dollar in the early 1980's. Journal of Monetary Economics 23:79-100. Lewis, K. K. 1989b. Changing beliefs and systematic rational forecast errors with evidence from foreign exchange. American Economic Review 79:621-36. Lyons, R. K. 1995. Tests of microstructural hypotheses in the foreign exchange market. Journal of Financial Economics 39:321-51. Sharfstein, D. S., and 1. C. Stein. 1990. Herd behaviour and investment. American Economic Review 80, no. 3:465-79. Suvanto, A. 1993. Foreign exchange dealing: Essays on the microstructure of the foreign exchange market. Helsinki: Research Institute of the Finnish Economy (ETLA).
4
One Day in June 1993: A Study of the Working of the Reuters 2000-2 Electronic Foreign Exchange Trading System Charles Goodhart, Takatoshi Ito, and Richard Payne
4.1
Introduction
This is a study of foreign exchange dealers' behavior as revealed in the working of Reuters 2000-2, a recently developed electronic foreign exchange trading system. It was launched in 1992 with twenty-three subscriber sites in two countries and by September 1993 had more than 230 dealing sites in twenty-eight cities in seventeen countries (Blitz 1993). The working of the system is described in more detail in section 4.2. This dealing system 2000-2 (henceforward termed D2000-2) is, however, still at the developing rather than a mature stage, and the snapshot that we have of its operations on one dayCharles Goodhart is the Norman Sosnow Professor of Banking and Finance and deputy director of the Financial Markets Group at the London School of Economics. Takatoshi Ito is professor of economics at Hitotsubashi University and senior advisor of the Research Department at the International Monetary Fund. Richard Payne is a Ph.D. student at the London School of Economics and a research assistant at the Financial Markets Group. This lengthy empirical exercise was conducted in a number of stages. After one of the authors, C. Goodhart, had obtained the original videotapes from Reuters, to whom we are most grateful, the data on the tapes were transcribed onto paper by two of the authors' wives, Mrs. Goodhart and Mrs. Ito, assisted by Yoko Miyao, a painstaking task beyond and above the normal requirements of matrimony. The data were then sorted and organized by T. Ito and R. Payne, separately in the United States and the United Kingdom. The graphic appendix is entirely Ito's work. The descriptive material in sections 4.1 and 4.2 was mostly written by Goodhart. The comparison ofD2000-2 and FXFX in section 4.3 had input from all authors, but mostly Goodhart and Payne. The comparable FXFX data were obtained from Olsen and Associates, to whom we are most grateful. Only the first three sections were ready in time for the July Perugia conference, so this is all that our discussants, to whom we are most grateful, then had before them. Section 4.4, completed thereafter, was entirely the work of Goodhart and Payne, with Payne responsible for the econometrics, apart from table 4.16 by Ito. Charles Goodhart and Richard Payne wish to thank the Economic and Social Research Council for financial support. Takatoshi Ito thanks Charles Kramer for technical assistance in producing the graphic appendix.
107
108
Charles Goodhart, Takatoshi Ito, and Richard Payne
16 June 1993-may have become outdated and obsolete by the time that this is published. 1 Reuters has become subject to competition in this marketplace, from Minex and from the Electronic Broking Service (EBS). The former was established in April 1993 by Japanese institutions and, according to Blitz (1993), is "much used in Asia," although, as of September 1993, it did not reveal the number of trades crossed or terminals used. EBS was founded on Wednesday, 21 September 1993. It cost, again according to Blitz, around £40 million to launch and has been backed by a dozen leading banks in foreign exchange-such as Citibank and Chase Manhattan-who formed a consortium with Quotron, an electronic information screen competitor with Reuters. In September 1993, Bob Etherington, Reuters' international marketing manager, would not reveal his dealing system's current volume levels, although Blitz (1993) did report that the "system has reached [its] initial target of 1,000 trades a day, each for a minimum 1 million units of currency dealt." 2 As noted, Minex was not then disclosing the number of trades, and EBS had not started but was going to invite dealers "to trade in standard amounts of $5 million in Drn/$ and £5 million in £/Dm." Such electronic dealing systems (as contrasted with informational pages supplying indicative bid-ask quotes, such as the Reuters FXFX page) are still in their early stages and are highly competitive. Moreover, they may have an important future: "Roughly 60 per cent of deals in the currency market are now done by traders in two banks-or counterparties-who call one another up directly. The remainder of deals are done through brokers, who bring together diverse buyers and sellers.... But they [the banks] complain that the commissions charged for broking a deal are very high. Automated brokerage terminals do the same job as humans at a reduced cost. ... The banks are attracted by the reduced cost of commission. But they fear that 2000-2 will help monopolize the market in electronic dealing systems. Mr. Bartko [chairman of the EBS partnership] admits that this is one of the principal motives for this week's launch of EBS" (Blitz 1993). Electronic trading systems have been in use for rather longer in other financial markets, notably in standardized futures and options markets. Instinet and Globex are two such that Reuters has again been developing. A useful taxonomy of the modus operandi of such electronic trading systems has been provided by Domowitz (1990, 1993). 1. Readers wanting more up-to-date information should refer directly to Reuters Limited, 85 Fleet Street, London EC4P 4AJ, United Kingdom. 2. The total amount thus traded is large in absolute amount but small relative to reported daily turnover in this market of some $900 billion or more. We find it hard to relate the data reported above to the BIS (1993) report in their 1992 survey that, "in the United States and the United Kingdom, the share of deals going through such [automated dealing] systems in April 1992 was 32 and 24% respectively" (table 1, p. 21, and p. 24). Probably definitions of automated dealing systems would have been somewhat wider, including Reuters D2000-1 as well as D2000-2, but, even so, the above percentage seems surprisingly high.
109
A Study of the Reuters D2000-2 Dealing System
Under these circumstances, details of the workings of such systems remain commercially sensitive. The database that we have studied, a videotape of all the entries over D2000-2 for almost exactly seven hours for the deutsche mark/dollar, and some sixteen minutes less for five other bilateral exchange rates, shown on the D2000-2 screen during European business hours on 16 June 1993 (from 08:31:50 to 15:30:00 British Standard Time [BST], i.e., GMT + 1), remains the copyright of Reuters. 3 Anyone wishing to use these data should refer to Reuters, not to us. We should like to emphasize that this videotape did not include, and we have not been given any access to, any information regarding the identity of any of the parties involved in trading; all the trades observed by us remain anonymous. Indeed, it is not possible for any observer, even in Reuters itself, to identify which are the individual banks using the system. Readers should keep in mind the shortcomings of these data. They represent a short snapshot of conditions in a rapidly changing market over a year ago. Trading undertaken over such electronic trading systems may well be, as discussed further below, not representative of the market as a whole; trading activity on D2000-2 on 16 June 1993 may have differed in some respects significantly from that in surrounding days and weeks; the volume and characteristics of electronic trading (over Reuters) in June 1993 may well be quite different from that now since over a year has passed. Given these disclaimers, why should anyone bother to read on? Despite these shortcomings, there are, however, several reasons why this study pr?:vides new insights in the literature of high-frequency exchange rate behavior. First, until now there have been virtually no continuous time-series data available at all on actual trades, prices, and volumes in the foreign exchange market. 4 The 60 percent or so of deals done directly by two bank counterparties over the telephone remain, naturally, private information. There has been little use made of data on foreign exchange transactions intermediated by specialist interbank brokers, no doubt partly because of commercial and confidentiality sensitivities. The only studies currently known to us making use of such data are by Lyons (1995, chap. 5 in this volume). Data of any kind on the characteristics and continuous time-series behavior of actual trading transactions on the foreign exchange market are, therefore, still rare. 5 Second, there have been so few data on transactions in the foreign exchange market that almost all the 3. We are most grateful to Reuters in general and to Mr. Etherington in particular for allowing us to record the quantitative details reported below. 4. There is, of course, the survey of foreign exchange business that has now been undertaken three times at three-year intervals in April 1986, 1989, and 1992 by central banks under the aegis of the Bank for International Settlements (BIS), but this does not provide time-series data. The volumes reported are aggregates for the month of April. 5. We have little doubt that such data will become more plentiful and easily available in the future. But for the time being at least they have rarity value. Also, as electronic trading systems mature, it should be of historical interest to observe how they looked and operated in the early stages of their development.
110
Charles Goodhart, Takatoshi Ito, and Richard Payne
studies on this market have used data on bilateral currency exchange rates that emanate from the indicative bid-ask prices shown on electronic screens by the specialist information providers, for example, Reuters, Telerate, Knight Ridder, and Quotron. There has, naturally, been some concern whether the highfrequency characteristics of such indicative quotes, for example, the negative auto-correlation and the fact that the size of the spread clusters at certain conventional values, are representative of the characteristics of firm (committed) bid-ask quotes at the touch. The touch, a term more commonly used in the United Kingdom than in the United States, is defined as the difference between the best (highest) bid and the lowest ask on offer, where these are (usually) input by different banks. Lyons, for example, expressed such concerns when he wrote, "Some of the shortcomings of the indicative quotes include the following. First, they are not transactable prices. Second, while it is true that the indicated spreads usually bracket actual quoted spreads in the interbank market, they are typically two to three times as wide.... Third, the indications are less likely to bracket true spreads when volatility is highest since there are limits to how frequently the indications can change. And finally, my experience sitting next to dealers at major banks indicates that they pay no attention at all to the current indication; rather, dealers garner most of their high-frequency market information from signals transmitted via intercoms connected to interdealer brokers [see Lyons 1993]. In reality, the main purpose of the indicative quotes is to provide non dealer participants with a gauge of where the interdealer market is trading" (1995, pp. 331-32; see also Flood 1994, esp. n. 6, p. 154). 00, for example, the frequency and volatility of the indicative quotes provide a reasonable proxy for the same characteristics both in the committed bidask quotes and in the associated transactions in the electronic trading systems? We provide an initial answer to such questions in section 4.3, where we seek to compare characteristics of the FXFX time series 6 with those of the 020002 data for the overlapping seven hours. As described in more detail in section 4.3, the 02000-2 series was not time-stamped, and our study of this relation is conditional on the assumptions and techniques used to match these two series temporally. Subject to that condition, and to anticipate some of our main findings in section 4.3, the averages of the bid-ask in both series (FXFX and 02000-2) are almost identical. A graph of the time path for the deutsche mark/dollar from the two sources looks like one line (see figure 4.1). Thus, the time path of the indicative quotes can, on this evidence, be taken as a very good and close proxy for that in the underlying firm series. Nevertheless, some of the characteristics of the bid-ask series, for example, the pattern of autocorrelation, are somewhat different. Even so, both series indicate a somewhat similar GARCH 6. We obtained the accompanying FXFX data series from Dr. M. Dacorogna of Olsen and Associates in Zurich.
111 '."5~_~
A Study of the Reuters D2000-2 Dealing System ,.......
-
1.66
1.655
1.65
TIme IMCOnd$)
Fig. 4.1
Average of bid-ask for FXFX and D2000-2 data: deutsche mark/dollar
pattern. As would be expected, the two series are cointegrated, with the indicative series responding more to deviations from the equilibrium (Le., a larger and more significant negative coefficient on the error correction mechanism). By contrast, the characteristics of the spreads in the FXFX as compared with the touch in D2000-2 are markedly different. The spreads in the FXFX series show clustering among a small number of standard values (e.g., 5, 7, and 10 pips for the deutsche mark/dollar), whereas the spreads at the touch show no such signs of clustering. After examining the relations between the quote series and associated spreads of FXFX and D2000-2 in section 4.3, we turn in section 4.4 to a more detailed study of the characteristics of D2000-2, in particular, the interaction between quotes and transactions in that data set. This long section has five subsections. First, in section 4.4.1, we examine the statistical characteristics of the transaction price series in 02000-2. Whereas for both 02000-2 and FXFX the quote series incorporate a first-order negative moving average, the transaction price data appear to follow a random walk. Our most interesting finding is that the series of runs of deals, sequences of trades at the bid and the ask, is not normally distributed but contains some very long consecutive sequences, another fat-tailed distribution. Second, in section 4.4.2, we examine the interrelations between the available data series, using nine main series from 02000-2, all of which, apart from the spread, can be separately obtained for the bid and the ask. These are the frequency of transactions (deals), their size, and whether such transactions exhausted the quantity currently quoted; then the frequency of quote revision, the change in the quoted prices, and the quantity quoted; and two measures of volatility, the absolute change in the quote and the standard deviation of the quotes. Our main finding is that there is a two-way interrelation between the
112
Charles Goodhart, Takatoshi Ito, and Richard Payne
frequency of quote revisions and the frequency of deals and that, when a deal exhausts the quantity on offer, this then affects (with one-way causality) a nexus of relations between volatility, spreads, and quote revisions. We also conduct similar companion studies on the (temporally associated) FXFX data using a smaller subset of data series (since we have no data on transaction characteristics or on posted quantities from FXFX), but these have less interesting results. Our finding that there is a strong two-way relation between the frequency of quote revisions and that of transactions within a period is, we believe, new, although the underlying cause, that both derive from the arrival of "news," is theoretically straightforward. Most studies of transactions in other asset markets (e.g., the New York Stock Exchange [NYSE]), have used data series calibrated in transaction (tick) time, with the result that one cannot then infer calendar-time frequency. Otherwise, with relatively low-frequency transactions on the NYSE, so many of the observations would exhibit zero change. With much higher-frequency transactions on foreign exchange markets, it seemed to us worthwhile to explore the form of these relations in both clock time and transaction time, although we feel that much remains to be done in clarifying the appropriate econometric usage in this field. Next, in section 4.4.3, we examine the ARCH (autoregressive conditional heteroskedasticity) characteristics of the quote series, in particular to discover whether their GARCH characteristics would be affected by the addition of transactions data. In this case, unlike most of the other main results in section 4.4.2, the results did appear sensitive to whether the exercise was run in clock time or tick time. Largely because much more data have been made available for the equity market, especially the NYSE, and its associated derivative markets, there has been much more empirical work on those markets than for the foreign exchange market. Moreover, the two markets are quite dissimilar in format and microstructure, as nicely described in Bessembinder (1994). Nevertheless, despite the comparatively very small size of our data set, its coverage of transactions as well as quotes brings it somewhat nearer to the richer data sets available on equity markets. In particular, our study here, examining the interaction between trades at the bid and ask and price quote revisions, has some features in common with that of Hasbrouck's (1991) study of such effects in the NYSE. So we then replicate his study as closely as we can, using our own data set and adding some variations of our own. We draw the conclusions of these exercises undertaken earlier in section 4.4 together in the final part, section 4.4.5. Throughout this work, the caveat that our data set lasts for only seven hours, a possibly atypical period, must always be kept in mind, despite the comparatively large number of data points. It is in this sense a very small sample. All our findings, both positive and negative, must be treated with caution.
113
4.2
A Study of the Reuters D2000-2 Dealing System
The Characteristics of D2000-2
Automated brokerage terminals do the same job as humans but at a reduced cost. A bank dealer who is a member of one of these electronic systems can enter her buy and/or sell price into them. Reuters D2000-2 and EBS show only the touch, the highest bid, and the lowest ask; these will normally, but not necessarily, be entered by different banks. This is different from the indicative foreign exchange pages (e.g., FXFX), which show the latest update of the bid and ask entered by a single identified bank. On all the electronic trading systems, the identity of the inputting bank is not shown. The quantity that the inputting bank is prepared to trade is also shown on D2000-2. This was then shown as integers of $1 million, and in some bilateral cases DM 1 million, from 1-5 and entered as M (medium) for a sum between $6 and $10 million and L (large) for sums above $10 million.? More than one bank may input the same best bid (ask) price, in which case the quantity shown is the sum of that offered by these banks. The limit orders, that is, those below the (best) bid and above the (best) ask, and their associated firm quantities are entered and stored in these systems but are not revealed over D2000-2 and EBS. Such reserve limit orders are shown on Minex. Another bank dealer and member of the trading system can then "hit" either the bid or the ask by typing instructions on his own machine. The first check is prudential. Banks in such systems may want to restrict the amount of dealing with certain other counterparties (in some cases refusing to deal at all with some counterparties). The computer first checks whether the deal is prudentially acceptable to both parties (who remain at this stage anonymous). If not, the deal is refused and the "hitter" so informed. We have no information as to how often this might happen, but we surmise that it might be fairly rare. Assuming that the "hit" is accepted and that several banks are offering the same best price, their offers are met on the basis of the time of entry, first in first out. When a new deal is made, the new transaction price enters on the right-hand column of the screen, 8 and there must be an associated change in the quantity of the bid (ask), depending on which is hit,9 and also in the price offered if the size of the deal exhausts the quantity offered at the previous price. In such cases, the bid price must move downwards if there was an exhaustive deal at the bid, and the ask price upward following an exhaustive deal at the ask, or indicate that there are no remaining limit bids (asks) in the systems, that is, no quote shown. 10 Note that, in an automatic system like this, a deal must be made 7. This classification has since been changed. 8. When a new deal has been made, the new transactions price initially for a few seconds shows purple, rather than the standard black, on the screen in order to alert traders to this. 9. When the deal is completed, both banks, the hitter and the quoter, will be sent details regarding to whom and where to make the payment, which is then settled in the standard fashion. So, ex post facto, the identity of the counterparty becomes revealed. 10. Unhappily, we had a few cases in our data where this directional constraint did not hold. While this could be due to new bid-ask inputs occurring at exactly the same moment, several of
114
Charles Goodhart, Takatoshi Ito, and Richard Payne
at either the posted bid or ask and cannot be made at an interior price between them, as can happen with nonautomated human dealers, which can cause problems in empirical studies. This has been a particular problem for empirical studies of the NYSE (see, e.g., Petersen and Fialkowski 1994; and Lee and Ready 1991). 02000-2 allowed traders to deal in some fifteen major bilateral exchange rates at the time of our exercise. The number and range of currencies covered have been changing over time, as is no doubt the case for EBS and Minex as well. The screen for 02000-2 is not big enough to show all fifteen at once, and in any case such a large number of separate rates might be distracting. So the dealer on 02000-2 can call up to six bilateral exchange rate onto the screen at anyone time. All this may be made somewhat easier to follow by seeing an example of what a dealer would see when looking at her screen. This is shown in table 4.1. Note, in particular, that not all the cells have entries. There are periods, especially in the less actively traded bilateral exchange rates, when no bank is making a firm offer. A bilateral currency can have a firm bid (ask) exhibited without there being any corresponding ask (bid) on the screen, as in this example for the deutsche marklFrench franc exchange rate; so there is no observed spread at such times. Any bid-ask price must be associated with an accompanying quantity offered (and vice versa). As electronic trading becomes more popular, such gaps in prices may be expected to become fewer. Note also that the representation of the bilateral exchange rate in the left-hand column is the reverse of what would be normally expected, that is, row 1 would in normal usage be described as the number of yen per dollar. (We thank a discussant for noticing this.) The reason, we understand, for this ordering is that all the volumes are denominated in units of 1 million of the first currency shown. Henceforth, however, we will revert to the standard representation of the bilateral rates. 02000-2 runs throughout the whole day during the week, apart from a short break from 2300 GMT to 0100 GMT. On 16 June 1993, a Reuters employee started to videotape the bilateral deutsche mark/dollar exchange rate at approximately 0830 hours BST. This is the dominant and most active of all exchange rates (see, e.g., Goodhart and Oemos 1990, 1991a, 1991b). About sixteen minutes, thirty seconds later, he also put the additional five bilateral exchange rates that were shown in table 4.1 up onto the screen. 11
these cases probably arise from mistakes in transcribing the videotape (see section 4.2). When we had identified these few errors, we removed them from the data set. 11. Reuters had decided to videotape a day (seven hours) of the working of D2000-2 for their own purposes. We do not know why their operator chose these other five bilateral exchange rates. There is some autocorrelation in volatility and activity in differing rates from day to day, and maybe the operator felt that these would provide either more interest or a better representation than the other nine available. But, basically, we do not know, just as we do not know how the characteristics of the observations in this seven-hour snapshot compared with the same hours on other days, or with other hours on the same day, or with other bilateral rates at the same time.
115
A Study of the Reuters D2000-2 Dealing System
Table 4.1
D2000-2: Screen at 10:17:40 on 16 June 1993
Currency
Bid
Ask
106.16
106.25
USD/JPY DEM/JPY USD/CHF
DEMICHF USDIDEM DEMlFRF
I
1.4672 0.8925 1.6439 3.3633
= U.S. dollar; JPY FRF = French franc.
Note: USD
I
1.4679 0.8933 1.6443 I
Quantity Columns
Blank Columns
Latest Price
2Xl
XX XX XX XX XX XX
106.26 64.59 1.4676 0.8929 1.6443 3.3634
II
42 32 21 MI
= Japanese yen; DEM = deutsche mark; CHF = Swiss franc;
It is this videotape, initially filmed for its own purposes, that Reuters was kind enough to let us observe, subject to confidentiality commitments. There are four Betacam tapes, which ran virtually continuously, subject to a future minor qualification, from 0832 BST to 1530 BST (on 16 June 1993). The screen does not show the clock time, and the entries are not time-stamped, but a time elapse (time passed since the start of videotaping) was entered onto the tape. 12 As might be expected, when the commitments made on screen are firm and deals are made at those prices, the original data are, as far as we can judge, remarkably accurate. We ended with only a couple of data points that we felt must be in error. This compares with errors that occur about once in every four hundred entries over FXFX (see Pictet et al. 1994, table 5). By contrast, we are conscious that there will be a number of transcribing errors. In particular, whether because of the need to copy the tapes or for some other reason, the final digit of the five-digit (in one case four-digit) number was often hard to decipher. In particular, it was difficult to distinguish zero from eight when these were faint on the videotape. 13 In one respect, fortunately, the data are self-checking. When a deal occurs, the transaction price in the right-hand column has to be the same as the prior (i.e., within seconds earlier) bid, or ask, that was hit and must change the quantity offered at that prior price, and also the price itself, should the quantity be fully taken up. The two series (i.e., of transactions prices, on the one hand, and bid-ask prices and their associated quantities, on the other) were transcribed at 12. We were working at Harvard University when we sought to take the details of the tape, every entry, from the video onto paper and then back onto electronic diskette. Since no Betacam video machines were available in the United States, the tapes were first copied onto S-VHS, and the entries on the S-VHS tapes were viewed over a special video player, with adjustable speeds, forward and backward, pause, etc. 13. The transcription from video to paper was primarily done by the wives of two of the authors, Mrs. Margaret Goodhart and Mrs. Keiko Ito, also with the assistance of Ms. Yoko Miyao, who did this extremely complex and difficult exercise in a dedicated, patient, and conscientious fashion, and we are most grateful to them. But there will inevitably be some errors in variables.
116
Charles Goodhart, Takatoshi Ito, and Richard Payne
separate times. By marrying these Up14 and reviewing in cases of errors, we can both cross-check the accuracy of our transaction data and get some idea of the remaining errors in variables for the entries (bid-ask and associated quantities offered) where no such cross-check was possible. 15 Turning now to the data themselves, the database divides into two separate parts. First, there is the deutsche mark/dollar market. This is the dominant exchange rate in the foreign exchange market overall, and its dominance of the electronic market in our snapshot is even more marked. There were 799 bid entries and 823 ask entries (note that these entries would usually come from separate banks). Quantities offered at the bid were entered on 802 occasions and at the ask on 841 occasions. (Note that the quantity offered can, and does, change quite frequently without an associated bid-ask price change. Similarly, the price can change without the associated quantity being altered; this happened on more occasions than we would have expected, perhaps because a bank changed the price for a given amount that it wanted to trade.) Although we cannot possibly deduce the total number of independently made entries, these might conservatively be put at around fifteen hundred in seven hours, or two hundred or so per hour. This compares with some thirty-five hundred entries over FXFX for the deutsche mark/dollar bilateral exchange rate in the same hours, about five hundred per hour. Considering that FXFX represents almost costless advertising and is the most commonly used indicative foreign exchange price screen, this shows just how busy the deutsche mark/dollar market on D2000-2 was during this snapshot. The number of deals in the deutsche mark/dollar was also quite large, relative to the commercial target, reported in section 4.1, of one thousand per day for deals in all fifteen exchange rates. During this snapshot, there were 186 deals done at the bid and 251 at the ask. Whether this ratio of deals to bid-ask entries is high, low, or normal, we cannot tell. We examine whether this ratio varied significantly from half hour to half hour over our data period in section 4.3. The depth of the deutsche mark/dollar on D2000-2 was fairly good, although it can, and no doubt will, improve further. Following a deal that exhausted the 14. There were a couple of cases when we could not marry the two data points, despite several reviews. It is this to which we referred earlier as the only examples of probable errors in the original data. 15. Thus, the cross-check revealed that the accuracy of visually timing the exact moment of an entry on a screen was to within about plus or minus three seconds. From the adjustments and reviews that had to be made to marry the transaction price data with the bid-ask (and associated quantity) data, it may well be that the final digit in the remaining data is incorrect about once every thirty observations and the penultimate digit incorrect once everyone hundred observations. Some of our statistical anomalies, e.g., the few zero and negative spreads and the incorrect direction of price movement following a deal, need to be seen in that context. Such inevitable human error could have been eliminated had the data been available in electronic disk form, but that was not on offer. Moreover, there are some advantages in getting to know the raw data thoroughly before proceeding to econometric testing.
117
A Study of the Reuters D2000-2 Dealing System
quantity offered or the removal of a bid-ask price, most of the time there was another limit order on the computer at a closely related price. Histograms of quantities offered at the bid measured over both frequency and duration of entry are shown in figures 4.2 and 4.3. The histograms for the ask are nearly identical and have been omitted to save space. From these it can be seen that the frequency and length of time during which no bid or ask price is on the screen for the deutsche mark/dollar are both few and brief. Note that the majority of the quantities offered, both at the bid and at the ask, are usually at or below 5. Consequently, the average size of deal here is also low. We cannot estimate it exactly because we cannot see the actual data lying behind M and L. If, however, we take M to be 8 on average and L to be 15, then the average size of deal at the bid was $2.51 million and $2.49 million at the ask, that is, of similar size. A recent paper by Garrett Glass (1994), examining all foreign exchange deals over the Multinet system, puts the aver300 250
~ c ~
0'
• ~
200 150 100 50 0 0
2
4
3
M
Ouantlty
Fig. 4.2
Bid quantity frequency:
d~utsche
mark/dollar
12000 1ססoo
..
8000
8 31
6000
'0
c
4000 2000 0 0
2
3
4
Ouantlty
Fig. 4.3 Bid quantity duration: deutsche mark/dollar
M
118
Charles Goodhart, Takatoshi Ito, and Richard Payne
age size of deals at about $9 million. 16 Be that as it may, it is the case that deals in the deutsche mark/dollar D2000-2 market were, by this standard, unrepresentatively small on average. Why this should have been so, we do not know, but Lyons (chap. 5 in this volume) reports that the average size of deals done through brokers is lower than that of customer deals, and his figure for the size of average broker deals is not that much larger than that shown here. One factor reducing the number/duration of occasions on which there might have been no entry in the deutsche mark/dollar ask series was that a participant, presumably a single bank, kept an off-market ask entry in the computer at 1.6475 when the market was actually running at about 1.6440. When no other entry was better, this was triggered (see fig. 4.4). As the graph shows, the U.S. dollar appreciated sharply thereafter, and the bank involved presumably disposed of its unwanted dollars. In the meantime, however, it represented a nuisance entry for us, distorting the true underlying pattern of the market. No deal was, naturally enough, done at such an off-market price, prior to the occasion of the dollar appreciation. We decided to remove these off-market asks (between observations 250 and 450 on the ask side). We did not remove the few asks at the same price earlier (around the fiftieth observation) since these were not seriously off market (nor did we remove two solitary occasions of offmarket bids at 1.6405). The resulting, adjusted ask series looks as follows, as shown for comparison in figure 4.5. As these charts clearly show, the major events in the foreign exchange market on 16 June 1993 were two brief periods of sharp appreciation in the U.S. dollar, the first lasting from about 1339 BST to about 1345 BST and the second from about 1443 BST to 1445 BST, as indicated by the time-stamp on the FXFX data series. The average price of FXFX quote entries in each minute during the course of these two jumps is shown in table 4.2. The underlying cause, from "news" arrival, of these dollar appreciations against the deutsche mark are clear enough, but their exact timing is difficult to relate to the news items coming over AAMM (the Reuters news page) on that day. The news on that day was "bearish" for Germany and "bullish" for the United States (see table 4.3). Possibly the 1338-1345 BST jump in FXFX could have been triggered by the U.S. housing figures (certainly the dollar opened firm in the United States) and the 1442-1445 BST jump by a delayed reaction to the German government report, but such links cannot be firmly established. The finding here is consistent with other findings in the literature that tend to experience difficulties matching news events to jumps in the asset price, and vice versa. Nevertheless, one can hardly query the time-stamp on the FXFX data, and the extent and timing of these jumps are very closely matched by the data on D2000-2, as will be discussed further below. One inter16. Considering that deal size is highly skewed, we wonder whether he meant median when he wrote average here.
A Study of the Reuters D2000-2 Dealing System
119
1.665
....
1.66
0
1.655
...
1.65
:s
0
~
1.645 1.64
.... .... .... '" .... '" '" ~'" '" '"'" '"'"'" '" .... '"'" '"'"'"
0
0
00
N
00
N
0
N
N
0
00
~
~
0 0 0 00 .... a; .... '" .... ....0 '"~ '"00'" '"'"00 00'" .... §~ '" '" '" '" ~ N N ~ ~ ~ ~ ~ ~
Time (seconds)
0 .... '" 00 00 N '" '"N
8 '"'".... '" '" '" 00
N N
N N
N
Fig. 4.4 D2000-2 ask data: deutsche mark/dollar 1.665 1.66
..... 0
1.655
:s
... 0
VI
CI:
1.65 1.645 1.64 0
'" '" '" '" '" ~'" '"'" '"'" '"'" '"~ '"'"'"
00
N
00
N
Fig. 4.5
N
~
0
00
'" '"'" '"'" lS '" '" ~ '" '" .... '"~ 00
N
0
00
00
N
~
0
~
~
~
N
N
'" '" 00
~ ~ Time (seconds) ~
~
0 0
N
00 .... N 00 '" ;D ;D .... '" '" '"N '"N0 ....N ~ '" '" 0 N N N N '"
N
'"'" N
Filtered D2000-2 ask data: deutsche mark/dollar
esting feature of these jumps in the value of the dollar is that they were associated with great activity on the ask side of the market and very little action, even in the guise of price revisions, on the bid. From 1337 BST to 1345 BST, there were seventeen deals at the ask in D2000-2 and none at the bid. Over the same period, there were some thirty-nine price revisions at the ask and thirteen at the bid, two of these remaining established and unchanged for almost two minutes each. From 1442 BST to 1446 BST, there were thirteen deals at the ask and none at the bid. There were some twenty-six price revisions at the ask. A few seconds after the start of the dollar appreciation, the existing bid price was removed from the screen, and for the remaining three and a half minutes of the appreciation no bid price at all was posted; this was the longest gap in having a firm price set for either the bid or the ask in our data set for the deutsche mark/dollar. Otherwise, price setting in the deutsche mark/dollar over D2000-2 was nearly continuous. A graphic representation of the bid-ask prices quoted, the occasion and price
120
Charles Goodhart, Takatoshi Ito, and Richard Payne Periods of Appreciation of the U.S. dollar
Table 4.2
BST
Bid
Ask
Number of Observations
13:38 13.39 13:40 13:41 13:42 13:43 13:44 13:45 13:46
1.6474 1.6486 1.6494 1.6500 1.6503 1.6525 1.6540 1.6553 1.6552
1.6481 1.6494 1.6503 1.6506 1.6512 1.6535 1.6550 1.6564 1.6560
7 6 6 5 6 7 6 7 7
14:42 14:43 14:44 14:45 14:46
1.6571 1.6575 1.6594 1.6600 1.6601
1.6580 1.6584 1.6602 1.6606 1.6611
8 8 5 5 5
News on U.S. and German Economies
Table 4.3 BST 12:13:18 12:34:40 12:53: 12 13:01 :44 13:32:04 13:37:04 13:46:54 13:56:30 14: 15:40 14:20: 12 14:32:04 14:33:48 14:41 :08
AAMM Report "German unemployment could top 4 million-Rexrodt" "Next Bundesbank rate cut seen most likely in July" "German industry says economy still declining" "German institute sees no recovery before mid-1994" "US May Housing Starts rose 2.4%" "US Home Building in May is strongest in 5 months" "Bonn can live with current mark-dollar rate" "German Govt source sees no danger for mark" "Dlr opens firm in US, surges on German comments" "US May Industrial Output rose, capacity use steady" "Bonn wants lower short-term rates-Source" "US May Housing Starts Rise is modest-analysts say" "Mark falls against dollar after govt comments"
of deals, and the quantities offered for the deutsche mark/dollar, for the first through the seventh hour, is shown in appendix figures 4A.1-4A.22. Such continuous price setting was not the case for the other five bilateral exchange rates exhibited on the screen during our seven-hour snapshot. Simple observation revealed that market activity in these rates on D2000-2 over our data period was far more patchy. Initially, the rates were not put onto the screen for some sixteen minutes after the deutsche mark/dollar was shown. Thereafter, during the following six and three-quarter hours, there were in some cases quite long gaps in setting bid-ask prices. The average quantities dealt ranged from just over $1 million (Swiss franc/dollar bid) to nearly $3 million (French
121
A Study of the Reuters D2000-2 Dealing System
franc/deutsche mark ask). The data are shown in table 4.4; the figures in parentheses in the table report the original average deutsche mark size when the deals were done in units of deutsche mark 1 million. Deals were, however, much fewer in number than for deutsche mark/dollar. When there are large price movements, the majority of the deals seem to be purchases of the appreciating currency, and the majority of quotes are on the strong side of the market (see table 4.5). We pursue this effect somewhat further in section 4.4 below. Data on these deals and the number of bid-ask price entries are given in table 4.4, and histograms of the bid quantities offered are shown in figures 4.6-4.10; again, the similar ask histograms are omitted to save space. These histograms show differing patterns. The quantities offered on the dollar-based bilaterals (i.e., deutsche mark/dollar, yen/dollar, Swiss franc/dollar) are predominantly for one or two units, with increasingly few offers made as
Table 4.4
Analysis of Deals and Quotes Number of Quotes Bid
Ask
Bid
Ask
93 99
127 54
12 15
17 2
Swiss franc/dollar Swiss franc/deutsche mark
142 121
134 168
18 19
33 45
French franc/deutsche mark
98
79
14
11
Yen/dollar Yen/deutsche mark
Average Size of Dealsa
Number of Deals
Bid 2.33 1.91 (3.15) 1.125 1.26 (2.08) 2.71 (4.45)
Ask 1.55 2.1~
(3.50) 1.67 2.71 (4.45) 2.97 (4.88)
Note: Figures in parentheses report the original average deutsche mark size when the deals were
done in units of DM 1 million. aBased on the assumption that M = 8, L = 15.
Table 4.5
Relation between Direction of Deals and Currency Change Number of Deals
Deutsche mark/dollar Yen/dollar Yen/deutsche mark Swiss franc/dollar Swiss franc/deutsche mark French franc/deutsche mark
Currency Value
Bid
Ask
Start
Finish
% Change
186 12 15 18 19 14
251 17 2 33 45 11
1.6450 106.25 64.63 1.4690 .8935 3.3648
1.6585 106.70 64.27 1.4840 .8953 3.3623
+.82 +.42 -.56 +1.02 +.20 -.07
30 25
~
20
r::
~
15
~
10
0-
o 2
4
QuantIty
M
123
A Study of the Reuters D2000-2 Dealing System 45 40 35
r; c
1!:
30 25
20 15
10
o 2
4
M
QuantIty
Fig. 4.9
Bid quantity frequency: deutsche marklFrench franc
50 45 40 35
~
!:
1
30 25 20 15
10 5
o 2
4
M
Quantity
Fig.4.10 Bid quantity frequency: deutsche mark/Swiss franc
size increases. The quantities offered on the deutsche mark-based bilaterals (i.e., yen/deutsche mark, Swiss franc/deutsche mark, and French franc/ deutsche mark) show many more (proportionately) larger offers, quite remarkably so for the French franc/deutsche mark (fig. 4.9). One possible explanation is as follows. Suppose that the European cross-rates tend to move less than the dollar-based bilaterals; then the risk involved in building up inventories for a dealer is less. Hence, a larger unit bid is offered. Now, the Swiss franc/deutsche mark and French franc/deutsche mark rates should move less than the correspondent currencies vis-a.-vis the dollar because the deutsche mark and the French franc are in the exchange rate mechanism (ERM), and the Swiss franc closely follows the deutsche mark historically. Even the yen/deutsche mark volatility tends to be less than in the yen/dollar or deutsche mark/dollar rates. We should again stress that we have no means of knowing whether these, somewhat patchy, results were representative of activity in these exchange rates at other times of the day (note that activity in the yen/dollar exchange rate might be expected to be somewhat muted in European market space) or on
124
Charles Goodhart, Takatoshi Ito, and Richard Payne
other days or whether they would have been representative of the nine other unshown bilateral exchange rates. Moreover, the use of electronic market systems is developing rapidly over time. Be that as it may, the somewhat occasional nature of the market, then, in these other five exchange rates means that we will concentrate most of our econometric studies on the deutsche mark/ dollar.
4.3
Comparison of FXFX and D2000-2
As described in the introduction, indicative screen prices, as provided over FXFX, constitute the basis for almost all current time-series studies of the foreign exchange market. While there is no doubt that these are close enough approximations to the underlying firm quotes for low-frequency studies (e.g., frequencies of one hour or longer), concern has been expressed as to whether they do necessarily provide sufficiently close approximations to the underlying firm data for very high-frequency studies. For example, Baillie and Bollerslev (1991) have conjectured that the negative moving average (MA) characteristics found in FXFX ultra-high-frequency data may be a facet of their indicative nature and that the underlying price(s) would not exhibit this characteristic (see also Zhou 1992; Bollerslev and Domowitz 1993; and Bollerslev and Melvin 1994). Now that we have a seven-hour snapshot of firm prices in D2000-2, we can, in principle, make a comparison of them with the bid-ask series from FXFX over the overlapping period for the three data sets deutsche mark/dollar, yen/ dollar, and Swiss franc/dollar. A problem, however, is that the D2000-2 data series is not time-stamped, although it does have a time elapse shown on the videotape. In practice, of course, the two series can be matched pretty closely by eye alone by matching the two occasions of short-term appreciation in the deutsche mark/dollar. To try to match the series even more closely, we constructed artificial series for both the D2000-2 and the FXFX deutsche mark/dollar, bid and ask, with observations evenly spaced every five seconds. (Note that in both cases the original series is irregularly timed and hence cannot be directly correlated.) We assumed, for the purpose of matching (D2000-2 and FXFX) only, that the existing price held until revised, for the purpose of interpolation, where necessary. When no price was exhibited on D2000-2, we treated the prior price as still holding, except for the gap in the bid price in the second jump, discussed in the preceding section, where we applied a linear interpolation (between 1.6565 and 1.6590).17 Alternative rules of thumb for interpolation could have
17. When we subsequently used this series for econometric work, we changed this rule of thumb so that, when a deal exhausted the quantity offered and no price was then shown, we took the next reported price as becoming effective. Otherwise, the estimated (absolute) price change, following a deal, would have been biased downward.
125
A Study of the Reuters D2000-2 Dealing System
been tried, but we are confident that doing so would have made no difference for this timing exercise. Our crucial assumption is that price adjustments on FXFX and D2000-2 would be synchronous. We believe that to be justified. Studies made by one of us (Goodhart 1989) of the reaction of FXFX bid and ask prices to precisely timed news announcements (e.g., U.S. "news" released at 0830 EST) show that these are virtually instantaneous (a few seconds at most), and we should surely expect no slower reaction where prices represent firm commitments (see, e.g., Ederington and Lee 1993). Accordingly, our strategy was to assume that prices in both series would move synchronously. Given this assumption, our approach was to compare the correlation of the two series for the deutsche mark/dollar as we varied their temporal overlap and see which temporal overlap gave the best fit. In practice, all the exchange rate action came in the second half of our data period (the last two tapes), and the market was so fiat in the opening hours (tapes) that we could not find any clear peak in the fit when starting from the front. We therefore worked from the back, fitting the final tape to the FXFX data, to the front. In the event, and slightly disturbingly, we found a twentysecond discrepancy between our best-fit timing for the comparison of the bid and the ask series (see table 4.6). However, given our exact knowledge of how the bid and ask series are timed relative to each other on D2000-2, we overrode this apparent discrepancy from the time-series fitting exercise and split the difference between the two with the result that the observations on D2000-2 are all properly aligned with each other. This then gave us the basis for comparison of the D2000-2 bid-ask series with the FXFX series over a closely matched data period (with the exact match uncertain by some fraction of a minute). We have to be careful, however, in using the interpolated five-second series themselves in econometric comparisons since the interpolations distort some of the characteristics of the raw data. There were some eight hundred observations in the basic D2000-2 series and about five thousand in the interpolated series for D2000-2. By construction, the extra forty-two hundred observations will exhibit no change, which must tend to drive any estimated autocorrelation toward zero and may also bias the Table 4.6
BST: Best Estimated Start Times for Tapes
Tape 4 Tape 3 Tape 2 Tape 1
Bid
Ask
13:40:47 11:59:10 10: 15:37 8:31:40
13:41:07 11 :59:30 10: 15:57 8:32:00
Note: In each case, the finish of tape t - 1 was about one second before the start of tape t. For
tape 1, the start time is given from the first quote, of deutsche mark/dollar bid and ask: the tape starts with a blank screen almost exactly eight minutes before.
126
Charles Goodhart, Takatoshi Ito, and Richard Payne
ARCH characteristics. We discuss some of the issues raised by the question of whether to scale the series by time or by tick activity at greater length in section 4.4 below. Subject to that condition, the means of the bid-ask in both series (FXFX and 02000-2) are almost identical. A graph of the time path for the deutsche mark! dollar from the two sources looks like one line (see fig. 4.1). Thus, the time path of the indicative quotes can, on this evidence, be taken as a very good and close proxy for that in the underlying firm series. Nevertheless, some of the characteristics of the bid-ask series (e.g., the pattern of autocorrelation) are somewhat different. Even so, both series indicate a somewhat similar GARCH pattern. As would be expected, the two series are cointegrated, with the indicative series responding more to deviations from the equilibrium (i.e., a larger and more significant negative coefficient on the error correction mechanism). By contrast, the characteristics of the spreads in the FXFX as compared with the touch in 02000-2 are markedly different. The spreads in the FXFX series show clustering among a small number of standard values (e.g., 5, 7, and 10 pips for the deutsche mark/dollar), whereas the spreads at the touch show no such signs of clustering. The basic characteristics of the temporally matched, filtered (but not interpolated) series are shown in table 4.7. The main pattern of results shows that the 02000-2 and the FXFX raw series are, in general, remarkably similar for the deutsche mark/dollar. 18 The differences between the first four moments of the various price series (bid, ask, and average of the bid and ask) in either levels or first differences are minor. The FXFX series in levels have a somewhat lower average value (probably owing to a larger proportion of their observations coming in the earlier part of the period; see table 4.7), an insignificantly lower volatility (standard deviation), and marginally higher skewness and kurtosis. The FXFX series in first differences have lower means, by a factor of one and a half in the mean and about two or three in the bid and ask (perhaps again because of more observations when little was happening in the early part of the period). These FXFX differenced series have a lower skewness and a slightly lower kurtosis. There is, however, a more marked difference in the autocorrelation data. The FXFX series exhibit stronger negative autocorrelation in all cases and at all lags, particularly after the first lag. This is least marked at the first lag of the bid and ask series, where the 02000-2 coefficient is about -0.61 compared with values of -0.62 (bid) and -0.67 (ask) for the FXFX series. In the average series, the first lag value for 02000-2 drops to -0.37, compared with -0.61 for FXFX. After the first lag, the absolute size of the negative coefficients, and of the t-values, drops much more rapidly for 02000-2 than for FXFX. The first 18. At some future date, we intend to construct similar tables for the raw data for the yen/dollar and Swiss franc/dollar on D2000-2 and FXFX, temporally matched. Time did not allow us to do so at this stage.
Table 4.7
Statistical Characteristics of the D2000..2 and FXFX Time Series Compared (deutsche mark/dollar)
D2000-2 1. Bid, number of observations: a Mean SD Skew Kurtosis 2. Difference of bid: Mean SD Skew Kurtosis Autocorrelation coefficients:
1 2 3 4 5 GARCH:c Ao Al Bo BI B2
3. Average of bid-ask, number of observations: Mean SD Skew Kurtosis 4. Difference of average: Mean SD Skew Kurtosis Autocorrelation coefficients:
1 2
(continued)
799 1.649007 .006060 .63670 -1.31504 798 .00000994 .000389 .57095 9.35931
FXFX
3,484 1.6482 .0058 .9392 2.1507 3,483 .000003646 .0004012 .0845 6.393
-.6173 (-17.3)b -.1437 (-3.44) - .1105 (-2.63) .0031 (.07) .0758 (2.13)
-.6236 (-36.77) -.3488 (-17.49) -.1917 (-9.32) -.0802 (-4.02) -.0365 (-2.16)
-.000 (-1.48) -.514 (-16.97) .000 (3.89) .198 (3.92) .728 (14.07) 1,581 d 1.649511 .006052 .55846 -1.40521 1,580 .00000515 .000192 .45549 13.3980
-.000 (- .22) -.481 (-31.92) .000 (4.10) .116 (6.96) .849 (38.14) 3,484 1.6486 .0058 .9400 2.1503 3,483 .000003646 .000371 .0920 9.1457
-.366 (-14.52) -.169 (-6.32) -.109 (-4.06)
-.6094 (-35.91) -.3278 (-16.51) -.1659 (-8.12)
Table 4.7
(continued) D2000-2
4 5 GARCH:e Ao Al
Bo BI B2 5. Spread, number of observations: Mean SD Skew Kurtosis Autocorrelation coefficients: 1 2 3 4 5 GARCH: Ao Al
Bo B1 B2
-.082 ( -3.08) -.043 (-1.72) .000 (3.04) -.179 (-9.19) .000 (.23) .536 (38.93) .540 (89.49) 1,556 6.8464 8.0955 4.034 27.063 .4686 (18.44) .1098 (3.91) .1322 (4.72) -.0027 (-.09) .0500 (1.97) 1.4778 (11.18) .6890 (29.68) 1.0234 (4.30) .6591 (43.67) .6454 (43.84)
FXFX
-.0586 (-2.56) -.0045 (- .26) .000 (1.56) -.026 (-1.40) .000 (6.75) .268 (9.01) .621 (16.05) 3,484 7.090 2.689 2.604 39.380 - .0118 (- .70) .0173 (1.02) .047 (2.81) .042 (2.49) .044 (2.58) .006 (185.8) .032 (5.58) .000 (.70) .287 (77.40) .643 (247.02)
Note: t-values are given in parentheses. aSince the results for the ask series are almost identical to those for the bid, we have omitted the former to save space. bThis is the first difference of the level. eWe ran the system ~x, = a o + al~x'_1 + s" s, I ['-I - N(O, h,), h, = ~o + ~lh'_1 + ~2S2,. dSince the bids and the asks were put in at separate times, the numbers of calculated means and spreads will be approximately equal to the sum of the number of bids plus the number of asks.
129
A Study of the Reuters D2000-2 Dealing System
Table 4.8
Error Correction Mechanism FXFX Dependent
Lagged: -1 Dependent: -2 -3 -4 -5
D2000-2
Coefficient
t-Value
Coefficient
t-Value
-.207
-12.8
-.009
-6.28
-.184 -.136 .002 -.001
-11.4 -8.7 -1.9 -.9
-.004 -.002 -.001 -.007
-2.64 -1.50 -.93 -4.81
Lagged: -1 Independent: -2 -3 -4 -5
-.107
-4.25
-.002
-2.25
-.004 -.000 -.003 -.004
-1.95 -.10 -1.48 -1.54
-.002 -.002 .001 .001
-1.92 -1.73 .64 .64
ECM
-.180
-15.47
-.006
-8.44
four lags in FXFX in each case have significant negative coefficients. This is so only for the averaged series of D2000-2, and the sum of the negative coefficients is always considerably greater in absolute size than -1 for FXFX, whereas it is between -0.75 and -0.90 for D2000-2. We find relatively little difference in the GARCH data, which approximate to IGARCH values, except that the FXFX series for the changes in the average and the level of the spread show-less persistence of volatility (a lower B 1 coefficient) than the D2000-2 series. One of the main findings about the characteristics of the continuous-time foreign exchange indicative quote series was that they appeared to have a negative moving average component. One supposition was that this could be due to the fact that they were indicative, not firm, quotes. Now that we can observe the firm quotes, the negative moving average does appear somewhat attenuated, especially for the average of the bid and ask, but it remains a highly significant feature of the time series. The main difference between the two series occurs in the case of spreads. The most distinctive difference relates to the numerical pattern of the spread, with the FXFX data showing the spread clustering around certain conventional values,19 while the D2000-2 spreads, being at the touch with the bid and ask prices being input usually by different banks, show no such clustering. Histograms of the frequency of spreads at various sizes for D2000-2 and FXFX are 19. This has been widely noted (e.g., Bessembinder 1994; and Bollerslev and Melvin 1994) and was more extensively described and analyzed in Goodhart and Curcio (1991).
130
Charles Goodhart, Takatoshi Ito, and Richard Payne 200 180 160 140
~
120
!
100
c
f
80 60 40 20 0 0
~
N
Spread (basIs poInts)
Fig. 4.11
Deutsche mark/dollar spread frequency: D2000-2 data
2000 1800 1600 1400
~ 1200
1
1000
i
800 600 400 200
o spread (basis poInts)
Fig. 4.12 Deutsche mark/dollar spread frequency: FXFX data
shown for deutsche mark/dollar in figures 4.11 and 4.12. The yen/dollar and Swiss franc/dollar charts, which show almost identical patterns, are available from the authors. One feature of the deutsche mark/dollar spreads in D2000-2 (fig. 4.11) is that there are a number of occasions of zero spread; that is, the best bid and the best ask are equal. In FXFX, when the quotes are input by the same bank, a zero spread would signal an input error. 20 These comparative tables possibly understate the extent to which the two quote series actually do move together. As shown in figure 4.1, when the two 20. Cohen et al. (1981) have persuasively argued that a dealer should always prefer to transact with certainty at a firm bid (ask) quote rather than set an ask (bid) quote at a zero, or tiny, spread distance from it with no immediate certainty of transaction, so on these grounds a zero spread in D2000-2 may also represent a transcription error or a dealer error; indeed, most of these occasions lasted for only a very few seconds. Nevertheless, we intend to discuss with practitioners whether there may be any rationale for the existence of zero spreads on D2000-2, e.g., asymmetric trading (execution) costs between the two sides, and, until we have done so, we have decided to let these data stand.
131
A Study of the Reuters D2000-2 Dealing System
Table 4.9
Regressions between FXFX and D2000-2 Series
Left-Hand-fRightHand-Side Variables FXFX mean/2000 mean 2000 meanlFXFX mean FXFX bid/2000 bid 2000 bidlFXFX bid FXFX ask/2000 ask 2000 askIFXFX ask
Constant
Coefficient on:
R2 (SE)
-.0018 (.0016) .0101 (.0016) -.0267 (.0022) .0397 (.0021) .0315 (.0021) -.0175 (.0022)
1.0011 (.0010) .9938 (.0010) 1.0162 (.0013) .9759 (.0013) .9810 (.0013) 1.0105 (.0013)
.995 (.0004) .995 (.0004) .992 (.0005) .992 (.0005) .991 (.0005) .991 (.0005)
Dickey-Fuller t-Statistic a -18.07 -18.04 -16.16 -16.16 -17.12 -17.11
Note: Standard errors are given in parentheses. aMacKinnon critical 1 percent value -3.896.
interpolated series are drawn on the same graph, there appears to be only one line. If we regress the two interpolated series for the deutsche mark/dollar together, after temporal matching, we get the results in table 4.9. As can be seen, the respective series, for the average of the bid-ask and the bids and asks separately, are all strongly cointegrated (as should be expected). Only in one case, however, when the average of the interpolated FXFX series is regressed on·the average of the 2000-2 series, do the coefficients take on their ex ante expected values with a constant insignificantly different from zero and the coefficient on the right-hand-side variable insignificantly different from unity. Otherwise, the constants are all significantly different from zero, with the D2000-2 bid on average just above and its ask Just below that on the FXFX series. As might be expected, the D2000-2 bid is slightly less variable than its FXFX equivalent, while the D2000-2 ask is a tiny bit more variable (perhaps a reflection of our treatment of outliers in the data?). Such a finding of strong cointegration enables us, always subject to our prior assumption that the two series are synchronous and our temporal matching procedure valid, to examine short-term dynamics and whether a deviation between the two series is corrected primarily by a shift in the FXFX series or in the D2000-2 series. Our hypothesis is that, since the D2000-2 series is the underlying firm series, the indicative FXFX series should adjust to it, rather than vice versa. When, therefore, examining the error correction mechanism (ECM), we expect a large, significant negative coefficient on the ECM when the change in FXFX prices is the left-hand-side variable and a much smaller, possibly insignificant coefficient when the change in D2000-2 prices is the left-hand-side variable. The ECM is taken, as appropriate, from the residuals of the equations in table 4.9.
132
Charles Goodhart, Takatoshi Ito, and Richard Payne
Taking the average of the bid-asks as our example (the results will not change much for the bid or ask series individually), we ran regressions, as follows:
L1 average series 1t
=
j(lags L1 average series 1, lags L1 average series 2, ECM).
The results can be seen in table 4.8. As expected, both the ECM and the effect of prior changes in the underlying D2000-2 series on the FXFX series are more strongly pronounced than the effect of the FXFX series, or the ECM, on the D2000-2 series, although the latter is still clearly significant, despite being much smaller. 21 Since time series on transactions (i.e., the number and value of deals) have not been available for the foreign exchange market, variations in either the frequency of entry or the volatility of indicative prices, or some combination of both, have often been taken as a proxy for the volume of unobservable transactions. Here, we examine whether this may have been a good proxy.22 Since we cannot, however, compare the profile of D2000-2 and total market transactions, we will proceed on the presumption that the former may be a good proxy for the latter. For this exercise, we divide our data period into half hours for the deutsche mark/dollar series. We take these periods from the start, with the result that the final period is not quite a complete period. Then we compare both the frequency and the size of deals in each half-hour period (as a percentage of the total) as compared with the frequency of quote entry (as a percentage of the overall number) and relative volatility (the standard deviation of the average of the bid-ask in the subperiod divided by the overall standard deviation). We also examine how the average size of spread related to these variables. The basic results for the D2000-2 and FXFX variables are given in table 4.10. Then simple regressions between these variables were run, as shown in table 4.11. The results are disappointing for those who would use the indicative FXFX data as a proxy to infer the underlying transactions series. The FXFX volatility series is an excellent predictor of the volatility in the firm quotes of 02000-2 (e.g., [4] in table 4.11); the spread series of FXFX is a mediocre predictor of the spreads on 2000-2, with the latter in this case being on average lower, but much more variable, by a factor of nearly five (cf. rows SS and FS in table 4.10, and see eq. [8] in table 4.11). This must raise some doubts about certain
21. Note that the coefficients will, however, be biased downward by the interpolation process, forcing the interpolations to take a no-change value. The t-values will be less affected by such time deformation. 22. We cannot, of course, yet observe any time series of total market transactions. All we have now is a short snapshot of data on transactions over D2000-2. If the temporal profile of transactions over D2000-2 should be an inaccurate and biased proxy for the total volume of transactions, then the question of whether the indicative FXFX data provide a good predictor of concurrent D20002 deals would not have much importance.
8.4
.05 4.04
7.68
8.3
.083 6.41
8.04
6.65
6.36
.070
1.64
1.42
.110
7.5
9.8
2
6.83
.061
7.78
.05 5.34
6.60
.057 6.89
.062 7.15
.042
7.02
7.02
7.95
.033 4.41
.066 5.15
.041 4.32
3.7
1.25
5.7
6
6.6
2.00
4.8
5
7.4
2.14
1.41
5.4
6.7
4
4.5
3
Deutsche MarkIDollar: Half-Hour Periods
aAs percentage of total. bDivided by volatility of whole period.
S2 deals a (SD) S2 deals, average size (SD) S2 frequencya price entry (SF) S2 volatilityb (SV) S2 spread (SS) FXFX frequencya (FF) FXFX volatilityb (FV) FXFX spread (FS)
Table 4.10
7.32
.079
5.09
.007 2.51
3.6
2.31
3.4
7
7.05
.091
6.42
.10 5.95
7.15
.059
6.48
.05 4.07
7.08
.058
7.92
.696
7.36
.152
7.91
7.32 7.38
.132 8.43
.612 17.5
.074 4.82
11.2
1.42
11.7
12
11.5
4.7
4.1
1.96
10.75
11
5.9
1.83
2.04
1.625
10 6.2
9 5.5
4.1
8
7.80
.374
7.25
.314 7.17
11.45
1.57
11.7
13
7.30
.140
6.65
.116 8.33
7.65
2.24
7.3
14
134
Charles Goodhart, Takatoshi Ito, and Richard Payne
Table 4.11
Deutsche MarkIDollar: Half-Hour Relations, D2000-2 and FXFX
Left-Hand-/Right-Hand-Side Variables (1) SFIFF (2) SDIFF
(3) SD/SF (4) SVIFV
(5) SD/SV (6) SD/SV, SF (7) SDIFF, FV (8) SSIFS (9) SS/SV (10) FSIFV (11) SDIFF, FV, FS
(12) SD/SF, SV, SS
Constant
bi
-6.8 (-1.1) -6.9 (-1.1) .4 (.8) -.0 (-.7) 5.7 (7.14) .3 (.3) -6.9 (- .15) -31.1 ( -2.3) 3.7 (7.7) 6.8 (64.4) -33.0 ( -1.9) .3 (.2)
1.95 (2.32) 1.96 (2.32) .94 (9.12) .88 (27.11) 11.30 (3.95) -.46 (- .17) 1.78 (2.83) 5.27 (2.80) 21.23 (8.67) 1.79 (3.82) 2.45 (3.34) .96 (6.06)
bz
b3
RZ .25 .25 .86 .98 .36
.96 (6.41) 9.30 (3.33)
.85 .59 .34 .85 .51
3.48 (.75) -.64 (-.11)
3.11 (1.54) .01 (.03)
.64 .84
Note: t-statistics are given in parentheses. Initial F stands for FXFX series; initial S for System D2000-2. Second letter F represents frequency of quote entry; D is number of deals; V is volatility; and S is spread. So SF is frequency of quote entry over System D2000-2; FF is frequency of quote entry over FXFX; SD is the number of deals on D2000-2; SV is the volatility of D2000-2, etc.
aspects of the results of recent empirical studies based on FXFX data (e.g., Bollerslev and Melvin 1994; and Bessembinder 1994). This is discussed further in section 4.4.2. The frequency of quotes series on FXFX was a relatively poor predictor of the quote frequency on D2000-2. Unfortunately, the importance of these series as a predictor of deals is largely in reverse. order in this data set. As can be seen (eqq. [3], [6], and [12]), the frequency of quote entries over D2000-2 is the dominant predictor of the number of deal entries, with neither volatility (whose coefficient was even wrong signed) nor spreads being significant. But FXFX entry frequency is a poor predictor of 02000-2 quote entry frequency. Thus, using FXFX data to predict the number of 02000-2 deals was not very successful. The frequency of entry (FXFX) was the most significant variable for predicting 02000-2 deals of the data series available over FXFX (eqq. [2], [7], and [11]), but both FXFX volatility and spreads made some positive contribution. We are fully aware of the small size of this sample among many dimensions, length of time, number of observations, etc.
135
A Study of the Reuters D2000-2 Dealing System
While more work is undoubtedly needed, we must warn that this preliminary exercise suggests that it would be dubious to try to infer transaction frequency from the more widely available FXFX indicative quote data. 23 To sum up, in this section we have sought to compare the characteristics of the D2000-2 and FXFX series over a temporally matched period. The main result is that the time paths for the prices quoted over the two series are extremely close and that most of the time-series characteristics of the two quote series are closely similar. The negative autocorrelation is somewhat attenuated, expected, the distribut still highly significant, in the firm D2000-2 series. bution of spreads is markedly different between the indicative series, which clusters at certain round numbers, and the touch with a much more even distribution. The size of spreads and the frequency of quote entry showed much more variation between subperiods in the D2000-2 series than in the FXFX, and the latter were not good predictors of their D2000-2 counterparts, unlike FXFX volatility, which like its mean value matched D2000-2 almost exactly. This meant that the FXFX data proved to be poor predictors of the frequency of deals over D2000-2 for the deutsche mark/dollar since this was most closely associated with the frequency of quote entries in that same data set.
As
4.4 The Interaction of Transactions and Bid-Ask Quotes on the Foreign Exchange Market 4.4.1
Characteristics of Transactions Data
In the preceding section, we asked how accurate a proxy the commonly available FXFX data were to the underlying firm D2000-2 quotes (excellent as a guide to price movements) and to the spreads and number of underlying transactions over the same data set (which suggested that a lot of caution would be needed). In this section, we test certain hypotheses about the determinants of the occurrence and size of such transactions and their effect in turn on quote revision. We concentrate solely on the deutsche mark/dollar series here because only in this series are there sufficient data points. Our first hypothesis is that the time series for transactions prices (returns) will be random walk. This is the standard efficient markets hypothesis. Most of the evidence of autocorrelation in returns in stock markets has related to discrete break points in markets, that is, market openings and closings, weekend effects, and end-tax-year effects (see, e.g., Dimson 1988; McInish and Wood 1991; Wood, McInish, andOrd 1985; Griffiths and White 1993). The 23. We also ran a similar exercise, using hourly data, for the Swiss franc/dollar series, but, with only fifty-one deals in our data period, this was too affected by small sample problems to provide a useful cross-check. Data on this are available from the authors.
136
Charles Goodhart, Takatoshi Ito, and Richard Payne
foreign exchange market exhibits fewer discrete break points; in any case, our sample is far too small, covering no such break points, to hope to test for any such anomalies. We exhibit the characteristics of the transactions data separately for transactions at the bid and the ask and also for the two series taken together (to see what the effect on the characteristics would be if, counterfactually, we could not distinguish between deals at the bid and the ask; see table 4.12). During our short snapshot, the deutsche mark/dollar traded upward (i.e., the dollar Table 4.12
Transactions in Deutsche MarkIDollar
Number Average size Levels mean SD Skew Kurtosis First Differences Mean SD Skew Kurtosis
Bid
Ask
Bid + Ask
186 $2.51 mn 1.64946 .0062 .5541 -1.4346
251 $2.49 mn 1.64978 .0061 .5571 -1.3617
437 $2.5 mn 1.6496 .0061 .5518 -1.3910
.000042 .00054 5.096 38.556
.000034 .00030 -0.575 10.164
.000019 .000269 1.273 15.326
Autocorrelation coefficient:
-1 -2 -3 -4 -5 GARCH: Ao Al
-.084 (-1.11) -.069 (- .90) .155 (2.07) -.009 (- .12) .003 (.04)
-.086 (-1.32) .138 (2.13) -.085 (-1.30) .050 (.77) .042 (.66)
-.000 (- .63) -.004
.000 (1.70) -.159 (-1.89) .000 (4.11) .572 (4.21) .246 (2.47)
.000 (.52) -.365 (-17.49) .000 (0.32) .553 (19.49) .478 (56.46)
-91.22
-301.3
(- .55) Bo BI B2 Dickey-Fuller test with 5 lags
.000 (2.89) .415 (20.26) .491 (44.87) -643.34
Note: t-statistics are given in parentheses.
-.1406 (-2.90) .0949 (1.96) .0185 (.38) .054 (1.12) .026 (.54)
137
A Study of the Reuters D2000-2 Dealing System
appreciated). So the mean change on all three series was positive, but less so for the composite series because of bouncing between deals done at the bid and the ask. Because of that same bounce, the absolute size of the negative autocorrelation on the first lag becomes larger (almost doubles) and becomes significant. Thus, we claim to be able to document here the statistical effect of the bounce. It would be possible to use these data to check the accuracy of the Roll (1984) model whereby the size of the bid-ask spread is estimated using only transaction prices. We leave that for later work, although we doubt whether that model would perform well, for example, because the direction of deals is autocorrelated and information asymmetry (volatility) is time varying. The positive coefficients at higher lags on the other two series may be owing to the large jumps in the dollar during our short data period. Bollerslev and Domowitz (1993, 1430-32; see also Bollerslev and Domowitz 1991) generate artificial transactions series from automated trade execution algorithms that exhibit positive first-order serial correlation; we find no sign of that outcome in our data set of actual transactions prices. Hasbrouck and Ho (1987) find that, for the NYSE, "the pattern consists of a large negative auto-correlation at the first lag, followed by positive autocorrelations of decreasing magnitude that are statistically significant ... through the fifth lag. The negative first order auto-correlation in transactions data is consistent with the findings of other studies. The positive autocorrelations, however, are (in transactions data) new" (1039). While the size and significance of our coefficients are considerably less, the general pattern in our data is exactly the same. With the significant negative first-order autocorrelation being caused by the bounce and none of the later positive autocorrelations being either large or significant, our results are, not surprisingly, consistent with efficiency. The Dickey-Fuller test indicates stationarity. This does not disturb us. The random walk characteristic of asset prices results from their subjection to a sequence of "news" shocks. At anyone point of time, the market price of an asset should have an equilibrium value, dependent on assessments of past "news" shocks. If the time period is short enough, here only seven hours, the amount of additional "news" is limited, so, over very short time periods, one might expect to observe stationarity. What we do feel remains to be clarified and modeled is the nature of the interaction between a quotes series that shows clear evidence of a negative moving average component and a transactions series that exhibits no such significant autocorrelations. This is the subject of our ongoing research. According to the simplified models wherein a single dealer undertakes one transaction of a standardized size per period, the dealer should adjust prices until the expectation of a transaction at the bid next period is equal to one at the ask. So the sequence of deals between bids and asks should be random (see, e.g., Admati and Pfleiderer 1988, 1989; and Hasbrouck and Ho 1987). If
138
Charles Goodhart, Takatoshi Ito, and Richard Payne
inventory effects are present, the sequence might be expected to show some negative autocorrelation. With many dealers posting limit orders and multiple orders possible in any finite period, we would, however, not expect that. Instead, we would expect runs of deals of each kind. We test that hypothesis, both by a histogram showing the lengths of sequences of deals of both kinds and by a formal runs test. The histogram, figure 4.13, shows that there are a number of runs of deals, at both the bid and the ask, that are much longer than one might normally expect to see. These are shown in table 4.13, together with their individual expected probability of occurrence. The probability of finding all such runs together is infinitesimal. Thus, rather like the kurtotic characteristic of the price change series, the run series for deals appears to have a fat tail. As noted earlier, there are indications that runs of similarly signed deals occur when the price series is trending in one direction, for example, dollar buying at the ask where the dollar is appreciating. We show the associated change in the relevant quoted price during each run over the same period in table 4.13. The formal runs test that we use is the Geary test. This concentrates attention on whether the number of runs observed in the sample is large or small relative to the number that one would expect to occur in a strictly random sample. According to this test, we are led to reject strongly the null that successive observations are independent since the test statistic is -7.11 compared to the standard normal critical value of -2.58 under the null. Some earlier empirical work has also found evidence that deals tend to run in sequences (bid deals followed by bid deals and ask deals followed by ask deals), for example, Hasbrouck and Ho (1987) and Lease, Masulis, and Page (1991) for the NYSE. Some of the reasons for this are straightforward, for example, a trader with a large order working up the limit order book. We would, however, conjecture, but have yet to do the work required to demon-
60
50
r;
40
c
!
f
30
20 10
o Run Length
Fig. 4.13
Deutsche mark/dollar deal runs: bid and ask combination
139
A Study of the Reuters D2000-2 Dealing System
Table 4.13 Run Length 21 17
15 14 12 11 9 8
8 7 7 7
Deal Runs, Price Changes, and Sample Probability of Occurrence Side of Market Ask Ask Ask Ask Bid Bid Ask Ask Ask Bid Bid Ask
Percentage Price Change .272
.516 .103 .122 ~.030 ~.097
.121 .055 .018 ~.090 ~.055
.018
Sample Probability
8.8 X 10- 6 8 X 10- 5 2.4 X 10- 4 4.2 X 10- 4 3.6 X 10- 5 8.4 X 10- 5 6.8 X 10- 3 .0018 .0018 .0025 .0025 .0205
Note: Percentage price change represents the percentage difference in the quotes on the stated side of the market at which the first and last transaction in each run took place. Sample probability is simply the probability, given the sample frequency of each type of deal, of observing n successive transactions on one side of the market, assuming that they are independent events.
strate, that the extent of autocorrelation revealed here is considerably beyond the explicable on the basis of such simple microstructural factors. The only theoretical explanation yet given for such positive autocorrelation is by Admati and Pfleiderer (1989). They suggest that market dealers may shade the costs of dealing, on one side of the market, to encourage liquidity traders to bunch together on that side, isolating and identifying informed traders on the other: "The intuition behind our results suggests that there will be periods in which prices rise at a slow rate when shares are purchased but fall at a more rapid rate when shares are sold. These periods will be periods of concentrated buying-periods in which it is expected that discretionary buyers will be trading" (Admati and Pfleiderer 1989,209). Our results are very different. In our data, buys concentrate together when prices are rising rapidly and spreads rising, but not enough to choke off the stream of purchases. At such moments, seller-initiated trades dry up altogether. Further research to check whether our results are typical of the foreign exchange market and, if so, what the reasons for this might be would be desirable. 4.4.2
The Interrelations between the Data Series
Given the existence of such long runs of deals at the bid and ask, one variable that may help predict the occurrence of a deal at the bid (ask) is whether there has been a prior deal at the bid (ask). Hence, we now turn to regression analysis to explore the interrelations between our series, separately for both D2000-2 and FXFX. For this purpose, we used our constructed five-second data set, where for D2000-2 a nonentry at either the bid or the ask is replaced by the prior entry, if no deal had occurred, or the subsequent entry following a
140
Charles Goodhart, Takatoshi Ito, and Richard Payne
deal. There was never more than one deal in any five-second period, but, of course, over longer periods (e.g., one minute) there were often several deals. For D2000-2 we had the data series shown in table 4.14 for both the bid and the ask; bid series are given the notation B and ask series A. There were thus seventeen basic series for D2000-2, eight bid, eight ask, and the spread. Initially, we used our five-second database, with lags covering the previous two thirty-second intervals and the two minutes before then, for example, BD l _ 6 , BD7 _ 12 , BD 13 _ 24 , BD25 _ 36 , Qoted as BD 6, BD 12, BD24 , BD 36 • In some cases, for example, for spreads and quote revisions, we also used shorter-unit (five-second) lags, noted as Lag 1, Lag 2, Lag 3, etc. For FXFX, we did not have the first three series, (BD, BDQ, BDE) or QB, so there were four basic series in this case, with similar notation (DB, BF, ADB, and BV), for bid quotes, four for asks, and the spread. This meant that we had over eighty-five basic series (including lags) for D2000-2 and a data set of five thousand observations. 24 Our basic approach was to regress each variable of interest on lagged values of all the variables (including the lagged dependent) separately and then include significant values from these first-stage equations in a larger equation to search for the best-fitting equation. There is a general problem in such exercises of how to scale the data. The two main alternatives are to use standard clock time or transactions (tick) time, whereby each activity observation is ordered consecutively, irrespective of the varying time gap between them. With very high-frequency series, for example, five-second intervals as here, a problem with the use of clock time is that most observations of price changes, deals, etc., are zero. Hence, the distribution of these variables is nonnormal, with a spike at zero. On the other hand, there are certain questions relating to the temporal relations between series, especially in multivariate analysis, that can be answered only using a clock-time scale. Several analysts have wrestled with this prob,lem, notably McInish and Wood (1990, esp. sec. 4.4) with respect to the NYSE and the various studies undertaken by analysts at Olsen and Associates (e.g., Muller et al. 1990; and Dacorogna et al. 1993) of the foreign exchange market. Most empirical work in both stock and foreign exchange markets, has been performed on an activity scale, utilizing tick-by-tick data. The studies (e.g., on price scaling laws), notably those carried out by Olsen and Associates, do suggest that this is probably preferable, where feasible, for the question under consideration. In our case, however, we are interested in multivariate intertemporal relations, so we have primarily used a clock-time scale but have, in certain cases, checked the result from these exercises against similar exercises on an activity scale. 24. Our computer could not handle a general to specific exercise with parameters of this size, although there was relatively little multicollinearity or autocorrelation (apart from the spread, S, which was strongly positively autocorrelated in D2000-2). We ran a simple cross-correlation matrix, which is too large to reproduce but is available from the authors.
141
A Study of the Reuters D2000-2 Dealing System
Table 4.14
Codes for Variables Bid Number of deals in period Quantity traded in deal Dummy if deal exhausted quantity Change in quote Quantity quoted Frequency of quote revision over period Absolute value of change in quote Standard deviation of changes in quotes Spread
Ask
AD
BD BDQ BDE DB QB BF
ADQ ADE DA QA AF ADA
ADB BV
AV S
T~e following exercises are quite detailed. The relevant tables are tables 4.15-4.32 below. Readers may prefer to skip first to the figures showing, qualitatively, the main directions of relations (figs. 4. 14a-c below) and also to the summary of main findings in section 4.4.5 before deciding how much detail in the next few pages they want to absorb. There were only some 186 bid deals in the deutsche mark/dollar during the more than five thousand five-second intervals. So, to examine the likelihood of a bid deal occurring, we used probit analysis. Our "best" equations for the probability of bid and ask deals occurring are shown in table 4.15. The main finding from this, which was foreshadowed in the results in table 4.11 above, is that the most important set of variables to determine bid (ask) deals is the frequency of bid (ask) quote revisions in the previous few minutes. This frequency, we believe, is probably a proxy for the extent of prior information. When lagged values of BF (AF), the frequency variable, are entered, lags of the dependent variable BD (AD) lose most of the significance they had when entered alone. Besides this frequency variable, in both cases, if there was a deal of the opposite sign (e.g., AD6 in the BD equation) in the previous thirty seconds, there is less likelihood of seeing a deal now. Bid deals in the deutsche mark/dollar are considerably more likely to occur where current spreads are low (i.e., prices are good) and when prices have recently been improving (DB6 is positive). This suggests that traders are doing their job effectively (i.e., hitting comparatively good prices). A comparison of average spreads when there is no deal and when there is a deal for the deutsche mark/dollar and the yen/ dollar is shown in table 4.16. 25 The AD (ask deal) results are more problematic, with some nonintuitive variables entering significantly, that is, a positive lagged spread (thirty seconds previous), positive changes in bid quotes, and a deal quantity variable, ADQ. We surmised that these results might have been due to many of the ask deals
25. Note that the split of the period into subdivisions differs slightly between table 4.10 and table 4.16.
142
Charles Goodhart, Takatoshi Ito, and Richard Payne Probability of Observing Deals
Table 4.15
A. Bid Probit estimates: Number of observations = 4,980 X2(7) = 96.1 Prob. > X2 = .000 Pseudo R2 = .061 Log likelihood
= -732.90769
Bid Deal bf6 bf12 bd36 db6 s ad6 cons
-
Coefficient
SE
.1663953 .1275713 .1151307 154.4183 -431.2308 -.2524728 -1.893544
.0329075 .0328465 .0427033 81.54656 91.80828 .1295843 .069593
p> It I 5.056 3.884 2.696 1.894 -4.697 -1.948 -27.209
.000 .000 .007 .058 .000 .051 .000
B.Ask Probit estimates: Number of observations = 4,980 X2 (l0) = 96.2 Prob. > X2 = .000 Pseudo R2 = .048 Log likelihood = -952.29651 Ask Deal
Coefficient
af6 afl2 adq6 Lag6 s bd6 bd24 db6 db12 db24 db36 - cons
.0737124 .0777921 .0508467 88.19644 -.212043 .1275771 107.5494 197.6548 154.0204 116.271 -1.940738
Note: bd
=:;
P> I tl
SE .0311327 .030167 .0205937 39.06158 .0811112 .0381741 68.76905 66.7621r 58.16457 54.50788 .0564778
2.368 2.579 2.469 2.258 -2.614 3.342 1.564 2.961 2.648 2.133 -34.363
.018 .010 .014 .024 .009 .000 .118 .003 .008 .033 .000
bid-side deals; ad = ask-side deals.
occurring in the latter part of the period, when spreads and volatility were high and both bid and ask quotes prices rising markedly. In order to test this, we divided our sample into two parts, the flat first half (observations 1-3560) and the upward-trended second half (observations 3561-5000), and redid the probit analyses for both the bids and the asks. The results for bid deals remained much the same. For ask deals, the spread becomes negative (as expected) in
143
A Study of the Reuters D2000-2 Dealing System
Table 4.16
Spreads: A Comparison of Spreads at Ordinary Times with Those at Transaction Times A. Deutsche MarkIDollar Bid-Ask Spread Bid-Ask, All Samples
Hour 0 1 2 3 4 5 6 7
Bid-Ask, Transaction Time Only
Mean Unit = DM/$
Median Unit = DM/$
Number of Observations
Mean Unit = DM/$
Median Unit = DM/$
Number of Observations
.0004214 .0004992 .0003791 .0005388 .0005110 .0011005 .0007651 .0007530
.00030 .00040 .00030 .00040 .00040 .00070 .00070 .00050
607 708 671 577 647 656 602 49
.0004125 .0003587 .0002391 .0003700 .0003113 .0010630 .0004777 .0004000
.00020 .00030 .00020 .00030 .00030 .00060 .00040 .00040
72 46 46 30 44 92 72 6
B. YenIDollar Bid-Ask Spread Bid-Ask, Transaction Time Only
Bid-Ask, All Samples
Hour 0 1 2 3 4 5 6 7
Mean Unit = Yen/$
Median Unit = Yen/$
Number of Observations
Mean Unit = Yen/$
Median Unit = Yen/$
Number of Observations
.11213 .10996 .14967 .18212 .15825 .14814 .10598 .08000
.15000 .13000 .20000 .20000 .19000 .15000 .10000 .08000
136 720 720 720 554 199 112 14
.01000 .08000 .08833 .14000 .04500
.02000 .09000 .09500 .14000 .05000
.08333
.03000
2 3 6 2 4 0 6 0
Note: (1) Each hour has a maximum of720 observations (five-second intervals). If an ask or bid is missing, then that bracket is not counted in the left-hand-side panels of "all" observations. (2) Transaction time bid-ask spread is the bid-ask spread of the five-second bracket, preceding the five-second bracket where a transaction occurs. There are instances where transactions occur even without one of the bid or ask being shown on the screen (just before the transaction is recorded). These are treated as missing observations in the right-hand-side panels.
the first half and insignificant in the second part; and the change in the ask price (DA36) also enters negatively, as expected, in the first half of the period. Apart from the insignificant spread, the results for the first part ask are similar to those of the contemporaneous bid. The table giving these two half-period results is available on request from the authors. Overall, however, the fit was rather poor. Perhaps it was expecting too much of the data to be able to predict the probability of a deal within a period as short as five seconds. So we lowered the frequency of analyzed periodicity to a minute. Within a minute, however, there were often several deals. So we used ordered probit analysis to estimate the interrelations. Somewhat to our surprise, the change of periodicities to the lower frequency of one-minute inter-
144
Charles Goodhart, Takatoshi Ito, and Richard Payne
vals made relatively little difference to the major apparent patterns of relations (see table 4.17). Given the probability of a deal, the next question is what will be the volume, the size of the deal. In 145 of 186 deals at the bid and 179 of 251 deals at the ask, the deal, however, exhausted the outstanding quantity offered. So the size of the deal was usually limited by the amount on offer. That means that it is more sensible to try to model the amounts offered by the dealers (BQ and AQ) than the amounts sought by the hitters (i.e., the supply function is better identified than the demand function). Similarly, of course, the price of the deal has to be at the price posted, either the bid or the ask, in the firm quotes. So we tum next to an analysis of the determinants of the changes in such prices, DA and DB. As noted earlier, when a quote is hit and exhausted, the price must change to the next limit order, if such exists. There is also known to be negative autocorrelation in the quote series. Our first basic exercise was, therefore, to regress DA and DB against their first six, t-1 to t-6, own lags and the dummy exhaust variable, BDE and ADE, taking the value 1 when the quote was exhausted by a deal. The results
Table 4.17
Ordered Probit Analysis on Data at One-Minute Intervals A. Bid-Side Deals Number of observations = 403 X2 (4) = 38.1 Prob. > X2 = .000 Pseudo R2 = .056 Log likelihood = - 320.24
Bid-Side Deals
Coefficient
SE
bf6 bf24 db24 Lagls
.0357 .0259 64.45 -180.0
.0133 .0079 31.34 89.27
P> t 2.67 3.28 2.05 -2.02
.008 .001 .041 .044
B. Ask-Side Deals Number of observations X2 (4) = 34.4 Prob. > X2 = .000 Pseudo R2 = .040 Log likelihood Ask-Side Deals af6 ada6 da12
Note: db
= 391
= -408.15
Coefficient
SE
.0351 198.20 -69.07
.0117 59.49 49.96
= change in bid quote; da = change in ask quote.
P> t 2.993 3.331 -1.383
.003 .000 .168
145
A Study of the Reuters D2000-2 Dealing System
are as shown in table 4.18. The value of the dummy exhaust variables (BDE and ADE) was in each case about ± .000375, showing that this is the average price revision (down following a bid exhaust, up after an ask exhaust), or alternatively the gap between limit orders, following a deal. The negative values for the lagged own values are consonant with the now-well-established highfrequency negative autocorrelation. The lower value of the coefficient on the first lag than in table 4.7 above is due to the fact that the series here are on clock time, five-second intervals, and not taken, as in table 4.7, by consecutive quotes. Consequently, most of the observations on price changes show zero. When we reran the exercise on exTable 4.18
Basic Determinants of Quote Revision
A. Bid Quote Revision Number of observations = 87.9 = .000 R2 = .110 Adjusted R2 = .109 Root MSE = .0002
= 4,976
F(7,4969) Prob. > F
Change in Bid Quote (db)
Coefficient
Lag1 db Lag2 db Lag3 db Lag4 db Lag5 db Lag6 db Lag1 bde - cons
-.11656
.0136231
- .1176993
.0137196
-.1320021 -.0546471 -.0210431 -.0776679 -.0003729 .0000157
.0138006 .0137991 .0137147 .0136245 .0000189 3.21e-06
P> I tl
SE -8.556 -8.579 -9.565 -3.960 -1.534 -5.701 -19.745 4.881
.000 .000 .000 .000 .125 .000 .000 .000
B. Ask Quote Revision Number of observations = 114.1 = .000 R2 = .138 Adjusted R2 = .137 Root MSE = .0002
= 4,976
F(7,4969) Prob. > F
Change in Ask Quote (da)
Coefficient
SE
Lag1 da Lag2 da Lag3 da Lag4 da Lag5 da Lag6 da Lag1 ade - cons
-.111342 -.08533 -.0669398 -.0322129 -.1609648 -.0400806 .0003769 -8.7ge-06
.0134768 .0133953 .0134363 .0134368 .0133957 .0134622 .0000161 3.06e-06
P> 1 tl -8.262 -6.370 -4.982 -2.397 -12.016 -2.977 23.339 -2.871
.000 .000 .000 .017 .000 .003 .000 .004
146
Charles Goodhart, Takatoshi Ito, and Richard Payne
actly the same basis but omitting those observations when prices changes were zero, we got the results shown in table 4.19. The absolute size of the coefficients of the lagged dependent variables increases by a factor of about five times (as the 80 percent of zero observations in the complete, clock-time, sample are removed), but the standard errors increase by as much, or slightly more, so the t- values actually decline, just, on balance. Since there virtually has to be a change in price after a deal exhausts the previous quote entry, coefficients of the deal exhaust dummies, BDE and ADE, rise only slightly, and, with a commensurately higher standard error, their t-values fall from around 20 to Table 4.19
Basic Determinants of Quote Revision: Zero Changes Omitted: Tick by Tick A. Bid Quote Revision Number of observations = 727 F(7,720) = 34.2
Prob. > F = .000 R2 = .249 Adjusted R2 = .242 Root MSE = .0005 Change in Bid Quote (db)
Coefficient
SE
Lag1 db Lag2 db Lag3 db Lag4 db Lag5 db Lag6 db Lag1 bde - cons
- .4204473 - .4810583 - .5284103 - .2811608 -.0902551 -.4268289 .0004502 .0000989
.0687554 .0717263 .0692662 .0843628 .0788967 .0869281 .0000513 .0000223
p> 1 tl -6.115 -6.707 -7.629 -3.333 -1.144 -4.910 -8.783 4.432
.000 .000 .000 .000 .253 .000 .000 .000
B. Ask Quote Revision Number of observations
=
747
F(7,740) = 43.1 Prob. > F = .000
R2 = .289 Adjusted R2 = .283 Root MSE = .0004 Change in Ask Quote (da)
Coefficient
SE
Lag1 da Lag2 da Lag3 da Lag4 da Lag5 da Lag6 da Lag1 ade - cons
-.7198426 - .3373846 - .4080851 -.2991646 -.53916 -.1875199 .0004132 -.0000571
.0813796 .0585086 .0765337 .075899 .0578206 .0644919 .0000445 .0000209
p> 1 tl -8.845 -5.766 -5.332 -3.942 -9.325 -2.908 9.295 -2.729
.000 .000 .000 .000 .000 .004 .000 .007
147
A Study of the Reuters D2000-2 Dealing System
about 9. The resultant series without the zeroes (i.e., in transaction time) is much more variable, so, although the fit of the series is much improved (the adjusted R 2 doubles from around .12 to about .25), the root MSE also doubles. We then explored to find other variables that might contribute significantly to the determination of quote revision, although the own lags out to t- 6 and the exhaust dummy remained the key variables. The main additional variables that entered in table 4.20 were the spread with a one-period lag, negatively for the ask and positively for the bid (i.e., where the spread was unusually large, someone would come forward with a more competitive quote); longer own lags (although this was more apparent in equations run without the spread, as shown in table 4.21); and some volatility variables. 26 When the spread variable is not included, changes in the ask price have a strong positive effect on changes in the bid price, whereas changes in the bid price had a weaker effect on changes in the ask prices (see the coefficients italicized in table 4.21). But the sum of the coefficients is well below unity. What this means is that, in this market, a change in the best bid (ask) has only a slight effect on the contemporaneous ask (bid). Most of the immediate effect becomes translated into a changed spread, which is highly positively autocorrelated. The spread returns toward normal only slowly. So, in this market, with best bids and asks being entered by different banks, the hypothesis that these two quotes will be revised closely and quickly in step with each other is convincingly refuted; instead, bids and asks vary somewhat independently, rather like two variables that are cointegrated in the longer run, with the spread acting as the error correction mechanism between them. We have no convincing explanation for the asymmetry whereby the change in the ask quote price had a stronger effect on the bid quote price than vice versa. We initially thought that this might be due to the surge in the value of the dollar in the second half of the period, affecting first ask deals and quotes and thereafter bid quotes, but, when we divided the period into two and reran, this hypothesis was refuted since, although the effect of DA on DB was slightly weaker than in the full sample, it was clearly stronger in the first, untrended part of the period than in the second part, when the dollar strengthened. We also looked for any signs that either the event or the size of deals influenced quotes, apart from the exhaust dummies, which, as already noted, were highly significant. We found generally rather weak effects, as in table 4.20 below, of these variables on quote revisions, but where significant usually of the expected sign. Thus, in some of the equations for bid quote revisions, DB, the event (BD) or the quantity (BDQ) of a deal in prior periods would enter
26. Such volatility variables were usually AV12, or sometimes BV60, in the ask price change, DA, equation and ADB in the bid price change, DB, equation. Rather counterintuitively, this latter variable was positive in the DB equation, and, when it entered, AV12 was negative in the DA equation, implying that higher volatility led to finer, more competitive prices being posted, but the significance level of this is not high.
148
Charles Goodhart, Takatoshi Ito, and Richard Payne
Table 4.20
The Determinants of Quote Revision A. Bid Quote Revision Number of observations F(13,4963) = 64.4 Prob. > F = .000 R2 = .144 AdjustedR2 = .142 Root MSE = .0002
= 4,976
Change in Bid Quote (db)
Coefficient
SE
Lag1 db Lag2 db Lag3 db Lag4 db Lag5 db Lag6 db Lagl bde db12 bdq6 adb6 adb24 adq24 Lagl s - cons
- .1199148 -.127044 -.1449427 -.0792908 -.0478669 -.1040509 -.0003474 -.0503687 -.0000108 .0375697 .0103388 5.36e-06 .0375859 -.0000196
.0139002 .0139954 .0140143 .0141885 .0140588 .0140001 .0000196 .0076551 3.32e-06 .0086841 .00735 1.70e-06 .0048246 4.76e-06
p> 1 tl -8.627 -9.078 -10.342 -5.588 -3.405 -7.432 -17.734 -6.580 -3.240 4.326 1.407 3.148 7.790 -4.116
.000 .000 .000 .000 .253 .000 .000 .000 .001 .000 .160 .002 .000 .000
B. Ask Quote Revision Number of observations F(l1,4964) = 76.5 Prob. > F = .000 R2 = .145 Adjusted R2 = .143 Root MSE = .0002
= 4,975
Change in Ask Quote (da)
Coefficient
SE
Lag1 da Lag2 da Lag3 da Lag4 da Lag5 da Lag6 da Lagl ade ad6 16db bv60 Lag1 s - cons
- .0946171 -.0699815 -.053245 -.018927 -.1497416 -.0278212 .0003646 .00001 -.0307674 .0000389 -.0278851 -2.63e-06
.0140177 .0138121 .0138313 .0138015 .0136565 .0135895 .0000176 5.83e-06 .0127157 .0000135 .0048746 4.68e-06
P> I tl -6.750 -5.067 -3.850 -1.371 -10.965 -2.047 20.663 1.723 -2.420 2.881 -5.720 -.563
.000 .000 .000 .170 .000 .041 .000 .085 .016 .004 .000 .574
149
A Study of the Reuters D2000-2 Dealing System
Table 4.21
Quote Revisions A. Bid Reaction to Changes in Ask Quotes Number of observations F(l6,4959) = 48.1 Prob. > F = .000 R2 = .134 AdjustedR2 = .131 Root MSE = .0002
Change in Bid Quote (db)
Coefficient
Lagl db Lag2 db Lag3 db Lag4 db Lag5 db Lag6 db db12 db24 db36 Lagl bde da6 da12 da24 da36 bv12 av12 - cons
-.1428408 -.1473166 -.1631221 -.0939637 -.0619189 -.1211814 -.067737 -.0103636 -.0191648 -.0003716 .0162051 .029355 .0231173 .0232844 .0001046 6.26e-06 -1.06e-06
=
4,975
P> I tl
SE .0136389 .0137856 .0139074 .0140831 .0140772 .0141262 .0084264 .0070925 .0065055 .0000187 .0072949 .0076464 .006233 .0060726 .0000154 .0000158 4.26e-06
-10.473 -10.686 -11.729 -6.672 -4.399 -8.578 -8.039 -1.461 -2.946 -19.894 2.221 3.839 3.709 3.834 6.778 .397 -.248
.000 .000 .000 .000 .000 .000 .000 .144 .003 .000 .026 .000 .000 .000 .000 .692 .804
B. Ask Reaction to Changes in Bid Quotes Number ofobservations F(l6,4959) = 51.6 Prob. > F = .000 R2 = .142 Adjusted R2 = .140 Root MSE = .0002 Change in Ask Quote (da)
Coefficient
Lagl da Lag2 da Lag3 da Lag4 da Lag5 da Lag6 da da12 da24 da36 Lagl ade db6 db12 (continued)
- .114878 -.088986 -.0702402 -.0350031 -.1631024 -.0421156 -.0055925 -.0156702 -.0039288 .0003776 -.0027777 .0160504
=
4,975
SE .013517 .0134761 .0135679 .0136022 .0136826 .0138858 .0074358 .0059864 .0058462 .0000163 .0075712 .0079987
P> 1 tl -8.499 -6.603 -5.177 -2.573 -11.920 -3.033 -.752 -2.618 -.672 23.221 -.367 2.007
.000 .000 .000 .010 .000 .002 .452 .009 .502 .000 .714 .045
150
Charles Goodhart, Takatoshi Ito, and Richard Payne
Table 4.21
(continued)
Change in Ask Quote (da)
Coefficient
db24 db36 av12 bv12 - cons
.0206124 .0064569 -.0000359 .0000147 -6.30e-06
P> I tl
SE .0068112 .0062616 .0000152 .0000148 4.08e-06
3.026 1.031 -2.361 .997 -1.542
.002 .303 .018 .319 .123
with a significant negative sign (and, even more occasionally, AD or ADQ lagged would enter with a positive sign), suggesting that stronger deal activity at the bid (ask) caused bid quotes to be lowered (raised). The same feature also occurs weakly for DA, with AD entering positively, as shown in panel b of table 4.20. Again, we examined how the results would change if we ran the regressions omitting all zero price change entries (80 percent of the sample). The results are shown in table 4.22. Our process of trying to eliminate insignificant variables resulted in almost identical "best" equations, with and without zero price changes, but the relative importance of the coefficients as measured by their t-values changed. 27 The fit, as before, improves sharply once zero price changes are omitted, with the adjusted R2 improving threefold in the bid price equation (to 0.43) and more than doubling (to 0.33) in the ask price equation. But, with a more variable series, the root MSE also again doubles. We next compared our results for the determination of quote revision over D2000-2 with a similar exercise for FXFX (see table 4.23). The results for DFXB and DFXA showed similar features for the lagged dependent variable with strong negative autocorrelation (a first-order negative moving average pattern) and a significant role for the spread (positive in the bid equation, negative in the ask). Again as in the D2000-2 equations, volatility variables appear to enter, but in rather a complicated way. Thus, the absolute change in the ask price enters the determination of the change in both the ask and the bid price at two separate lags with reversed signs. Tests over a longer run of data are needed to resolve whether, and how, prior volatility affects price quote revi27. The absolute size of the coefficients on the lagged dependent variables increased by a factor of over three times for bid prices but nearer eight times for ask prices. With their standard errors rising by a factor of over five times in both cases, the t-values of the lagged dependent variables fell for bid price changes but rose for ask price changes (relative to those in table 4.20). As before, the t-values of the deal exhaust variables fell from nearly 20 to about 5. By contrast, the coefficient on the lagged spread variable rose sharply in the bid price equation, where the size of the coefficient rose by a factor of ten and the t-value also increased. (Note that we did test that the spread with six lags entered more strongly than the spread lagged once in the ask price change equation.) Otherwise, the residual variables that entered significantly changed around slightly; a variety of volatility variables still entered weakly without any clear, or intuitive, direction of effect, and, again, the effects of previous large quantities of ask deals (AD6 and ADQ24) tended to raise both bid and ask quotes.
151
A Study of the Reuters D2000-2 Dealing System
Table 4.22
The Determinants of Quote Revision: Zero Changes Omitted: Tick by Tick A. Bid Quote Revision Number of observations = 727 F(l4,713) = 40.6 Prob. > F = .000 R2 = .444 Adjusted R2 = .433
Root MSE = .0004 Change in Bid Quote (db)
Coefficient
SE
Lag1 db Lag2 db Lag3 db Lag4 db Lag5 db Lag6 db Lag1 bde db12 bd6 Lag1 s adb12 adq24 ada6 ada24 - cons
-.2573243 - .4178523 -.4402707 - .253194 -.0483614 -.3135134 -.0002782 -.1727642 -.0000829 .3735075 -.1342426 .0000298 -.0906233 .1076248 -.000137
.0662203 .0650541 .0644304 .0779863 .0720124 .0784248 .000057 .0426325 .0000325 .0334393 .0479387 9.6ge-06 .0417378 .0440883 .0000306
P> I tl -3.886 -6.423 -6.833 -3.247 -.672 -3.998 -4.881 -4.052 -2.554 11.170 -2.800 3.075 -2.171 2.441 -4.471
.000 .000 .000 .001 .502 .000 .000 .000 .011 .000 .005 .002 .030 .015 .000
B. Ask Quote Revision Number of observations F(lI,736) = 33.7 Prob. > F = .000 R2 = .335 Adjusted R2 = .325 Root MSE = .0004
= 747
Change in Ask Quote (da)
Coefficient
SE
Lag1 da Lag2 da Lag3 da Lag4 da Lag5 da Lag6 da Lag1 ade adq24 ad6 Lag6 s adb6 - cons
-.7993277 - .4331602 -.4928082 -.422294 -.5945421 -.1588786 .0003121 .0000411 .0000926 -.125264 .1099008 -.0000683
.0805063 .0587267 .076277 .0766422 .0571305 .0634126 .0000558 8.78e-06 .0000305 .0230924 .0436925 .0000277
P> I tl -9.929 -7.376 -6.461 -5.510 -10.407 -2.505 5.595 4.684 3.040 -5.424 2.515 -2.466
.000 .000 .000 .000 .000 .012 .000 .000 .002 .000 .012 .014
152
Charles Goodhart, Takatoshi Ito, and Richard Payne
Table 4.23
The Determination of Quote Changes over FXFX A. Bid Prices Number of observations = 4,983 F(10,4973) = 96.1 Prob. > F = .000 R2 = .162 Adjusted R2 = .160 Root MSE = .0002
Change in FX Bid Quote (dfxb)
Coefficient
SE
Lag1 dfxb Lag2 dfxb Lag3 dfxb Lag4 dfxb Lag5 dfxb Lag6 dfxb fxb12 Lag6 s adfxa6 adfxa24 _cons
-.3652674 -.298141 -.2499072 - .1231102 -.0775751 -.05442 -.0198112 .2664089 -.0256695 .0275647 .0001859
.0140455 .0148757 .0153796 .0155578 .0154255 .015174 .0096738 .0266012 .011956 .0097646 .000019
P> I tl -26.006 -20.042 -16.249 -7.913 -5.029 -3.586 -2.048 10.015 -2.147 2.823 -9.770
.000 .000 .000 .000 .000 .000 .041 .000 .032 .005 .000
B. Ask Prices Number of observations = 4,983 F(9,4974) = 117.5 Prob. > F = .000
R2 = .175 Adjusted R2 = .173 Root MSE = .0002 Change in FX Ask Quote (dfxa)
Coefficient
SE
Lag1 dfxa Lag2 dfxa Lag3 dfxa Lag4 dfxa Lag5 dfxa Lag6 dfxa adfxa6 adfxa24 Lag6 s - cons
-.3707797 - .3252713 -.2448494 -.1190115 -.0820493 -.0497289 -.0317825 .0523204 -.2650965 .0001846
.0141493 .0149596 .0155451 .0155936 .015018 .0142814 .0130013 .0105903 .0292663 .0000209
P> I tl -26.205 -21.743 -15.751 -7.632 -5.463 -3.482 -2.445 4.940 -9.058 8.831
.000 .000 .000 .000 .000 .000 .015 .000 .000 .000
sion, either over D2000-2 or over FXFX. The other variables tested (i.e., the prior frequency of quote revision, the absolute change in lagged bid prices, etc.) were not significant. In D2000-2, unlike FXFX, changes in the bid (ask) price initially become incorporated into the spread, which is highly positively correlated. Indeed, the
153
A Study of the Reuters D2000-2 Dealing System
first-order autocorrelation with the spread in the previous five-second period has a coefficient of about 0.88 and a t-value in excess of 50, as will be shown below. In order to lessen the power of this relation and show the effects of other variables, we mostly worked with a lagged dependent variable with a thirty-second lag, Lag 6 s. Once again, a deal that exhausts a quote will force a price revision and an increase in the spread as the price shifts to the next limit order, so BDE t _ 1 and ADEt _ 1 were always entered. Thus, the basic equation was 5 = .000220 + 0.620 5_ 6 + .000179 BDE_ 1 + .000313 ADEt _ 1 (0.000015) (0.011) (0.000047) (0.000042) R2 = 0.398.
As earlier noted, an increase in the bid price will reduce the spread, and an increase in the ask price will increase it. These results came through strongly in the equations. The standard finding is that volatility will increase spreads, and this was also strongly supported, as shown by the significant t-values on AV and BV. Our basic equation, using 5t - 6 as the lagged dependent variable, is shown in table 4.23. When 5t - 1 is introduced instead, the fit improves, but the significance of all the other variables· weakens dramatically, and even the sign of the other independent variables often goes wrong since almost all their influence is incorporated into 5t - 1, as shown in panel B of table 4.24. Besides the exhaust dummies, price revision, and volatility variables, we also looked to see whether either the event or size of deals or the frequency of quote revisions affected the spread. The answer is generally no, once the significant variables above are also entered. As can be seen from table 4.23, the number of bid deals in the thirty seconds from t-30 to t-60 (i.e., BD12) enters with a negative significant coefficient. There is some uncertainty in the literature about what relation to expect between the volume (number) of transactions and the spread. On theoretical grounds, Admati and Pfleiderer (1988) and Foster and Viswanathan (1990) expect liquidity trading to cluster together so that low adverse selection trading costs should occur at times of high volume; yet there is evidence in both the NYSE (Foster and Viswanathan 1993) and the foreign exchange market (Glassman 1987) that the intraday pattern is for spreads to be positively correlated with volume. Bessembinder (1994) seeks to resolve this conflict by distinguishing between expected and unexpected volumes, with the expected signs on these being found to be, as hypothesized, negative and positive. We do not, however, feel that our relatively weak finding of a negative coefficient on a volume variable helps resolve this problem; we are inclined to dismiss this finding as possibly occurring by chance; its significance, along with that of many other variables, was cut back sharply when 5t - 1 was entered as the lagged dependant variable. By contrast, there is no uncertainty in the literature that information asym-
154
Charles Goodhart, Takatoshi Ito, and Richard Payne
Table 4.24
Spreads A. With Lagged Dependent 5t - 6 Number of observations F(16,4964) = 953.3 Prob. > F = .000 R2 = .754 Adjusted R2 = .753 Root MSE = .0003
Lag6 s Lag1 bde Lag1 ade bv12 bv60 av12 av60 bd12 db6 db12 db24 db36 da6 da12 da24 da36 - cons
Coefficient
SE
.4768366 .0002891 .0003563 .0002113 .0002813 .0001099 .0003531 -.0000393 -.672806 -.2537867 -.1797322 -.0684526 .7448953 .3830915 .2522268 .1455865 .0001214
.0120753 .0000298 .000027 .0000277 .0000269 .000028 .000024 .0000107 .0127785 .0154748 .0128506 .0107952 .0117689 .0148889 .0118793 .0104181 8.33e-06
= 4,980
P> I tl 39.489 9.691 13.184 7.634 10.450 3.919 14.688 -3.679 -52.652 -16.400 -13.986 -6.341 63.293 25.730 21.232 13.974 14.578
.000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000
B. With Lagged Dependent, 5t - 1 Number of observations F(16,4964) = 1460.7 Prob. > F = .000 R2 = .824 Adjusted R2 = .824 Root MSE = .0003
Lag1 s Lag1 bde Lag1 ade bv12 bv60 av12 av60 bd12 db6 db12 db24 db36 da6
Coefficient
SE
.8807202 .0003409 .0003596 -.0000139 .000057 .0000297 .0000282 -8.43e-06 .042907 .0280221 -.0019636 .0107807 -.0089642
.0136244 .0000252 .0000228 .0000239 .0000233 .0000237 .0000216 9.02e-06 .0150085 .0144781 .0115459 .0092818 .0153301
= 4,980
P> I tl 64.643 13.517 15.750 -.582 2.449 1.252 1.310 -.934 2.859 1.935 -.170 1.161 -.585
.000 .000 .000 .560 .014 .210 .190 .350 .004 .053 .865 .245 .559
155
A Study of the Reuters D2000-2 Dealing System
Table 4.24
(continued) Coefficient
da12 da24 da36 - cons
.0435953 .021478 .009385 .00003
P> I tl
SE .0146797 .0112872 .0093128 7.33e-06
2.970 1.903 1.008 4.099
.003 .057 .314 .000
metries and high volatility will be associated with high spreads. 28 This has been found in two recent articles using FXFX data. Bollerslev and Melvin (1994) and Bessembinder (1994). We have, however, shown earlier (tables 4.10 and 4.11 above) that the form of the (numerical) relation (the coefficients) between volatility and spreads differs depending on whether D2000-2 or FXFX data are used. So next, for comparison, we examined the determination of spreads on FXFX for the same deutsche mark/dollar exchange rate over the same period. The results of this (see table 4.25) show that, besides positive autocorrelation (although much weaker than in D2000-2, the coefficient on the first lag drops from 0.88 to 0.38), the spread is again positively related to volatility (ADFXB24). There is also a weak relation with the frequency of quote entry, but the coefficients are of equal and opposite sign, so the net effect is negligible. Most of the variation in spreads in FXFX is just noise, with an adjusted R2 of 0.15, as compared with over 0.75 for D2000-2. We then looked at the factors affecting the absolute change in prices (a rpeasure of the volatility) of bid (ADB) and ask quotes (ADA) both in D2000-2 and in FXFX. The results of this part of the exercise were not particularly exciting (see tables 4.26 and 4.27 as well as n. 29). 28. Much of the literature on spreads, especially for spreads in the NYSE, seeks to distinguish between the effects of trading costs, inventory costs, and information asymmetry (e.g., Madhavan and Smidt 1991). We cannot attempt a similar exercise as we have no measure of inventories, unlike Lyons (1995). 29. Obviously, the exhaustion of the quote by a deal would cause a jump in prices, so, in the equations to explain the absolute change in prices in D2000-2, ADB and ADA, BDE and ADE were entered into their respective equations. The lagged dependent variable and the absolute change in quote revision on the other side (e.g., ADA in the ADB equation) were quite strongly significant. The prior event of deals (AD and BD), but not their size (BDQ and ADQ), and the frequency of price revision (AF and BF) were also a positive, but somewhat weaker, influence on the absolute value (volatility) of price change. The size of ask quotes (AQ) appeared to affect the absolute value of ask price changes, although the two lags that entered had offsetting effects. Two of our (better) representative equations are given in table 4.26. Again, we undertook the companion exercise of looking at absolute price changes on FXFX. Apart from the lagged dependent variables, the spread entered with a significant positive coefficient. Presumably, this is picking up some (expected) determinants of volatility (not otherwise caught by the lagged dependant variables). The change in ask prices enters the equation explaining the absolute change in ask prices, whereas the frequency of quote entries enters the similar equation for the bid prices. With price movements in the bid and ask being much more closely tied together and similar for FXFX than for D2000-2, here we show only the former equation in table 4.27 since the latter (apart from the substitution of FXBF for DFXA) is almost identical.
156
Charles Goodhart, Takatoshi Ito, and Richard Payne
Table 4.25
Determination of FXFX Spreads Number of observations = 4,927 F(5,4967) = 180.3 Prob. > F = .000
R2 = .153 Adjusted R2 = .152 Root MSE = .0002 Coefficient Lagl s Lag4 s fxaf12 fxaf36 adfxb24 - cons
.377904 .0300787 4.58e-06 -4.94e-06 .0353665 .0004243
P> I tl
SE .0131287 .013117 2.76e-06 1.76e-06 .0089404 .0000179
28.785 2.293 1.660 -2.803 3.956 23.654
.000 .022 .097 .005 .000 .000
As described earlier, the frequency of quote revision (BF and AF) Grangercauses the event of deals. The reverse causal relation also holds, with the number of recent deals influencing the frequency of quote revision. This is consistent with the hypothesis that trading activity itself generates revisions of prior information and hence further trading (e.g., French and Roll 1986). Thus, BD6 is the dominant influence on BF and AD6 on AF. Besides this, there is a weak positive effect from the lagged dependent variable and from the lagged frequency of the other quote (AF in the equation for BF, and vice versa), some positive effect of higher price volatility on the frequency of quote revision, and, finally, a weak and rather uncertain (the lagged variables usually had an offsetting effect) effect from the quote size variables (BQ and AQ). We show two of our better representative equations in table 4.28. 30 Once again, largely for the record, we ran associated regressions to examine the determinants of the frequency of quote entry over FXFX. This showed that, apart from own lagged values, the only variable, from the set of FXFX data available examined here, that influenced the frequency of quote entry over FXFX was a lagged volatility variable. 31 In order to save space, the table is not shown but is available from the authors on request. Finally, in this set of studies of activity on D2000-2 (and FXFX), we explored the determinants of the quantities posted, BQ and AQ. (Recall that we chose not to seek to examine the determinants of the size of deal, BDQ and ADQ, since these most often just exhausted the quantity already on offer.) A noteworthy feature of our results is that the quantities posted, BQ and AQ, did not significantly affect most of the preceding variables (e.g., probability of 30. We have no good explanation for the negative values for AD24 or ADA24 in the equation shown in panel A of table 4.28, and we would again be inclined to regard these as chance findings. 31. This volatility variable was the absolute change in prices over the preceding half minute (ADFXA in the ask equation and ADFXB in the bid equation).
157
A Study of the Reuters D2000-2 Dealing System
Table 4.26
The Determinants of Absolute Price Changes A. In Bid Prices (adb) Number of observations = 4,980 F(7,4973) = 78.3 Prob. > F = .000 R2 = .099 Adjusted R2 = .098 Root MSE = .0002
In Bid Prices (adb) adb6 adb12 abd24 Lag1 bde ada12 ada36 af24 - cons
Coefficient
SE
.0638526 .0190271 .0366817 .0003333 .0178956 .0156567 6.13e-06 3.55e-06
.0082624 .0083937 .0073459 .0000181 .0078319 .0064971 2.02e-06 5.04e-06
P> I tl 7.728 2.267 4.994 18.460 2.285 2.410 3.037 .705
.000 .023 .000 .000 .022 .016 .002 .481
B. In Ask Prices (ada) Number of observations = 4,980 F(l1,4969) = 82.2 Prob. > F = .000 R2 = .154 Adjusted R2 = .152 Root MSE = .0002 In Ask Prices (ada)
Coefficient
SE
ada6 ada12 ada24 ad24 ad36 ade1 af6 adb36 bf36 aq12 aq24 - cons
.0621732 .0244177 .0184217 .0000145 .0000113 .0003508 5.86e-06 .0210127 5.46e-06 8.91e-07 -7.22e-07 -6.84e-06
.0079386 .0076472 .0063198 3.37e-06 3.35e-06 .0000155 3.17e-06 .0069293 2.00e-06 3.7ge-07 2. 14e-07 7.87e-06
P> I tl 7.832 3.193 2.915 4.308 3.392 22.667 1.851 3.032 2.726 2.348 -3.365 -.869
.000 .001 .004 .000 .000 .000 .064 .002 .006 .019 .000 .385
deal, quote revision, spread) and only weakly affected, if at all, volatility and the frequency of quote entry. Anyhow, the main factors affecting the quantities offered, BQ and AQ, as shown in table 4.29, are the respective lagged dependent variables, with strongly significant first-order positive autocorrelation (but in the case of BQ thereafter a somewhat complex dynamic process), and the number of prior deals (BD in the BQ equation, AD in the AQ equation), which
158
Charles Goodhart, Takatoshi Ito, and Richard Payne
Table 4.27
The Determinants of Absolute Price Changes on FXFX (adfxa) Number of observations = 4,972 F(5,4966) = 51.7
Prob. > F = .000 R2 = .049 Adjusted R2 = .048 Root MSE = .0002 Determinants of Absolute Price Changes on FXFX (adfxa) adfxa6 adfxa12 adfxa36 s6 dfxa6 _cons
Coefficient
SE
.1366876 .0338006 .0267606 .0990903 .0375036 .0000414
.0117699 .011855 .0095021 .0258231 .008251 .0000187
P> I tl 11.613 2.851 2.816 3.837 4.545 2.220
.000 .004 .005 .000 .000 .026
reduces quote size. Other activity variables, such as BF, AF, and BDQ, enter weakly and often with offsetting signs, so their net effect is negligible. A volatility variable (BV12) enters the BQ equation positively. The only factors, however, about which we have some confidence are those for the lagged dependent variable and the negative effect of deal activity on quote size. This extended series of results and tables must seem quite complicated, and so in a manner it is. We try to simplify by illustrating, in figures 4. 14a-c, the main interrelations (excluding interactions whereby bid variables affect ask variables, and vice versa), with the direction of causality given by the arrow, the strong relationships displayed in figure 4.14a, the weak relationships in figure 4.14b, and the questionable relationships in figure 4.14c. A key point is that deals mainly affect quote (price) revisions, spreads, and volatility if they have exhausted the amount then on offer, but with a much weaker effect otherwise. This deal exhaustion effect is the main link from the deal occurrence/ frequency of quote revision nexus (one-way) to volatility and to the quote revision/spread nexus. The exercises, whose results were reported in these figures, were mostly, except for tables 4.19 and 4.22, done on a clock-time scale. We were both encouraged and slightly surprised to find that, when we changed the periodicity (table 4.17 compared with table 4.15) or the scale (table 4.18 compared with table 4.19 and table 4.20 compared with table 4.22), the patterns of the basic relations, as measured by the t-values on the key variables, remained quite robust. 4.4.3
Conditional Heteroskedasticity in D2000-2
Most asset price series exhibit ARCH, autoregressive conditional heteroskedasticity. We next turned to examine whether our price series, DB and DA,
159
A Study of the Reuters D2000-2 Dealing System
Table 4.28
The Frequency of Quote Entry on D2000·2 A. Of Bid Prices (bf) Number of observations F( 11 ,4969) == 26.6 Prob. > F == .000 R2 = .055 Adjusted R2 = .053 Root MSE = .3435
Of Bid Prices (bf) bf6 bf12 bf24 bf36 bd6 af24 ad24 ada24 bv12 bq24 bq36 - cons
= 4,980
Coefficient
SE
-.0093551 -.0000455 .0109553 .0078274 .1440359 .0149482 -.0157613 -24.13405 .0900122 .001001 -.0008411 .0607582
.0059631 .0053486 .0034244 .0033847 .0111967 .0037679 .0062909 10.76621 .0254729 .0003442 .0003457 .0143452
P> I tl -1.569 -.008 3.199 2.313 12.864 3.967 -2.505 -2.242 3.534 2.908 -2.433 4.235
.117 .993 .001 .021 .000 .000 .012 .025 .000 .004 .015 .000
B. Of Ask Prices (af) Number of observations F(8,4972) = 32.2 Prob. > F = .000 R2 = .049 Adjusted R2 = .047 Root MSE == .3484
= 4,980
Of Ask Prices (af)
Coefficient
SE
af6 ad6 bf12 bf36 adq24 adq36 av60 aq24 - cons
.0166517 .0963283 .0118653 .0109036 .0049921 .0070069 .0510773 -.0006868 .0705923
.005561 .0092722 .0051196 .0033706 .0026929 .0027153 .0199153 .0003179 .0133874
P> I tl 2.994 10.389 2.318 3.235 1.854 2.581 2.565 -2.160 5.273
.003 .000 .021 .001 .064 .010 .010 .031 .000
also had such characteristics, either in clock (five-second) time or on a tickby-tick (activity) scale. We could also explore whether the addition of transaction data (e.g., BO, BOE) would influence the GARCH coefficients. Having already examined the relation between the GARCH coefficients of the interpolated 02000-2 and FXFX series in section 4.3, we now focus solely on the former to investigate whether, in clock time or using a data set constructed solely using quote and transaction activity, the series exhibit signs of condi-
160
Charles Goodhart, Takatoshi Ito, and Richard Payne
Table 4.29
The Determinants of Quote Size A. Bid Quote Size (bq) Number of observations = 4,980 F(9,4971) = 329.2 Prob. > F = .000 R2 = .373 Adjusted R2 = .372 Root MSE = 1.393
Bid Quote Size (bq)
Coefficient
SE
bq6 bq12 bq24 bd6 bf12 bf24 bv12 af24 af36 - cons
.1505502 -.0257189 .0081696 -.1266831 -.0550551 .0343499 .217296 -.0533222 .0394126 .6453671
.0032501 .0030404 .0014368 .0414378 .0215234 .0134897 .0993405 .0133771 .0135053 .058449
P> I tl 46.321 -8.459 5.686 -3.057 -2.558 2.546 2.187 -3.986 2.918 11.042
.000 .000 .000 .002 .011 .011 .029 .000 .004 .000
B. Ask Quote Size (aq) Number of observations = 4,981 F(11,4970) = 235.6 Prob. > F = .000 R2 = .342 Adjusted R2 = .341 Root MSE = 1.458 Ask Quote Size (aq)
Coefficient
SE
aq6 aq24 ad6 ad24 af6 bf6 bdq6 bdq12 bdq36 bq6 bq12 - cons
.1242625 .0095996 - .1941901 -.0502285 .0641108 -.0892569 .053556 .0597338 -.0569668 .0089876 -.0097294 .7090902
.0029287 .0014215 .0391663 .0234261 .0231159 .0230461 .0222602 .0208646 .0140248 .0034053 .0029193 .0635584
P>ltl 42.430 6.753 -4.958 -2.144 2.773 -3.873 2.406 2.863 -4.062 2.639 -3.333 11.157
.000 .000 .000 .032 .006 .000 .016 .004 .000 .008 .000 .000
tional heteroskedasticity. The basic specification that we used is shown below. Quote revisions are assumed to depend on their own first lag and a dummy indicating a deal that exhausted the quantity on offer at the prevailing price in the previous period. The volatility expression is based on a simple GARCH(l,l), extended subsequently to examine the effect of deals on volatility:
161
A Study of the Reuters D2000-2 Dealing System
Quantity Quoted Quote Revision
Frequency of Revision
Volatility Deal Event Quantity of Deal Fig.4.14a
Strong relationships: main transmission channels
Quantity Quoted Quote Revision
Frequency of Revision
Exhaust
Spread
Volatility
Deal Event Quantity of Deal Fig.4.14b
Weak relationships: but some clear effect
Quantity Quoted Quote Revision
Frequency of Revision /
Exhaust
Spread
Deal Event : : ; : _ - - - - - - / - - - - - - Volatility
~ Quantity of Deal Fig. 4.14c
Questionable relationships
162
Charles Goodhart, Takatoshi Ito, and Richard Payne ~bt
== a o +
ht == (30
+
a1~bt-1 (31 s ;-1
+
+
a 2BDE t - 1
(32 ht-1
+
+
St'
St
(33 BDE t-1
I It - 1 - N(O, ht ),
+ (34 BD t-1'
For brevity's sake, we report the results only for the bid side of the deutsche mark/dollar, presented in tables 4.30 and 4.31. Taking first the estimations for the quote revision equations, note that the autoregressive parameter, aI' is negative in all cases but with a consistently greater magnitude for the activity scale data, as previously reported when comparing tables 4.18 and 4.19 and tables 4.20 and 4.22. As before, the deal exhaust indicator has the expected negative effect on quote revisions. Then, next inspecting the volatility estimations, GARCH effects are present in both data sets. As shown in tables 4.30 and 4.31, the parameters (31 and (32 are significantly different from zero in a standard GARCH(l,l). We then examined whether deals affected quote revisions in an indirect manner through the underlying volatility series. This was done by adding lagged deal and deal exhaust indicators to the simple GARCH framework. For our activity scale data, we could not uncover any real effect of deals on volatility. Neither of the previously defined variables, BDt - 1 and BDEt _ 1, entered significantly into our
GARCH Estimation, Including Transactions, Calendar Time Data
Table 4.30
GARCH +bde,_1 +bd,_1
0. 0
0. 1
0. 2
~o
1.2e-5 a l.4e-5 a 1.3e-5 a
-.166a -.144a -.148 a
-2.2e-4a -2.6e-4a -2.8e-4a
1.2e-ge 5.7e-9 b 7.le-9 a
f31
f32
~3
.29ta .113 a .213 a
.874a .734a .667 a
2.4e-8 b
~4
2.7e8 b
Note: Presentation of the estimated parameters of the specification described in section 4.4.3 for the five-second data set plus extended specifications including lagged deal and deal exhaust indicators in the volatility expression. aSignificantly different from zero at I percent. bSignificantly different from zero at 5 percent. elnsignificantly different from zero.
Table 4.31
GARCH +bde,_1 +bd'_1
GARCH Estimations Including Transactions, Tick-by-Tick Data 0. 0
0. 1
0. 2
~o
~I
~2
~3
5.4e-5 a 5.4e-5 a 5.5e-5 a
- .436 a -.423 a - .433 a
- 3.3e-4a -3.le-4a - 3.le-4a
1.le-8 a 1.3e-8 e l.4e-8 e
.332a .252a .241 a
.68ta .724a .730a
-1.2e-8 e
~4
-1.2e-8 e
Note: Presentation of the estimated parameters of the specification described in section 4.4.3 for the activity-based data set plus extended specifications including lagged deal and deal exhaust indicators in the volatility expression. aSignificantly different from zero at I percent bSignificantly different from zero at 5 percent. elnsignificantly different from zero.
163
A Study of the Reuters 02000-2 Dealing System
estimation, and, indeed, their negative sign is implausible. But, when we moved to the five-second data set, the results were markedly different. Both of these variables had a positive effect on volatility, significant at the 5 percent level. Maybe when the quiet no-change observations are excluded in the activity-based data, the slighter persistent effects of deals on volatility become drowned out in the noisier "news." Indeed, in all cases, the GARCH estimates are far more significant in the five-second, clock-time data set. Comparison of the t-statistics between the activity and the five-second data shows those in the latter case to be far greater. This does not, however, imply that GARCH-type phenomena are better addressed in clock time than in activity time. It has been suggested that GARCH effects apparent in clock-time data may be the result of the transformation of a uniform, latent process that evolves on a different (activity) scale (see Stock 1988). This could underlie the diminished significance of the GARCH parameters in the tick-time results. This, however, is a subject for further research and is not pursued further here. 4.4.4
A Comparison with Hasbrouck's (1991) NYSE Study
Finally, in his 1991 study of the NYSE, Hasbrouck studied a bivariate VAR of the interrelations between deals (and/or deal quantities) and price revision (taking the middle of the bid-ask quote), scaling by activity, tick time. Here, we show his main results (which he gives in his table lIon p. 194) and our replication from our own data, both in tick time as he ran the regressions and in clock time, here reported in table 4.32. Since the scales of the price changes in the two markets (NYSE and foreign exchange) are markedly different, the differences between the absolute sizes of the coefficients should be ignored and are not shown (but are available from the authors). What matters is the size and pattern of the t-values, as shown in table 4.32. This shows that the equation for price quote revisions, the a and b t-values in the first columns, are qualitatively similar. In both cases, although considerably more strongly in the foreign exchange data, both in the clocktime and in the activity-time equations, there is significant negative autocorrelation, and in both cases quote revisions are strongly positively related to prior deals (i.e., a sell causes a drop in prices and a buy an appreciation). Like several other economists, Hasbrouck tended to dismiss the negative autocorrelation, noting that it "may simply arise from measurement error" (1991, 195); our repeated findings of such negative autocorrelation on high-frequency foreign exchange data make us believe that this ~nding cannot be brushed aside in this fashion. In the activity-based foreign exchange equation (column B), we can explain considerably more of the fluctuations in quote revisions than Hasbrouck, but this is primarily due to the stronger autocorrelation. (Hasbrouck does not report the F -statistic showing the combined effect of the X o variable on r [in our case it is 10.25], but a look at the comparable t-statistics suggests that the combined effect may be somewhat stronger in his equations.) The main, qualitative difference between his results and our own comes in
164
Charles Goodhart, Takatoshi Ito, and Richard Payne
Table 4.32
Estimates of the Bivariate Vector Autoregressive Model B
A aI a2 a3 a4 as bo
bI b2 b3 b4 bs
R2
-7.22 -.67 -.17 -1.31 -.14 15.15 6.83 .46 .87 -.30 .94 .096
A
B
C
-13.44 -6.05 -1.80 -.46 .41 10.16 7.20 4.66 1.24 2.03 .085
-1.47 -.07 -.72 .57 1.21 4.82 2.45 4.15 .67 .87 .038
-1.33 .94 .59 -.25 .79 2.07 2.53 1.25 3.73 1.85 .005
C
-8.19 -7.09 -7.22 -1.65 -5.65 -.69 13.53 .42 1.87 .69 3.55 .068
-16.16 -7.38 -5.36 -6.39 -3.00 7.57 2.87 3.09 .13 2.61 2.36 .175
c1 c2 c3 c4 Cs
dl d2 d3 d4 ds
R2
Note: We estimate a five-lag near-VAR involving rr' the revision in the quote midpoint, and X Or' the trade indicator variable. The VAR is not exact as the trade indicator is assumed to have a contemporaneous effect on quote revisions, as shown in the system below: 5
rr
=
L i=I
5
aJr-i
+ L bixor - i + Vir' i=O
5
X Or
=
L i=I
5
CJr-i
+ L dixor - i + V 2r · i=I
t-statistics are reported for each of the estimated parameters. Column A reproduces Hasbrouck's results, column B gives our equivalent activity-scale results, and column C gives our results on a clock-time basis.
the second set of equations for the event of deals. In Hasbrouck's equation, price quote revisions have a significant negative effect on deals; the first two c coefficients have t-values well below -2. In our own work, as reported in tables 4.15 and 4.16 above, price quote revisions have little, or no, effect on the probability of deals occurring, and this (negative) finding recurs also here. Both in Hasbrouck's results and in our own, there is positive autocorrelation in deal events, slightly stronger in his case than in either of our two runs. So, although the fit in all cases is close to zero, Hasbrouck can "explain" rather more of deal eventuality than we can. Hasbrouck notes that "a negative relation between trades and lagged quote revisions is consistent with inventory control effects since a monopolistic marketmaker with an inventory surplus would reduce his quotes to elicit more purchases. It is also consistent with the price experimentation hypothesis of Leach and Madhavan (1989) in which the marketmaker sets quotes to extract information optimally from the traders. These possibilities are deserving of further study" (1991, 295). In this further study, we find that, in our data sample from a market with many competing marketmakers, there was no indication of any significant (negative) effect between trades and lagged quote revisions. In addition, we examined whether our results were robust to a longer lag structure (ten instead of five); the answer was yes. We were also able to replicate with our data the exercises done by Hasbrouck (1991) in his tables III (p. 198) and V (p. 203). In table III, Hasbrouck examines the interrelations be-
165
A Study of the Reuters D2000-2 Dealing System
Table 4.33
Comparison of Coefficients
x
Own Lags Hasbrouck Our data
32.29 138.69
.98 2.72
5.89 -.018
-3.24 .118
tween r, price quote revisions, x o' the event of a deal, x, the size of a deal (- for sales, + for purchases), and x 2 = x o• 1 X 12 • Our results (the data are available on request from the authors) show generally less effect of the deal (x) variables on quote revisions. Unlike in Hasbrouck, neither the x nor the x 2 variables have a significant effect on r; only the X o variables do. Like Hasbrouck, we find that none of the variables, even the lagged dependent variables, can help much to explain x, the size of deals; indeed, as in his study, the lagged event of a deal X o is very slightly better at explaining x, the size of deals, than lags of x themselves. Again, as in Hasbrouck's work, neither the size nor the squared size of deals, x and x 2 , has any effect on the eventuality of deals, x oindeed, even less in our data than in his. Finally, we look at the determinants of the spread. In his table V (1991, 203), Hasbrouck regresses the spread, for his particular equity share, Ames Department Stores, on its own (five) lags and the absolute values of the current and five lagged values of x o' x, and x 2 • The t-statistics for the sums of these variables-and for comparison on our data (activity scale only)-are as shown in table 4.33. This shows that the extent of the positive autocorrelation of spreads is even larger in our data than in his. Otherwise, the significance of deals in our data set is rather less than in his and works in our own case primarily through x o' the event of a deal, rather than its size (or squared size). In particular, Hasbrouck finds some general tendency for the effect of x (on quote revisions and spreads) to be positive and for x 2 to be negative, which we do not find in our data set; but this is very likely because of the manner in which deal size was limited by the usually small size of the quote on offer in our data set, as earlier described. 4.4.5
Conclusions
It is now time to summarize this long, and often quite complex, study of the interrelations and determinants of the variables that can be extracted from D2000-2: for example, event, price, and size of deal, and whether an order exhausts the prior quote; the frequency of entry, price, size, and volatility of prices for both the bid and the ask; and the spread between them. Let us do so by reviewing our main findings. 1. Unlike the price quote series, which exhibits highly significant negative autocorrelation at high frequencies, the transaction price series exhibits no strong signs of autocorrelation (there is an insignificant negative first-order
166
Charles Goodhart, Takatoshi Ito, and Richard Payne
autocorrelation balanced by just significant higher-order positive autocorrelation). If one could not observe the "bounce" between deals at the bid and ask, the transactions series would then appear to exhibit weak negative first-order autocorrelation. 2. Tests of length of runs of deals at the bid and ask suggest that these have a fat (long) tail, which in this data set appears to be associated with strong price trends. 3. Studies of interactions between the many variables available from D2000-2 suggest a close interrelation (nexus) between quote frequency and deals (two-way causality) and between quote (price) changes and the spread (two-way causality). These two nexuses are linked, in that a deal that exhausts the amount offered at a previously quoted price will cause a price change both directly and indirectly via its effect on the spread (both directly and again indirectly by raising volatility). Deals that do not exhaust the amount on offer have a much weaker effect. There are only weak relations (in either direction) between the quantities (posted) and any of the other variables in the system. 4. Unlike a single dealer system, where the dealer will normally adjust both bid and ask quotes simultaneously, in this multiple competitive dealer system the bids and asks are normally input by different banks. There is no automatic reason why bid quotes should be revised in response to changes in ask quotes (or deals). In practice here, they did not respond much to such activity on the other side. Instead, price changes on one side primarily affected the spread and thence gradually the quote on the other side, with the spread acting as an error correction mechanism between the cointegrated bid and ask series. 5. The main pattern of relations reported in point 3 above appear to be encouragingly robust, as evidenced by the similarity of t-values, to changes in either the periodicity or the scale over which the regressions were run. 6. On the other hand, the GARCH equations varied considerably when run in clock-time rather than on an activity scale. The results for the former were more intuitive. 7. We were able to run an exact comparison, and replication, of Hasbrouck's (1991) study of transaction/quote relations in the NYSE. The main difference between us is that in his study lagged quote revisions have a significant (negative) effect on deals, whereas there is no such interaction in our data set.
4.5
Tailpiece
We have already summarized our main findings at the ends of sections 4.3 and 4.4. Here, we wish to emphasize again how short our data period was, only seven hours. Our findings should, therefore, be treated with due caution. By the same token, there would be considerable value, not only to academics but also to practitioners, in obtaining additional data of this high-quality format. We hope, and expect, that such data will become more widely available soon.
167
A Study of the Reuters D2000-2 Dealing System
Appendix Reuters D2000-2 Data 1.6480,------------------------------------,
1.6470
I----------~------------------------l
1.6460
f--------------------------------~-l
,)A../l.
1.6450 J . - - - - - - - - - - - - - - : o : - - l - - - r r - - - : - - - - - - - - - - - - - - : - - - : - - - - + - - - - i - - - l
..:... "'\~.~.::.:.~.: . .:.~.~~~~~~~~~.~.~~~.~.~~~~~~~~ :·~...T·~·~·::~·~·~·~·~·~·~:·~·~·7.:.: ......... .
1.64401-------------------------------------+ 1.6430
1--------~====================-------1
6
7
8
9
10
11
11
~!~. P..~~ Bid
Ask ~action ~~.~~.~
12
13
14
15
16
17
18
19
20
traractlon
4.5,------------------------------------, . ··-.. .1·-.-..· · · -·-· . ·--..- . . . , . . . . - - - - - - , - - - - - - - - , - , - - - - 1
4.01------~=----------·-··
3.51--------------------H,...---1-----_+_--------+--l1--__;
3.0 1----------------,.~----:T"i~~.'~·,I------1I------+-----:-~--.--...,...---+---1I----; 2 . 5 1 - - - - - - - - - - - - - - - ; . > - - - " ' - + i:~,--'--,--+------+---~__II____1t___+__+_-__;
:----J.---,--.. . . . . . .
....
2.0 f----------~----~:..,..: ----:-:-.;~\ . ;,.,. ",;-,;---."\ .,---,r---~: \--+----+--I............-+-N-r-:--n-f:\' 1 . 5 1 - - - - - - - - - - - - - - - 6 - 1 ; : -.......... \ -+\----+/~::---+-----IIJ---+-i --1--.........;:+\-+--+-+I-I--ttI~--fT\--I\\j T-:
,T-,
1.01----------------~_.....;..c,...---I.----l~-_---J'-----.J..--.:...'O""""......:....-:.-L~--L.....I---.u..L-...b\L..J
0.5 0
8 9
-7
4
1'0
1'1
1'1
12
Ask units offered Bid Urt~.91fered
Fig. 4A.l
Deutsche mark/dollar hour 0: 0-20
1.6480,--------------------------------------,
1.6460
....... 1.6450 1.6440
~u.': : :.:>"<:.: .:.:.: ; : ~ ::::::::e
\.:i\..
\r~;
; .-e:~.·p~. -·.!i.:~' .- ~,!!D~=:l·~~~~~\Vt~:~:~5~;;:::\! . ,: . .
..
.·.·.•. .·.•.
~
---c:\~cA....0"=!'.. i~:- - - - - - - - - - - - - - - - - - - - +
1=--
"=i\
···.~·. ·····V
1.6430 r - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - I
21
21
22
23
24
25
26
27
28
29
30
31
32
32
33
34
35
36
37
38
39
40
35
36
37
38
39
40
Ask tra~aetion ~~ J?~_~ ~!~..P..~~ Bid tra~action
6.0 5.0 4.0 3.0 2.0 1.0 0'°21
21
22
23
24
25
26
27
28
29
30
31
32
32
33
Ask units offered Bid Urt~_91fered
Fig. 4A.2
Deutsche mark/dollar hour 0: 20-40
34
168 1.6480
Charles Goodhart, Takatoshi Ito, and Richard Payne
r----------------------------------,
1.6470/---------=-================------------j 1.6460
t---------~-~~~~===============---===t
:~\.~~~\,::~~;~·;;~;::~::{:~;;~;~ft::;,,;\-"'?"'i7:!3L:::ei}1)"~;;~:.::.:.;.""~':.~.::;)::::::,';~
1.6450
1.64401--------------"'----'---------------------i....-------l 1.64301----------------------------------1
41
41
42
43
44
45
46
47
48
49
50
51
52
52
53
54
55
56
57
58
59
Ask tra~aetion ~~.J?~~ ~!~ . ~~ Bid ~aetion
6.0.-------------------------------------. 5.01-----+-----------~~~---__t------_____,_o;.______-______r--___;
4.0 I-------t+r..,...-----rr--n-----.,,----f-----+.+--+-,.------+I--------i----,...,~i..-.;\i\-----; 3.0 1-------r-T+-L-r+r-r-..L-.L...L..t--n--rr--r-rH-r---i----'-'------'f-'r-r---+t--~----__!_..;..........~~'-------___;
\f\(iig
2.0 1--r+.-,..u....-:...:...J-.¥-----.----t--++--H-~J_+f.l____._---1;...-----'--:.+" 1.0 t----''--L...!-_---'---'-'-_--'----'--'''"......a...-.I----Jr....u---I
~
--L-...L.--I~ .........~__L.._.&...!.___'_'_:..L.l..__""'_ _
0.0 41
41
42
43
44
45
46
47
48
49
50
51
52
Ask units offered Bid
Fig. 4A.3
1.6480
52
53
54
55
56
57
58
_........'''--'-_ __ _ _ ;
59
ur}~.pffered
Deutsche mark/dollar hour 0: 40-59
r-------------------------------------,
1.64701-'----------------------------------·1
1.6460 ~--------=~--=~=----------~------·I
1.6450 1.6440 1.6430
~--------------------------------I
5
6
8
9
10
11
11
12
13
14
15
16
17
18
19
20
Ask tra~aetion ~~ .~~.~ ~~~..e~~ Bid ~aetlon
6.0r--------------------------------......_---.
~:~
_. ,. -+ !l~.-~ ._ _ _ - - +f-. .,. - - r. ,-,- - -,.£_ +_ ",. .,. .,_-_ _._ _,. . _LJ-.,. . ,J_!. .LI. - -:. .-_J_-I
\.Jli!i...'----r-""'""r-'i--i...!!,/-+\
1.0
0.0 0
0
3
4
7
8
9
10
11
11
12
Ask units offered Bid ur}~. pffered
Fig. 4A.4 Deutsche mark/dollar hour 1: 0-20
13
14
15
16' 17
18
19
20
169
A Study of the Reuters D2000-2 Dealing System
1.6480 . - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - ,
1.6470 1.6460
f-----------------------------------------j
..
1.6440
......
....
1.64301---------------------------------------1
21
21
22
23
24
25
26
27
28
29
30
31
32
32
33
34
35
36
37
38
39
40
Ask tra~action ~~ .~~.~ ~~.~ ..P~~ Bid trai8ction 6.0,..---------------------------------------,
5.01-------------------------------F==t······ 4.0
2.0 1.0
21
0.0 21
22
23
24
25
26
27
28
29
30
31
32
32
33
34
35
36
37
38
39
40
Ask units offered Bid ul1.~.pffered
Fig.4A.5
Deutsche mark/dollar hour 1: 20-40
1.6480,..--------------------------------------.,
1.6470
1.6460 1.6450 f - - - - - - - - - - -
1.6440
_
__ _
---;.1-------+
_
-'.·--·~~t~~.·_·J\.· -0·.!\_O:6· .>~ ·.~P:·:.?~.~ ~ ~cr.~.~ :.:·.;>:;.L=-:l·r:~· .~!.~=~l.o:.::;~::..£;::.:i'{ .
!>-!.\...-
:/ v..·..
':] ~.:'
1.64301=--------==:..::..:.......--------------------------
41
41
42
43
44
45
46
47
48
49
50
51
52
52
53
54
55
56
57
58
59
Ask tra~8ction ~~.~~.~ ~!~. p..~~ Bid ~action 6.0..-----------------------------------------.
0.0 41
41
42
43
44
45
46
47
48
49
50
51
52
52
53
Ask units offered Bid ul1.~.pffered
Fig.4A.6
Deutsche mark/dollar hour 1: 40-59
54
55
56
57
58
59
170 1.6480
Charles Goodhart, Takatoshi Ito, and Richard Payne
r--------------------------------------,
1.64701---------------------------------------1 1.6460 1 - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - 1
1.6450 t - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - I
1.6440 1.6430 1.6420
L . . L - - - . . J ' - - - - - I - - - . . I _ . . . . I . - - - L _ . . . . I . - - - L _ - ' - - - - L _ . . . . I - . ........._.....L.--...L._.....L.---....._...L.----L._...L.----L._...L.---L.._...L.-~
5
7
9
10
11
11
12
13
14
15
16
17
18
19
20
Ask tra~8ction ~~.~.~ ~~~ ..~~ Bid ~action
7.0.----------------------------------------.
6.0t--------------yr-------------------------1 5.0f--------------++-------------------------1 4.0t---------~---r'-L..--__._-~
3.0
~:~
JI. .!
0.00
,
0
"." 3
:1'• •••_ ..""
4
""",,,,, , " ""."._
5
6
""•••••"••".._.._
,
".."
"""
,,:.••__
..._..._....IrRACT
.-.
7 8 9 10 11 11 12 Ask unl!l..Pffered Bid urtlJ!l.p1f8red
13
14
15
1ti
":7
18
19
20
Fig.4A.7 Deutsche mark/dollar hour 2: 0-20
1.6480
r------------------------------------,
1.64701--------------------------------------+ 1.6460 1 - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
1.64501-------------------------------------.1
1.6440 1.6430
21
21
22
23
24
25
26
27
28
29
30
31
32
32
33
34
35
36
3i'
::I(J
39
40
$
~
~
~
~
~
Ask tragsaction ~~J!~~ ~~~..~~ Bid tra~8ction
6.0 5.0 4.0 3.0 2.0 1.0 0.0
21
21
~
~
~
~
~
~
~
~
~
~
~
~
~
Ask units offered Bid urtlJ!l.p1fered
Fig. 4A.8
Deutsche mark/dollar hour 2: 20-40
M
171
A Study of the Reuters D2000-2 Dealing System
1.6480.-----------------------------------, 1.6470
f----------------~-----------------I
1.6460 t - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - I 1.6450 t - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - I 1.6440
_
·· ·· · ·· · · · - - - - - - - - - - 1
:·:::,::.:~,~::~~·~:'.::~~,~::!i,f.'1.::f?:::::::~.:!~:~:~-~~=:-!.~-~~·:~:=/:::::-~i~~~·::::-::\j:\l?:::::C'::~,,:,:.:0.,:,:Q
1,&130
41
41
42
43
44
45
46
47
48
49
50
51
52
52
53
54
55
56
57
58
59
Ask tragJaction ~~_~~~ ~~~..~~ Bid traraetlon 4.5 4.0 3.5 3.0 2.5 2.0 ~n 1.5 _.,u), 1.0
: ::
..... _........ /
I I :.~\: I\ ·1
.l__i
~~
~
\~
~
~
:1\
~ ~
I
d
I
\ \
~
~
~
~
~
~
~
\ I
1
Iii: ~
1
:j
; i
I f\
Q
I I
~
~
M
\ \
I
~
$
~
A
1\
~
Ask urillLPffered Bid u.,~_ pffered
Fig. 4A.9
Deutsche mark/dollar hour 2: 40-59
1.6480 , . . . . - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - ,
1.6470
~---------
,.,
-.__
_-
__..,_ ,_ __ _-._-_
__
·· · . · ·····
1.6460
I-----------------------------~--=~·I
·1
1.64501----------------------------------1 1.64401--..!r---------------------------------1
L~D -~:~=;~::~~~~:j~--:::~ 5
6
7
8
9
10
11
11
Ask tra8'8ction ~~.~~.~ ~~~..p.~~
12
13
14
15
16
17
18
19
20
19
20
Bid traasaction
6.0 5.0
~
3.0 V
2.0 1.0
\
,
4.0
A
0.00
\
3
4
7
8
9
10
11
11
12
Ask units offered Bid u.,~. pftered
Fig. 4A.I0
"roo' ::
Deutsche mark/dollar hour 3: 0-20
13
1·_·
i
~
:f
i
1\ _._.._..__..._----..\
\
:j
i
14
A.
:~ N
15
: 16
'.17
18
172
Charles Goodhart, Takatoshi Ito, and Richard Payne
1.6480,..----------------------------------------, 1.6470 r - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - I 1.6460 1 - - - - - - - - - - - - - - 1.6450
r---------------
1.6440
t------------
"
"
--
r··..:_ _
- •.;.:.:==
.........-6, ~ ~~.. ~.~. ~.~............................
,!~_
_ _"
:.~.~:;.(p . "r-e.> '.'. . . . . . . .
=
21
22
23
24
25
-- - - - - - - - - - - - - - I 'O'__._•• _•.• _._._•.• ""=T
..!\
1.6430 1 - - - - - - - - - - -
21
~---------4
"
26
27
28
29
30
31
32
32
_.b!·
···
"
33
·'..····,,·····,,·..0
.
34
-----'--'-------f
35
36
37
38
39
40
.
Ask tra8'actlon ~~J?~~ ~!~ ~~ Bid trafactlon
7.0 ~---------------------------_._---------, 6 . 0 1 - - - - - - - - - - - -..· ·.. ·.. · -------------------+ 5.01-----------++-----------------r-------:-=-=-'~------t
++
4.0 )---, . - - - - - - - - 3.0
·
•..- ••. . - . - - - - - - - - / . ,
· ·· · ·
. . - - - - , ,,-,
· ·
T ..
--·····-··..·---·····
1-',
·-----+'-,...---..;..--;.-------1 ---+-+------.+----+-..,----1
·
2.0 1.0 0.0 21
21
~
~
~
~
~
~
~
~
~
~
~
~
M
~
~
$
~
~
~
$
Ask units offered Bid un~.9ffered
Fig. 4A.ll
Deutsche mark/dollar hour 3: 20-40
1.6480 , . . - - - - - - - - - - - - - - - - - - - - - - - - - - . - - - - - - - - - - - - - - - , 1.6470 1 - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - + 1.6460 1 - - - - - - - - - - - - - - - - - - - - - - - - - - - . - - - . - - - - - - - - . . - - - - - - 1 1.6450 1 - - - - - - - - - - - - - - - - -
.;"
:._-..- .._
~:::.::.:.:.:
. : :.:.:
.
1.6440 1.6430
I------...:.......-'====="-'~~~--------------------------+
1.6420
L..-L-_.l.--.......I--_..I.----L.,._-'----L.,._...L.-----L._....I.----L._.....L-----J._....L----I.-.-;.-L----l_--L...._'-----'--_.l.---.l...._..l...-----l
41
41
42
44
43
45
46
47
48
49
50
51
52
52
53
54
55
56
57
58
59
.
Ask tra8'action ~~J?~~ ~!~ p"~~ Bid trafaction
6.0,--------------------------------------.,
f.. . .'.. . . . ,
5.01-------..;......---'-----=---············· ······ . ···;.·~··,···
i idi
4.0 1 - - - - - - - - ,
3.0
~::! 0.0 41
-rT--------y-! !
---It-/+-+--:-
.~L!~'U__~::1...
-r--__.
41
42
43
44
45
46
~."
;
~1
/r.-.\
~--l.:--u.---,/ . . . "\'L-J
_.L<..-I:
47
,
:! ------......------
~
;..! !
-------------/
:.~
i;···_·..T
48
49
50
51
52
_---++-----1
:Io..oo--.-_...L....-J.-J':}jJ
L-..-
1 ... I,
Ask units offered Bid
Fig. 4A.12
..;.;
I ;
~.:
52
53
un~.9ffered
Deutsche mark/dollar hour 3: 40-59
54
55
56
57
58
59
l
173
A Study of the Reuters D2000-2 Dealing System
1.6480
r-------------------------------------,
1.6440
~-----
1.6430 1 - - - - - - - - -
_
8
10
11
_
_ _ _.._
11
12
-----------1
13
14
15
16
17
18
19
20
Ask tragsaetion ~~.~~ ~~~..P..~~ Bid ~action
6.0 5.0 4.0
I
3.0
I
...
!, ------, \ \U\~
;!.. L
2.0
,
1\
\
0.00
4
5
6
7
,
I
I \I \ L~~~ ____._. .
t_L\
1.0
I
I
~ I
I\
8
9
10
11
11
12
13
14
1
1:\
I 15
16
'17
18
19
20
Ask units offered Bid un~_9ffered
Fig. 4A.13
Deutsche mark/dollar hour 4: 0-20
1.6480,----------------------------------=-;
~ L~~tl~r~~~·~r.t:~\:·:~=:~=·=·;.;z=:~~~::c::: · · · · · · · : : :o:.:<: :~·
: : : ..o 1.6450
I----------------~----
1.6440
~------
1.6430
I---------~.~------------_
21
21
22
23
24
-------------1
25
26
27
28
29
30
31
32
32
Ask ~action ~~.I?~.~ ~~.~..p..~~ Bid
33
_--------------I
34
35
36
37
38
39
40
~
~
~
trawaction
6.0..--------------------------· 5.0 4.0
3.0 2.0 1.0 ~
~
~
~
~
v
~
~
~
~
~
~
~
Ask units offered Bid un~_9ffered
Fig. 4A.14 Deutsche mark/dollar hour 4: 20-40
M
$
~
~
174
Charles Goodhart, Takatoshi Ito, and Richard Payne
1.6480 , . . - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - , 1.6470 f - - - - - - - - -
,
_
-
---------------1
1.6460 1.6450 f - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - f 1.6440 f - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - I 1.6430 1 - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - 1
41
41
42
43
44
45
46
47
48
49
50
Ask tra8'8ction ~~ .I?~~
51
52
52
53
54
55
56
57
58
59
~~~..~~ Bid traa'8etlon
7.0r---------------------------------------, 6 . 0 1 - - - - - - - - - - . . - - - - - - - - - - ~--~-------~----------I 5.0 I---------io"""'~
4.01---------P.----T-'ri<>----, 3.0 I------~~"--'-'h--2.0 t---~...--,._,_,.___._-'----'-+---+-t-r--rr.,..,
m;"n'" ·
• ,.....-.
·
·,..,
···
50
51
1+ _ ·
,-L
1.0 41
0.0 41
42
43
44
45
46
47
48
49
52
53
52
54
55
56
57
58
59
Ask units offered Bid Ull~. pffered
Fig. 4A.15
Deutsche mark/dollar hour 4: 40-59
1.6580 . . . - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - , 1.6560
I-------------------...,----------------~·O~i :~.;t__t ..
1.6540
l---------------------------------~fJ~--1
1.6520
~--------------------------------:.=-.GQ-"--,'-----:--------j
etP
!.?:T···.. h'?::=:f!i~:·J'
1.6500
\nf
t
~~~.-=-~-~-:~-~-~-.~-:-=-~-;~-=~:-=-2-~-E-~-=Y-e-~-i-i-=-~-~-~-~-~-;~-:_-.:-;-=-=-~-~-~-=~~:A__ 5
6
7
8
10
11
.
Ask ~8ction ~~.r:>~.~ ~~~ p..~~
11
12
13
14
G 15
16
17
18
19
20
Bid tra~action
7.0.---------------------------------------, 6.0~--~-----------------------------.----1
5.01-----=+-------------------t-i"'t---r-n----:-==rw.--:-.o.=o.:-tr-------j 4.0~
·
---+-~-----------.......".,.---
·1
· .. ··
·It
3.01-..·------,,.............-.---. 2.0 1.01-------U---U----I-J.. . . .'----'-..L..o-...1-I-..L----'
0.00
4
7
8
9
10
Ask units offered Bid
Fig. 4A.16
11
11
12
Ull~. pffered
Deutsche mark/dollar hour 5: 0-20
13
14
15
16
17 - 18
19
20
175
A Study of the Reuters D2000-2 Dealing System
1.6580 , . . . . - - - - - - - - - - - - - - - . . - . . . . - - - - - - - - - - - - - - - - - - - - - - - - - ,
: : : :
1.6570 0 1.6560 1.6550
J\ Ii
:{) . "
-1-~dr:1.~:=1-tLif: :·.,·. ·.• !.:~ ~:.~ :.
,·tJ~
:
e:.·
\.I -.•....-..i;."r..
v
1-------+-'---4't'-+-.I,..~H-··~~_b--:---- L..._.,.;.... .
.•;-
.J.:!.\:,
oi.:.·,.,·." : n ·iT~1';=; ..y.' '~....'....:"
,::~
.......... ....-:-._......
.
:....
\
,.
...
:'.;: :
....
VH \:
!
if
::: :~~~~~~:. ~. . ~.i~ :\'~.-..- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~:~~~:~~~= : ;..-j;..-·
21
21
22
23
24
25
26
27
28
29
Ask tra~action
29
31
~~-'?~.~
32
32
33
34
35
36
37
:18
39
40
~~~ ..p"~~ Bid trawact10n
7.0,-----------------------------------------, 6 . 0 1 - - - - - - - - - - - - - - - - - ·..· . - - - - - - - - - - - - - - - - - - - - - - - - 1
5.0 ! - - - - - - - + - - t = = ' \ - - - - - - , R - - - - - - -
!
-------------.---:-----l
,.
4.0 1------11----1----.--+-----,.--.-++------
,
3.0 f---o------+J.----t------;.+r---lHtt-+-------.. !:
r-+,
·..-----------T.-~-I
·'·.. If-:;·.·....,
~,...-
,..,.. ------,--.--~--,-------~7-----f
2.0
1.0 21
0.0 21
22
23
24
25
26
27
28
29
29
31
32
32
33
34
35
36
37
38
39
40
Ask units offered Bid ull~.pffered
Fig.4A.17
Deutsche mark/dollar hour 5: 20-40
1.6580 , - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - , 1.6570 t - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - t
1.6560 \------..,----,-------,-.-......,------------.
1.6550
1.6540 1.6530
....... f - - - - - - - - - - - - - - - - - .....- -.............- - - - - - - - - - - - - - - - - - - - - j
41
41
42
43
44
45
46
47
48
49
Ask tra~action
50
51
51
~~ .~~.~ ~~.~ ..P..~~
52
53
54
55
56
57
58
59
Bid traiaction
7.0,....----------------------------------------, 6.0 f - - - - - - - - - - - , - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - I
. ·. ·.. ·.. . .·.. .·.. ·.. I""",.,Fl·· n .. l l r l - - - - - - - - - - - - +
5.0f------~-----------------
4.01------+-¥.'------.------.--------------,-......----------·-+--+-f--+---+-I--tI++------------I
3.0 2.0
1.0 0.0 41
41
42
43
44
45
46
47
48
49
50
51
Ask units offered Bid
Fig. 4A.18
51
52
53
Ull~. pffered
Deutsche mark/dollar hour 5: 40-59
54
55
56
57
58
59
176
Charles Gotluhart, Takatoshi Ito, and Richard Payne
1.6620...---------------------------------------r......--, 1.6600
I------------------------------~------
1.6580
~---------------------
5
6
7
8
9
10
10
11
",.'
12
13
14
15
16
17
18
19
b~' 20
Ask tra~ction ~~.~~ ~~..~~ Bid trarctlotl 7.0.....-------------------------------------, 6.01--------------,..,.......,·,1"·········· · · · · · ·..·········· _-------------i . ···· . ·············· . ···(···;··",.. t
5.0F--------------~·1'1··················
4.01-----------.-----.....,...-- · · ·. · ·
111 ·
3.0 I-------~--,--_ . . ·.~···.··r
.
2.0 1.0
'-.-.L,o..;'--l
f-_-..l.-~_-"'~~_~--'-L-'--J.
0.00
6
4
··
····..···· ..···· ·
········ ···..· ,."':...." ;....t- _+..;
_
···········..- - - - - - - - - - - - - - - - - - " 1
_
_
-------,-------------"1
···J+h ·..··· ·....·· ..··· ......··,; ..··lf··r+·m··.. ··~,. !·····..·..····· ..·..···..··"!.. 'r ···..· ·..· · · · · . ·.. _-- ~-------n----(
I·..,·.. ···· ..
7
_,~_--I
8
9
10
.:..
10
:L..J.
11
--'--
12
-"
13
14
15
16
17
18
19
20
Ask urUJLpffered Bid ufl~.pffered
Fig. 4A.19 Deutsche mark/dollar hour 6: 0-20
1.6630
r------------------------------------,
1.6620
r-------=
_.._ _
-----------1
;:: lV\.b-.-~!\-.o;:":~nr~]f~- :-:~: :~ -!'{\~: ·:·;~-!~i~ ~fi~ 9]:\: : ~ -.- -:-~- .- - .•~-~ 1.6580
. _
21
21
22
23
24
25
26
27
28
29
30
31
__
32
32
_.__.. _
-_
33
34
L.i - - - - - - - . . . . . - ' - - - - - - - + 35
36
37
38
39
40
~
~
~
~
~
~
Ask trBBJ8ctlon ~~.f?!i.~ ~~~..~~ Bid tnyactlon
6.0 5.0 4.0 3.0 2.0 1.0 0.0
21
21
~
~
~
~
~
~
~
~
~
~
~
~
~
Ask ullU!..9ffered Bid Ufl". pffered
Fig. 4A.20 Deutsche mark/dollar hour 6: 20-40
M
177
A Study of the Reuters D2000-2 Dealing System
1.6630.-------------------------.....--------------, 1.66201------------------------------------1
41
41
42
43
44
45
46
47
48
49
50
51
52
52
53
54
55
56
57
58
59
Ask tra8'action ~~o~~o~ ~!~. p'~~ Bid tra~aetlon 7.0r----------------------------------------.
::~I--------------.-----n====:r==-=-=a~ .
i:~t~.\ i·\f--~\~=:==~T~t~f~M_t~-=[~=-L~-. -_.-.-.
-------l
0.0 41
41
42
43
~
~
~
Q
~
G
~
~
~
~
~
~
M
~
~
~
~
Ask units offered Bid un".p1fered
Fig. 4A.21
Deutsche mark/dollar hour 6: 40-59
1.6600,---------------------------- --------, 1.6595 r-----..............-----------------
"'
------------1
1.6590
f--------T---------~~~~-""'------------------------I
1.6585
I - - - - - - - - T - - - - - - - -..;-o;.
E~>,·· ·.bB--_--_------_ _--------------/
'0''''' ...,
1.6580 1.6575
5
Ask traBBactlon ~~-'?~.~ ~!~..P..~~ Bid tra~aetlon 6.0,--------------------------------------------.
Ask units offered Bid unl~.pffered
Fig.4A.22 Deutsche mark/dollar hour 7: 0-20
178
Charles Goodhart, Takatoshi Ito, and Richard Payne
References Admati, A., and P. Pfleiderer. 1988. A theory of intraday patterns: Volume and price variability. Review of Financial Studies 3:593-624. - - - . 1989. Divide and conquer: A theory of intraday and day-of-the-week mean effects. Review of Financial Studies 2, no. 2: 189-223. Baillie, R. T., and T. Bollerslev. 1991. Intra-day and inter-market volatility in foreign exchange rates. Review of Economic Studies 58:565-85. Bank for International Settlements (BIS). 1993. Survey offoreign exchange market activity. Basle: BIS, Monetary and Economic Department. Bessembinder, H. 1994. Bid-ask spreads in the inter-bank foreign exchange markets. Journal of Financial Economics 35, no. 3 (June): 317-48. Blitz, J. 1993. Foreign exchange dealers enter the 21st century. Financial Times, 13 September 1993, 19. Bollerslev, T., and I. Domowitz. 1991. Price volatility, spread variability and the role of alternative market mechanisms. Review of Futures Markets 10:78-102. - - - . 1993. Trading patterns and prices in the interbank foreign exchange market. Journal of Finance 48, no. 4 (September): 1421-43. Bollerslev, T., and M. Melvin. 1994. Bid-ask spreads and volatility in the foreign exchange market: An empirical analysis. Journal of International Economics 36:355-72. Cohen, K., S. Maier, R. Schwartz, and D. Whitcomb. 1981. Transaction costs, order placement strategy, and existence of the bid-ask spread. Journal ofPolitical Economy 89, no. 2:287-305. Dacorogna, M. M., U. A. Muller, R. J. Nagler, R. B. Olsen, and O. V. Pictet. 1993. A geographical model for the daily and weekly seasonal volatility in the FX market. Journal of International Money and Finance 12, no. 4:413-38. Dimson, E., ed. 1988. Stock market anomalies. Cambridge: Cambridge University Press. Domowitz, I. 1990. The mechanics of automated trade execution systems. Journal of Financial Intermediation 1 (June): 167-94. - - - . 1993. A taxonomy of automated trade execution systems. Journal of International Money and Finance 12, no. 6 (December): 607-31. Ederington, L. H., and 1. H. Lee. 1993. How markets process information: News releases and volatility. Journal of Finance 48, no. 4 (September): 1161-91. Flood, M. D. 1994. Market structure and inefficiency in the foreign exchange market. Journal ofInternational Money and Finance 13, no. 2 (April): 131-58. Foster, F. D., and S. Viswanathan. 1990. A theory of interday variations in volumes, variances and trading costs in securities markets. Review of Financial Studies 3:593-624. - - - . 1993. Variations in trading volume, return volatility and trading costs: Evidence on recent price formation models. Journal of Finance 48, no. 1 (March): 187-211. French, K. R., and R. Roll. 1986. Stock return variance: The arrival of information and the reaction of traders. Journal of Financial Economics 17:5-26. Glass, G. R. 1994. Multinet's FX netting solution. Proceedings of the International Symposium on Banking and Payment Services, 152-67. Washington, D.C.: Board of Governors of the Federal Reserve System. Glassman. D. 1987. Exchange rate risks and transactions costs: Evidence from bid-ask spreads, Journal of International Money and Finance 6:479-90. Goodhart, C. 1989. "News" and the foreign exchange market. Pamphlet. Manchester: Manchester Statistical Society, 17 October.
179
A Study of the Reuters D2000-2 Dealing System
Goodhart, C., and R. Curcio. 1991. The clustering of bid/ask prices and spreads in the foreign exchange market. Discussion Paper no. 110. Financial Markets Group, London School of Economics, January. Goodhart, C., and A. Demos. 1990. Reuters screen images of the foreign exchange market: The deutschemark/dollar spot rate. Journal of International Securities Markets 4 (Winter): 333-48. Goodhart, C., and A. Demos. 1991 a. The Asian surprise in the forex markets. Financial Times, 2 September, 13. - - - . 1991b. Reuters screen images of the foreign exchange market: The yen/dollar and sterling/dollar spot market. Journal of International Securities Markets 5 (Spring): 35-64. Griffiths, M. D., and R. W. White. 1993. Tax-induced trading and the turn-of-the-year anomaly: An intraday study. Journal of Finance 48, no. 2 (June): 575-98. Hasbrouck, J. 1991. Measuring the information content of stock trades. Journal of Finance 46, no. 1 (March): 179-207. Hasbrouck, J., and T. S. H. Ho. 1987. Order arrival, quote behaviour and the returngenerating process. Journal of Finance 42, no. 4 (September): 1035-48. Leach, C., and A. Madhavan. 1989. Price experimentation and market structure. Working paper. Wharton School, University of Pennsylvania. Lease, R., R. Masulis, and J. Page. 1991. An investigation of market microstructure impacts on event study returns. Journal of Finance 46: 1523-36. Lee, C. M. C., and M. J. Ready. 1991. Inferring trade direction from intraday data. Journal of Finance 46:733-46. Lyons, R. 1993. Information intermediation in the microstructure of the foreign exchange market. Business School, University of California, Berkeley. Typescript. - - - . 1995. Tests of microstructural hypotheses in the foreign exchange market. Journal of Financial Economics 39:321-51. Madhavan, A., and S. Smidt. 1991. A Bayesian model of intraday specialist pricing. Journal of Financial Economics 30:99-134. McInish, T. H., and R. A. Wood. 1985. An analysis of transactions data for the Toronto Stock Exchange. Journal ofBanking and Finance 14:441-58. - - - . 1990. A transactions data analysis of the variability of common stock returns during 1980-1984. Journal of Banking and Finance 14:99-112. - - - . 1991. Autocorrelation of daily index returns: Intra-day-to-day versus close-toclose intervals. Journal ofBanking and Finance 15: 193-206. Milller, U. A., M. M. Dacorogna, R. B.Olsen, O. Pictet, M. Schwarz, and C. Morgenegg. 1990. Statistical study of foreign exchange rates, empirical evidence of a price scaling law, and intraday analysis. Journal of Banking and Finance 14: 1189-1208. Petersen, M. A., and D. Fialkowski. 1994. Posted versus effective spreads: Good prices or bad quotes. Journal of Financial Economics 35, no. 3 (June): 269-92. Pictet, O. V., M. M. Dacorogna, U. A. Milller, and C. G. De Vries. 1994. The distribution of extremal foreign exchange rate returns and extremely large data sets. Preprint. Zurich: Olsen and Associates Research Group, 22 June. Roll, R. 1984. A simple implicit measure of the effective bid-ask spread in an efficient market. Journal of Finance 39: 1127-39. Stock, J. 1988. Estimating continuous time processes subject to time deformation: An application to postwar U.S. GNP. Journal ofthe American Statistical Association 83, no. 401 (March): 77-85. Wood, R. A., T. H. McInish, and 1. K. Ord. 1985. An investigation of transaction data for NYSE stocks. Journal of Finance 40, no. 3 (July): 723-39. Zhou, B. 1992. High frequency data and volatility in foreign exchange rates. Department of Finance, Sloan School of Management, Massachusetts Institute of Technology. Typescript.
180
Charles Goodhart, Takatoshi Ito, and Richard Payne
Comment
Richard K. Lyons
The authors do a lovely job with an important topic. The paper provides much information. Keeping it in perspective, however, is crucial. Accordingly, the first part of my comment provides perspective on precisely where these data fit in. The second part addresses the specific results of the paper.
Some Perspective This paper is about spot trading. It is important to keep this straight. For example, when the Bank for International Settlements writes of a $1 trillion daily "foreign exchange market," many market segments are being lumped together: spot, forward, swaps, futures, and options (see BIS 1993). Care should be exercised when using aggregated BIS statistics to discuss the spot segment in particular. The authors themselves occasionally lapse (e.g., when discussing the market share of automated dealing systems, they refer to BIS data that are not from the spot segment alone). Let me telescope further. Spot trading accounts for about half the foreign exchange total. Mark/dollar is the largest spot market by a margin, accounting for about a third of trading. Now, within spot markets, there are two main types of participants: dealers and customers. By customers I mean here any participant that does not provide two-way prices (e.g., corporate treasurers, investors, hedge-fund managers, liquidity traders, central banks, etc.). About 85 percent of spot mark/dollar trading is between dealers. 1 It is this interdealer trading that produces the D2000-2 data in this paper. Moreover, the data come from a particular type of interdealer trading, namely, brokered trading. There are two basic types, direct and brokered. Direct interdealer trades involve communication between the counterparties only. Price and quantity from these trades are not observed by others. In contrast, brokered trading involves prices that are advertised to dealers generally, as described in their section 4.2 (customers do not have access to interdealer brokers, electronic or otherwise). In spot mark/dollar, about two-thirds of interdealer trades are direct, and the remaining third are brokered. 2 It is important, in my judgment, not to overemphasize the distinction between electronic (screen-based) trading and voice-based trading. More imRichard K. Lyons is associate professor in the Haas School of Business at the University of California, Berkeley, and a faculty research fellow of the National Bureau of Economic Research. 1. Table 1-B in BIS (1993) reports that customer-dealer trades account for about 12 percent of the total in spot mark/dollar. Calling the remaining 88 percent interdealer would be an overestimate, however. That report includes a third category of participant called other financial institutions that accounts for another 12 percent. This category includes nonreporting banks, which in many countries includes investment banks, some of which are important in dealing. (That dealers are included in this third category is evidenced by the significant brokered trading of this category; in general, only dealers have access to brokers.) Since this third category does include some "customers" by my definition (e.g., insurance companies and pension funds), 85 percent is a reasonable conjecture for the interdealer share. 2. Note that table VI of BIS (1993) does not report brokered shares for the spot market alone. For spot market data, individual central banks provide more information.
181
A Study of the Reuters D2000-2 Dealing System
portant is the above distinction between brokered and direct trading, both of which have electronic and voice-based options. The D2000-2 system that the authors track competes with traditional voice-based brokerage. 3 Another Reuters system, called Dealing 2000-1, is an electronic means for direct trading. Like voice-based direct trading, only the counterparties communicate when using Dealing 2000-1. Thus, speaking of "electronic dealing systems" without separating direct and brokered trading can be misleading since they involve very different dissemination of information. With the above as background, here is a concentric rings model to organize the data sources referred to in the paper. There are three rings. The inner ring is direct interdealer trading. The Dealing 2000-1 data used in Lyons (1995; see also Lyons, chap. 5 in this volume) is from this inner ring. In mark/dollar, spreads in this inner ring are typically three to four pips for large banks when trading is active (London afternoonlNew York morning). The second ring is brokered interdealer trading. The authors' D2000-2 data are from this second ring. Spreads in this ring are typically five to six pips when trading is active (here, I have large brokers in mind, which D2000-2 was not in 1993).4 The third ring is customer-dealer trading. Although transaction data from this ring are not currently available, my experience with dealers is that spreads are in the seven to twelve pip range for large customers (circa 1993). I view the indicative FXFX data as targeted at this third ring. That is, for most customers, this indicative series is the best real-time indicator of where the market is trading. Clearly, FXFX is not targeted at dealers since live broker quotes are more informative and they are easy to monitor continuously. This leads to an issue that I do not believe has been addressed adequately in past work using FXFX: Exactly who inputs these indications at any given bank? And how? Stop to think for a moment about how rational it would be to pay a veteran dealer to input indicative quotes while trading, say, a billion dollars a day. No. It is much less expensive to hire a dependable young person to sit within earshot and intermittently type in a five or ten pip price based on where the dealer is actually trading. Better yet, why not build in some automaticity? For example, write a program that captures the dealer's firm Dealing 2000-1 quotes and widens them for customer consumption. A better understanding of this entry process would shed light on why the series has the properties it does. Let me suggest that the clustering of the FXFX spreads at 3. When I asked a spot mark/dollar dealer what market share broker systems like D2000-2 would have in five years, his response was essentially the following: "Currently, traditional brokers have about a third of the interdealer market, the rest being direct. Five years from now, I would guess that electronic brokers will have about half that third, traditional brokers the other half of that third, and direct will remain about two-thirds. There will always be a need for direct dealing." Admittedly, this is just one person's view, but tempered with experience nonetheless. 4. So why use a broker if direct prices are tighter (and brokers also charge a commission)? Smaller banks often do not have access to the tighter spreads among large banks. Large banks often prefer wider advertisement of their prices than bilateral direct quoting provides. Keep in mind that a large bank inputting the best bid, e.g., still buys at the bid side if a second bank hits that bid (it is the second bank that sells at the bid). Pretrade anonymity may also be valuable.
182
Charles Goodhart, Takatoshi Ito, and Richard Payne
five and ten pips indicates that considerable automaticity is indeed built in (fig. 4.12).
Their Results The authors present a fearsome array of results. In my judgment, this is appropriate given that their paper is the first study of its kind. This requires, however, that readers draw their own conclusions about what is most important to take away. The following is my take on it. First, the paper brings to sharp relief the fact that there is no monolithic entity called the spot market. Even within mark/dollar, there are different ways to trade and different classes of participants. Consequently, there are many different sources of data. No one source provides a complete picture. The idea that D2000-2 is the market and therefore the ultimate benchmark is overwrought (and the authors are duly cautious here). That said, these are transactions data, and in that sense they represent market activity in a way indicative FXFX data cannot. With the authors' caution in mind, the three central take-ways appear to be the following. First, FXFX provides an excellent image of the level of market price as it evolves over time. Second, FXFX is a poor indicator of how market spreads vary over time. Third, FXFX provides little information regarding trading volume (whether through entry frequency or otherwise). Another result that I find interesting is their finding that the negative autocorrelation in returns disappears when transaction prices are used. The negative autocorrelation in FXFX quotes is well documented and piqued enough interest that people had begun theorizing as to why it occurs. The fact that transaction prices do not exhibit this will surely affect how we think about it. Of course, as the authors point out, the reason that the quotes are autocorrelated while the transaction prices are not is an interesting topic in itself. Two ways in which the paper might be clearer are the following. First, the text bounces a bit too much from comparative mode (02000-2 vs. FXFX) to focus mode (the properties of 02000-2 data per se). Although section 4.3 would appear to contain the comparative analysis, in fact the authors frequently compare the series elsewhere. This makes it difficult at times to know when the text is referring only to 02000-2. Second, in various places the text discusses "negative moving average" and "negative autocorrelation" without intending any distinction (as far as I can tell). Further, the term reversal is now commonly used in this literature to describe negative autocorrelation and would help readers less familiar with time-series work on returns.
References Bank for International Settlements (BIS). 1993. Survey offoreign exchange market activity. Basle: BIS, Monetary and Economic Department. Lyons, R. K. 1995. Tests of microstructural hypotheses in the foreign exchange market. Journal of Financial Economics 39:321-51.
5
Foreign Exchange Volume: Sound and Fury Signifying Nothing? Richard K. Lyons
Volume in the spot foreign exchange market dwarfs that in any other financial market. But is all this trading informative? This paper provides some empirical evidence. At the broadest level, my results help clarify why trading volume in this market is extraordinarily high. At a narrower level, I provide some sharp results regarding the relation between the intensity of trading and the informativeness of trades. Specifically, I provide results that discriminate between polar views of trading intensity, to which I refer as (1) the event-uncertainty view and (2) the hot potato view. The event-uncertainty view holds that trades are more informative when trading intensity is high; the hot potato view holds that trades are more informative when trading intensity is low. In general, theory admits both possibilities, depending primarily on the posited information structure. To understand the event-uncertainty view-that trades are more informative when trading intensity is high-consider the work of Easley and O'Hara (1992). In contrast to earlier models where new information is known to exist, in Easley and O'Hara (1992) new information may not exist. That is, there is some probability, say p, of new information and probability (1 - p) of no new information. In the event of new information, there is some probability, say q, that an informed trader has received good news and probability (1 - q) of having received bad news. They demonstrate that, if there is no trade at time t, then a rational dealer raises the probability that she attaches to the noRichard K. Lyons is associate professor in the Haas School of Business at the University of California, Berkeley, and a faculty research fellow of the National Bureau of Economic Research. The author thanks the following for helpful comments: George Constantinides, Mark Flood, Jeffrey Frankel, Antonio Mello, Carol Osler, and seminar participants at Berkeley, LSE, North.. western, NYU, UBC, MIT, the IMF, and the NBER. He also thanks Jeff Bohn for valuable research assistance and Merrill Lynch and Lasser Marshall for access to dealers and brokers while trading. Financial assistance from the National Science Foundation and the Berkeley Program in Finance is gratefully acknowledged. Any errors are his own.
183
184
Richard K. Lyons
information event and lowers the probability of news having occurred. Put differently, if trading intensity is low, an incoming trade of a given size induces a smaller update in beliefs since it is less likely to be signaling news. On the flip side, trades occurring when intensity is high should induce a larger update in beliefs. To understand the term the hot potato view-that trades are more informative when trading intensity is low-consider the ideas of Admati and Pfleiderer (1988). Key to their model is the presence of discretionary liquidity traders: in order to minimize their losses to informed traders, rational liquidity traders clump together in their trading. (The reason that informed traders cannot fully offset this advantage to clumping is that information is short-lived.) Owing to this clumping of liquidity traders, trades occurring when intensity is high tend to be less informative. The metaphor of the hot potato offers a link between this discretionary liquidity trading and foreign exchange trading. Foreign exchange dealers use the metaphor in referring to the repeated passage of idiosyncratic inventory imbalances from dealer to dealer following an innovation in customer order flow. These interdealer liquidity trades are clearly discretionary as to timinghence the connection between discretionary liquidity trading and the hot potato view of order-flow information. To clarify the hot potato process, consider the following crude but not unrealistic example. (Keep in mind that roughly 85 percent of foreign exchange trading is interdealer, a much higher share than in other multiple-dealer markets.) Suppose that there are ten dealers, all of wholJl are risk averse, and each currently with a zero net position. A customer sale of $10 million worth of deutsche marks is accommodated by one of the dealers. Not wanting to carry the open position, the dealer calculates his share of this inventory imbalance-or one-tenth of $10 million-calls another dealer, and unloads $9 million worth of deutsche marks. The dealer receiving this trade then calculates his share of this inventory imbalance-or one tenth of $9 million-calls another dealer, and unloads $8.1 million worth of deutsche marks. The hot potato process continues. In the limit, the total interdealer volume generated from the $10 million customer trade is $9 million /(1 - 0.9) == $90 million. Thus, the example produces an interdealer share of 90 percent, roughly matching the empirical share. Here are two possible reactions to the example given above, neither of which vitiates its message. (a) Shouldn't the multiplier be infinite since risk-averse dealers would not choose to retain any of the imbalance? The answer is that, in equilibrium, price will adjust to induce dealers to hold some of the perceived excess supply. The 10 percent rule of the example is a crude approximation of a much richer short-run clearing mechanism. l (b) Interdealer trades can reduce idiosyncratic inventory imbalances-which reduces idiosyncratic risk rather 1. For an optimizing model in which hot potato trading arises between dealers, see Lyons (1995a). Flood (1992) examines simulation experiments that allow for hot potato trading.
185
Foreign Exchange Volume
than simply bouncing it-and this will mute the multiplier. This is true, particularly if the trades are brokered. It is therefore more reasonable to think about the example in terms of net customer orders rather than gross. The role of time in the empirical microstructure literature has only recently emerged. Two important contributions are Hasbrouck (1991) and Hausman, Lo, and MacKinlay (1992). Hasbrouck decomposes the variance of stock price changes into trade-correlated and trade-uncorrelated components and finds that trades are more informative at the beginning of the trading day. Also working with stocks, Hausman et al. test for exogeneity of the length of time between transactions, which they reject at conventional significance levels. However, they argue that their estimates do not change when endogeneity is addressed using instrumental variables. On the basis of this, they forge ahead with the assumption of exogenous intertransaction times. Empirical microstructure work in foreign exchange has been constrained until recently by a lack of transaction-level data. The paper most closely related to the analysis here is Lyons (1995b), which uses a transactions data set that is a subset of the data used here (namely, it uses dealt quotes only). That paper presents evidence supporting both of the two branches of microstructure theory: the asymmetric-information branch and the inventory-control branch. Although many papers have provided evidence supporting the asymmetricinformation branch, little or no direct evidence had previously been found in support of the inventory-control branch (see, e.g., Madhavan and Smidt 1991; Manaster and Mann 1993; and the overview in O'Hara 1995). The fact that they are both present provides further impetus for the application of microstructural models to the foreign exchange market. The application here extends previous work by addressing the informational subtleties of order flow. The chapter is organized as follows: section 5.1 presents a model of transaction prices that includes a relation between trading intensity and the information content of trades; section 5.2 describes the data; section 5.3 presents the results; and section 5.4 concludes.
5.1
A Model in Which Time Matters
The following model extends the model of Madhavan and Smidt (1991) by incorporating a role for intertransaction time. As they do, I will exploit the model's ability to disentangle the information effects of trades from the inventory-control effects. The result is a richer characterization of the effect of trades on price. There are two assets in a pure exchange economy: one riskless (the numeraire) and one with a stochastic liquidation value (representing foreign exchange). The foreign exchange market is organized as a decentralized dealership market with n dealers. Here, we focus on the pricing behavior of a representative dealer, denoted dealer i. A period is defined by a transaction effected against dealer i's quote, with periods running from t == 1, 2, ... , T.
186
Richard K. Lyons
-
Signal St _ Signal Cjt
Receive Trade Q jt
Quote Pit
Observe Increment rt
------t----t--+----+-~I Fig. 5.1
Sequencing in each period
Note: St is a public signal of the full information value Vt; Cjt is dealer j's private signal of Vt, where j denotes the dealer requesting the quote from dealer i; Pit is dealer i's bilateral quote to dealer j, a schedule matching each transaction quantity with a price; Qjt is the signed quantity traded, positive for dealer j's purchases, negative for sales; and rt is the period t increment to Vt "
Let dealer j denote the dealer requesting dealer i's quote in any period. Figure 5.1 summarizes the timing in each period. 5.1.1
The Information Environment
The full information price of foreign exchange at time T is denoted by V, which is composed of a series of increments-for example, interest differentials-so that V == "Li=o where rois a known constant. The increments are i.i.d. mean zero. Each increment rt is realized immediately after trading in period t. Realizations of the increments can be thought to represent the flow of public information over time. The value of foreign exchange at t is thus defined as Vt == "L~=o rio At the time of quoting and trading in period t, that is, before rt is realized, Vt is a random variable. In a market without private information or transaction costs, the quoted price of foreign exchange at time t, denoted P t , would be equal to Vt - l , which is the expected value of the asset price conditional on public information available at t. The following two signals define each period's information environment prior to dealer i's quote to dealer j:
'i'
(1) (2)
St
== Vt + ilt'
Cjt ==
Vt
+ Wjt ,
where the noise terms TIt and W jt are normally distributed about zero, are independent of one another and across periods, and have variances (J~ and (J~, respectively. At the outset of each period t, all dealers receive a public signal St of the full-information value Vt • Also at the outset of each period t, dealer j the dealer requesting a quote-receives a private signal Cjt of Vt • In the foreign exchange market, one potential source· of private signals at the dealer level is order flow from nondealer customers; because each dealer has sole knowledge of his own-customer order flow, to the extent that this flow conveys information it is private information, which can be exploited in interdealer trading (see, e.g., Goodhart 1988, 456; and Lyons 1995a).
187
Foreign Exchange Volume
Dealer i conditions on St and then quotes his schedule as a function of possible Qjt. The schedule's sensitivity to Q jt ensures that any realization of Q jt will be regret free for the quoting dealer, in the sense of Glosten and Milgrom (1985). That is, the quote takes account of the adverse selection arising from C·t • Of course, the realization of Q't still prodealer j's additional information _ vides dealer i a signal of C jt • As is standard, the signed quantity that dealer j chooses to trade is linearly related to the deviation between dealer j's expectation and the transaction price, plus a quantity representing liquidity demand Xjt that is uncorrelated with Vt : }
}
(3)
where J.Ljt is the expectation of Vt conditional on information available to dealer j at t, and the value of Xjt is known only to dealer j. (The demand function that supports eq. [3] requires either exponential utility defined over a single period or mean-variance optimization over multiple periods.) I introduce a role for time in the model via equation (3) and the liquidity demand Xjt . The hot potato hypothesis of order-flow information associates liquidity demand Xjt with inventory-adjustment trading. In foreign exchangeaccording to the hypothesis-innovations in nondealer order flow spark repeated interdealer trading of idiosyncratic inventory imbalances. This rapid passing of the hot potato generates a relatively large role for liquidity trades in periods of short intertransaction times. The event-uncertainty hypothesis, in contrast, associates short intertransaction times with a relatively large role for informative trading: in the presence of event uncertainty, intense trading is a signal that an information event has occurred. To summarize, for given precisions of the signals C jt and St' we can characterize these views as follows: Hot potato hypothesis: 2
CTXj
{high when. intertransac~ion ~imes are short; low when IntertransactIon tImes are long.
Event-uncertainty hypothesis: 2
CT Xj
{1~W when i~tertransacti?n ti~es are short; hIgh when IntertransactIon tImes are long.
This change in the relative intensity of liquidity trading will alter the signal extraction problem faced by the quoting dealer, to which we now turn. 5.1.2
The Formation of Expectations
Dealer i's quotes depend on his conditional expectation of Vt at the time of quoting, which- I denote J.L it • This expectation, in turn, is a function of the variables described above: St and Qjt; the third variable described above, Cjt , is communicated (noisily) to dealer i via Qjt.
188
Richard K. Lyons
I now address the determination of this expectation J.Lit. Dealer i's prior belief regarding Vt is summarized by the public signal St. Dealer i then considers the "what if" of various possible Qj/s. In particular, from any Qjt dealer i can form the statistic Zjt (see the appendix):
(4)
Z. Jt
== Q;/8 + Pit - 'ASt == V + w. + [1/8(1 1 - 'A
t
Jt
'A)]X. jt'
where 'A == (J'~/( (J'~ + (J'~). This statistic is normally distributed, with mean Vt and variance equal to the variance of the last two terms, both of which are orthogonal to Vt . Via Xjt , the variance of the second of these two terms is a function of intertransaction times, per above. Let (J'~s denote the variance of the statistic Zjt when intertransaction times are short, and let (J'~[ denote the variance of ~t when intertransaction times are long. Since Zjt is statistically independent of St' dealer i's posterior J.L it , expressed as a function of any Qjt' takes the form of a weighted average of St and Zjt: (5)
where K s == (J'~s/( (J'~s + (J'~), and K[ == (J'~/( (J'~[ + (J'~). This expectation plays a central role in determining dealer i's quote. Note that K s > K[ if (J'~s > (J'~[' that is, if liquidity trading is relatively important when intertransaction times are short. 5.1.3
The Determination of Bid/Offer Quotes
Consider the following prototypical inventory-control model. Here, the transaction price is linearly related to the dealer's current inventory-a specification that is optimal in a number of inventory control models: (6)
where J.L it is the expectation of Vt conditional on information available to dealer i at t, I;t is dealer i's current inventory position, and I; is i's desired position. The inventory-control effect, governed by ex, will in general be a function of relative interest rates, firm capital, and carrying costs. The variable D t is a direction-indicator variable with a value of 1 when a buyer-initiated trade occurs and a value of -1 when a seller-initiated trade occurs. Thus, the term 'YDt picks up (half) the baseline spread: if dealer j is a buyer, then the realized transaction price Pit will be on the offer side and therefore a little higher, ceteris paribus. (To be precise, 'YDt picks up half the spread for trade quantities approaching zero, i.e., for which there is no adverse selection effect on J.Lit") This term can be interpreted as compensation resulting from execution costs, price discreteness, or rents. Consistent with the regret-free property of quotes, I substitute dealer i's expectation conditional on possible Qj/s-equation (5)-into equation (6), yielding:
189
(7)
Foreign Exchange Volume
Pit
KkSt + (1 - Kk)Zjt - a«(t - I;)
=
+ 'YDt'
k
= s,
1,
which is equivalent to (see the appendix) (8)
_ + (1~ -
k) Qjt -
Pit - St
(a) *
'Y ) Dt , + (
where k == (K k - A)/(1 - A) and 0 < k < 1 since 0 < Kk < 1, 0 < A < 1, and K k > A. 5.1.4
An Estimable Equation
Equation (8) is not directly estimable because St is not observable to the econometrician. My assumptions about the signals available and the evolution of Vt allow me to express the period t prior St as equal to the period t - 1 posterior from equation (6) lagged one period, plus an expectational error term Bit:
Substituting this expression for St into equation (8) yields Pit
= [Pit- + a(lir-l 1
1 - k] +[~
- I;) - ,,/Dt- 1
* Qj, - [a]
+
Sit]
+ [
D"
which implies:
This corresponds to a reduced-form estimating equation of (11)
!1Pit = (30 + (3IQjt + (32(t + (33(t-1 + (34 Dt + (35 Dt-1 + Bit·
Thus, the change in the transaction price from t - 1 to t is linearly related to (i) the signed incoming order at t, (ii) the inventory level at t, (iii) the inventory level at t - 1, (iv) whether Pit is at the bid or the offer, and (v) whether Pit-I is at the bid or the offer. Note that the last two regressors-the indicator variables D t and Dt_I-are accounting for bid-offer bounce. The model predicts that {r31' (33' (34} > 0, {(32' (35} < 0, 1(321 > (33' and (34 > 1(351, irrespective of the intertransaction time. (The latter inequalities derive from the fact that 0 < k < 1.) These more general predictions are borne out in the data and are presented in Lyons (1995b). Here, the focus is on the information in order flow measured by (31' which in tum is a function of the structural parameter K from equation (5). That is, I want to test whether the coefficient (31 is sensitive to
190
Richard K. Lyons
intertransaction time and, if so, in which direction. The hot potato hypothesis predicts a lower J31 when intertransaction times are short; the event-uncertainty hypothesis predicts a higher J31 when intertransaction times are short. These predictions derive from the relative importance of liquidity trading (ui) in the signal extraction problem. My final comment on the model concerns the assumption of a time-invariant desired inventory. First, note that with a slight reinterpretation the model can accommodate variability in desi~ed inventories, that is, an I; that varies through time. Consider the model I: = I; + O(f-1it - St)' which is consistent with the linear demands arising from negative exponential utility, where the public information St represents the market price away from dealer i. Further, Qjt is the only information available to dealer i that is not reflected in St. Under the assumptions of my model, (f-1it - St) is proportional to Qjt. Accordingly, I ~rite (f-1it - St) = 7rQjt· I!ence, I can express the desired inventory as I: = Ii + 07rQjt. In estimation, Ii is absorbed in the constant. The estimate of J31 now represents
whose significance still evinces an information effect, although I have to be more careful in interpreting its magnitude.
5.2 Data My data set has significant advantages over foreign exchange data used in the past, in particular, Reuters FXFX indications data (see, e.g., Goodhart 1989; and Bollerslev and Domowitz 1993). The main shortcomings of the Reuters indications are three: first, these are only indications, not firm quotes at which dealers can transact; second, there is no measure of order flow or transaction prices; and, third, the spreads in the indications data set are two to three times the size of firm quotes in the interdealer market. My data set consists of two linked components, covering the five trading days of the week 3-7 August 1992, from the informal start of trading at 8:30 EST to roughly 1:30 EST. The first component includes the time-stamped quotes, prices, and quantities for all the direct interdealer transactions of a single deutsche mark/dollar dealer at a major New York bank. The second component comprises the same dealer's position cards, which includes all indirect (brokered) trades. 5.2.1
Dealer Data: Direct Quotes and Trades
The first component of the data set includes the dealer's quotes, prices, and quantities for all direct transactions. The availability of this component is due
191
Foreign Exchange Volume
to a recent change in technology in this market: the Reuters Dealing 2000-1 system. This system-very different from the system that produces the Reuters indications-allows dealers to communicate quotes and trades bilaterally via computer rather than verbally over the telephone. 2 Among other things, this allows dealers to request up to four quotes simultaneously, whereas phone requests are necessarily sequential. Another advantage is that the computerized documentation reduces the paperwork required of the dealers. Although use of this technology differs by dealer and is currently diffusing more widely, this dealer uses Dealing 2000-1 for nearly all his direct interbank trades: less than 0.4 percent of all transactions were conducted over the phone over my sample week (as indicated on the position cards). Each record of the data covering the dealer's direct trading includes the first five of the following seven variables; the last two are included only if a trade takes place: 1. 2. 3. 4. 5. 6. 7.
The time the communication is initiated (to the minute, with no lag). Which of the two dealers is requesting the quote. The quote quantity. The bid quote. The offer quote. The quantity traded (which provides Qjt)' The transaction price (which provides Pit)'
This component of the data set includes 952 transactions amounting to $4.1 billion. Figure 5.2 provides an example of a dealer communication as recorded by the Dealing 2000-1 printout (for more details, see Reuters [1990]). The first word indicates that the call came "From" another dealer. Then comes the institution code and name of the counterparty, followed by the time (Greenwich Mean, computer assigned), the'date (day first), and the number assigned to the communication. On line 3, "SP DMK 10" identifies this as a request for a spot deutsche mark/dollar quote for up to $10 million. Line 4 provides the quoted bid and offer price: typically, dealers quote only the last two digits of each price, the rest being superfluous in such a fast-moving market. These two quotes correspond to a bid of 1.5888 deutsche marks/dollars and an offer of 1.5891 deutsche marks/dollars. In confirming the transaction, the communication record provides the first three digits. Here, the calling dealer buys $10 million at the deutsche mark offer price of 1.5891. The record confirms the exact price and quantity. In the current data set, transactions never take place within the spread; the transaction price always equals either the bid or the offer. 2. Dealing 2000-1 is also very different than Dealing 2000-2. The former is wholly bilateral, while the latter is akin to an electronic broker, where multiple dealers participate. See Goodhart, Ito, and Payne, chap. 4 in this volume.
192
Richard K. Lyons
From CODE FULL NAME HERE * 1250GMT 030892 */1080 Our Terminal: CODE Our user : DMK SP DMK 10 # 8891 BUY # # # #
#
10 MIO AGREED VAL 6AUG92 MY DMK TO FULL NAME HERE TO CONFIRM AT 1.5891 I SELL 10 MIO USD TO CONFIRM AT 1.5891 I SELL 10 MIO USD VAL 6AUG92 MY USD TO FULL NAME HERE AC 0-00-00000 THKS N BIFN
# #
# A
A
#END LOCAL# ## WRAP UP BY DMK DMK 1250GMT 3AUG92 #END#
( 265 CHARS)
Fig.5.2 Example of a Reuters Dealing 2000-1 communication Note: "From" establishes this as an incoming call; the caller's four-digit code and institution name follow; "GMT" denotes Greenwich Mean Time; the date follows, with the day listed first; "SP DMK 10" identifies this as request for a spot, deutsche mark/dollar quote for up to $10 million; "8891" denotes a bid of 88 and an offer of 91 (only the last two digits are quoted); the confirmation provides the complete transaction price and verifies the transaction quantity; "THKS N BIFN" is shorthand for "thanks and bye for now."
5.2.2
Dealer Data: Position Cards
The second component of the data set is composed of the dealer's position cards over the same five days covered by the direct-transaction data, 3-7 August 1992. In order to track their positions~ spot dealers record all transactions on handwritten position cards as they go along. An average day consists of approximately twenty cards, each with about fifteen transaction entries. There are two key benefits to this component of the data set. First, it provides a very clean measure of the dealer's inventory It at any time since it includes both direct trades and any brokered trades. Second, it provides a means of error checking the first component of the data set. Each card includes the following information for every trade: 1. The signed quantity traded (which determines It)' 2. The transaction price. 3. The counterparty, including whether brokered. Note that the bid/offer quotes at the time of the transaction are not included, so this component of the data set alone is not sufficient for estimating the model. Note also that each entry is not time-stamped; at the outset of every card, and often within the card too, the dealer records the time to the minute. Hence, the exact timing of some of the brokered transactions is not pinned
193
Foreign Exchange Volume
down since these trades do not get confirmed via a Dealing 2000-1 record. Nevertheless, this is not a drawback for my purposes: the observations for the empirical model are the direct transactions initiated at the dealer's quoted prices; since the timing of these is pinned down by the Dealing 2000-1 records, and since these transactions appear sequentially in both components, the intervening changes in inventory due to brokered trades can be determined exactly. 5.2.3
Descriptive Statistics
Table 5.1 presents the data in the form of daily averages to convey a sense of the typical day's activity. This is masking some daily variation in the sample: the heaviest day (Friday, 7 August) is a little less than twice as active as the lightest day (Wednesday, 5 August). Note that this dealer averages well over $1
Table 5.1
Overview Statistics, 3-7 August 1992 Direct 1. Average number of transactions daily: a) Incoming b) Outgoing 2. Average value of transactions daily:a a) Incoming b) Outgoing 3. Median transaction size: b a) Incoming b) Outgoing 4. Average number of quotes daily: a) Made b) Received 5. Median quoted spread: dealt: a) Made b) Received 6. Median quoted spread: not dealt a) Made b) Received C
C
Brokered 77
190 170 20
.8
.4
.65 .15
3 3 5
4
924 502 422 .0003 .0003 .0003 .0003 .0003 .0005
Note: Data for the dealer's direct (interdealer) quotes and transactions are from the Reuters Dealing 2000-1 communications. Incoming refers to transactions initiated by another dealer; outgoing refers to transactions initiated by my dealer. Made refers to quotes made by my dealer; received refers to quotes received by my dealer. The trades in these two columns reflect more than 95 percent of this dealer's trading; the trades that make up the remainder are executed either (i) over the phone, (ii) with a nondealer customer, or (iii) in the futures market (lMM). Data for the dealer's brokered transactions are from the dealer's position sheets; it is not possible to identify the aggressor from these data. The dealer's trading day begins at 8:30 A.M. EST and ends around 1:30 P.M. on average. a$Billion. b$Million.
cDM.
194
Richard K. Lyons
billion of interdealer trading daily (brokered trades are necessarily interdealer). With respect to quoting, because this dealer is among the larger in this market, he has $10 million "relationships" with many other dealers; that is, quote requests from other high-volume dealers that do not specify a quantity are understood to be good for up to $10 million. Note the tightness of the median spread. For comparison, the median spread in the Reuters indications data set is two to three times as large. A bid/offer spread of three pips is less than 0.02 percent of the spot price. A natural concern is whether this dealer is representative of the larger dealers in the spot market. While I cannot answer this definitively, I offer a few relevant facts. First, he has been trading in this market for many years and is well known among the other major dealers. Second, in terms of trading volume, he is without a doubt one of the key players, trading well over $1 billion per day and maintaining $10 million quote relationships with a number of other dealers. Although this would probably not put him in the top five in terms of volume, he is not far back, possibly in the fifth to fifteenth range somewhere. In the end, my view is that he is representative, at least with respect to the issues addressed here. There is no doubt, however, that different trading styles exist. 5.2.4
Relevant Institutional Background
Here, I highlight two institutional factors relevant to my analysis: (i) trading limits imposed on dealers and (ii) trading on the International Money Market (IMM) futures market. As for trading limits, there is an important distinction between intraday limits and overnight limits. At my dealer's bank, which is typical of major banks, there are no explicit intraday limits on senior dealers, although dealers are expected to communicate particularly large trades to their immediate supervisor (about $50 million and above for many banks in the current deutsche mark/dollar market). In contrast, most banks impose overnight limits on their dealers. Currently, a common overnight limit on a single dealer's open position is about $75 million, considerably larger than the largest open position in my sample. Most dealers, however, close their day with a zero net position; carrying an open position means monitoring it through the evening, an unattractive prospect after a full day of trading. My dealer ended his day with a zero net position each of the five days in the sample. Finally, although broader risk-management programs are in place at the bank for which my dealer trades, it is rare in foreign exchange that a dealer's position is hedged because it aggregates unfavorably with others; when this does occur, it is typically without the participation of the individual dealer. As for trading on the IMM futures market while dealing spot, this differs by dealer. I stress, however, that, unlike equity markets, the spot foreign exchange market is many times larger than the futures market: in 1992 the average daily volume in New York in spot deutsche mark/dollar was roughly $50 billion (New York Federal Reserve Bank [1992], adjusted for double counting); in the
195
Foreign Exchange Volume
same year, the average daily volume on all IMM deutsche mark/dollar contracts was less than $5 billion. As for my dealer, his position cards show that he traded less than $1 million daily in futures over the sample period, which is negligible relative to his daily spot volume. Like other spot dealers, he does listen to an intercom that communicates futures prices. However, this intercom is less important to a spot dealer than the intercoms connected to interdealer brokers in the spot market.
5.3 Estimation Results I begin with results from direct estimation of the model in equation (11), which are presented in table 5.2. Although these estimates do not include any role for intertransaction time, they provide a benchmark for the later results regarding the hot potato and event-uncertainty hypotheses. Note that these estimates are essentially a replication of a result presented in Lyons (1995b). Accordingly, I refer readers to that earlier work for more detailed interpretation. Given these benchmark results, henceforth I present only those coefficients that bear on the information content of order flow-namely, variations of J31 from equation (11). All nonreported coefficients remain significant at at least the 5 percent level, with the predicted signs and relative magnitudes. Presenting the results this way allows me to focus on the informational subtleties outlined in section 5.1. 5.3.1
The Core Model of Trading Intensity
Table 5.3 presents my estimates of the information content of order flow, distinguishing between short and long intertransaction times. This is achieved via the introduction of dummy variables St and It (see the equation heading the table). The dummy St equals 1 if intertransaction time is short, 0 otherwise; the dummy It equals 0 if intertransaction time is short, 1 otherwise. Short interTable 5.2
Estimated Predicted
Benchmark Results (11) Mit = Ilo + III
Qjt
+ 1l2l it + 1l3l it-l +
1l4 D t
+ IlSD
t-
l
+ E it
~o
~l
~2
~3
~4
~5
R2
-1.37 ( -1.07)
1.34 (2.80) >0
-.92 (-3.03) <0
.72 (2.46) >0
10.85 (5.69) >0
-9.14 (-6.04) <0
.22
Note: t-statistics are given in parentheses. ~Pitis the change in the transaction price (DM/$) from t - 1 to t. Qjtis the dollar quantity transacted directly at dealer i's quoted prices, positive for buyerinitiated trades (i.e., effected at the offer) and negative for seller-initiated trades (at the bid). It is i's position at the end of period t. D t is an indicator variable with the value 1 if the trade is buyer initiated and the value - 1 if seller initiated. The units of Qjt' lit' and I it - 1 are such that a coefficient of unity implies a price effect of DM 0.0001 for every $10 million. The units of the indicator variable D t - 1 are such that a coefficient of 10 implies DM 0.0002/$ between bid and offer at quantity zero. Estimated using OLS, with heteroskedasticity- and autocorrelation-consistent (firstorder) standard errors. Sample: 3-7 August 1992, 842 observations.
196
Richard K. Lyons
Table 5.3
Is Order Flow Less Informative When Intertransaction Time Is Short?
~l = ~;, ~l
Intertransaction time short if: Less than 1 minute Less than 2 minutes
(short)
-.01 (- .01) .76 (1.63)
~;
(long)
2.20 (3.84) 2.60 (3.40)
Fraction Short
P-value
262/842
.000
506/842
.009
Note: t-statistics are given in parentheses. The coefficient ~l measures the information effect of trades for which the time from the previous transaction is short (Sf = 1 and If = 0 in the equation in the heading), where short is defined in the first column. The coefficient ~; measures the information effect of those trades for which the time from the previous transaction is long (Sf = 0, If = 1), where long is defined as not short. The "Fraction Short" column presents the fraction of observations satisfying the corresponding definition of short intertransaction times. In each case, the remaining observations fall into the long category. The P-value column presents the significance level at which the null ~l = ~; can just be rejected. ~Pif is the change in the transaction price (DM/$) from t - 1 to t. Qjf is the dollar quantity transacted directly at dealer i's quoted prices, positive for buyer-initiated trades (i.e., effected at the offer) and negative for seller-initiated trades (at the bid). The units of Qjf are such that ~l = 1 implies a price effect of DM 0.0001 for every $10 million. If is i's position at the end of period t. D f is an indicator variable with the value 1 if the trade is buyer initiated and the value -1 if seller initiated. Estimated using OLS, with heteroskedasticity- and autocorrelation-consistent (first-order) standard errors. Sample: 3-7 August 1992, 842 observations.
transaction times are defined two ways: less than one minute from the previous transaction and less than two minutes. The time-stamps on the data are very precise since they are assigned by the computer; however, they do not provide precision beyond the minute. Hence, less than one minute includes trades with the same time-stamp; less than two minutes includes trades with time-stamps differing by one minute or less. These categories bracket the mean intertransaction time of 1.8 minutes. The second category corresponds to a break at the median intertransaction time. The results provide strong support for the hot potato hypothesis over the event-uncertainty hypothesis. The coefficient r3 1-which measures the information effect of incoming trades with short intertransaction times-is insignificant at conventional levels. In contrast, the coefficient r3 ~ -which measures the information effect of incoming trades with long intertransaction times-is significant. Moreover, a test of the restriction that r31 = r3 ~ is rejected at the 1 percent level in both cases. In summary, trades occurring when transaction intensity is high are significantly less informative than trades occurring when transaction intensity is low. This is the main result of the paper. 5.3.2
The Pattern of the Market
There is an additional testable implication of the hot potato hypothesis: it follows directly from the story of bouncing inventories outlined above that
197
Foreign Exchange Volume
these discretionary liquidity trades will tend to be in the same direction (i.e., have the same sign). The obverse is that clumped trading is more likely to be hot potato (liquidity) trading if trades follow in the same direction. The implication for prices is that, even if martingales, they are not necessarily Markov. The test presented in table 5.4 addresses the question, Is clumped order flow less informative when transactions follow the same direction? Again, I introduce dummy variables, in this case St' 0t' and it (see the equation heading the table). The dummy St equals 1 if (i) intertransaction time is short and (ii) the previous incoming trade has the same direction, 0 otherwise; the dummy 0t equals 1 if (i) intertransaction time is short and (ii) the previous incoming trade has the opposite direction, 0 otherwise; the dummy it equals 0 if intertransaction time is short, 1 otherwise. A short intertransaction time is defined as less than the median of two minutes. Once again, the results support the hot potato hypothesis. The coefficient ~l-short intertransaction times and same direction-is insignificant. In contrast, the coefficient ~~ -short intertransaction times and opposite directionis significant. A test of the restriction that ~ 1 == ~ ~ is rejected at the 1 percent level. To summarize, clumped trades occurring in the same direction are significantly less informative than clumped trades occurring in the opposite direction.
Table 5.4
Is Clumped Order Flow Less Informative When Transactions Follow the Same Direction? Mit = flo + fllStQjt + fl~OtQjt + fl~ltQjt + fli"it + fl3 I it-l + fl4 D t
+ flSD t - 1 + ~l (short and same)
-.06 (-.11)
~; (short and opposite)
1.90 (3.01)
~~
E it
(long)
2.64 (3.46)
Fraction Short and Same
Fraction Short and Opposite
~l = ~;,
276/842
230/842
.009
P-value
Note: t-statistics are given in parentheses. The coefficient ~l measures the information effect of trades that (i) have short intertransaction times, defined as less than the median of two minutes, and (ii) have the same direction as the previous trade (St = 1, 0t = 0, and It = in the equation in the heading). The coefficient ~; measures the information effect of trades that (i) have short intertransaction times, defined as less than the median of two minutes, and (ii) have the opposite direction of the previous trade (St = 0, 0t = 1, it = 0). The coefficient ~~ measures the information effect of trades that have long intertransaction times, defined as greater than or equal to the median of two minutes (St = 0, 0t = 0, it = 1). The "Fraction Short and Same" column presents the fraction of observations satisfying the corresponding definition of short and same (similarly for the "Fraction Short and Opposite" column). The remaining 336/842 observations fall into the long category. The P-value column presents the significance level at which the null ~l = ~; can just be rejected. ilPitis the change in the transaction price (DM/$) from t - 1 to t. Qjt is the dollar quantity transacted directly at dealer i's quoted prices, positive for buyer-initiated trades (i.e., effected at the offer) and negative for seller-initiated trades (at the bid). The units of Qjt are such that ~l = 1 implies a price effect of DM 0.0001 for every $10 million. It is i's position at the end of period t. D t is an indicator variable with the value 1 if the trade is buyer initiated and the value - 1 if seller initiated. Estimated using OLS, with heteroskedasticity- and autocorrelation-consistent (firstorder) standard errors. Sample: 3-7 August 1992, 842 observations.
°
198
Richard K. Lyons
5.3.3
Another Measure of Market Pace: Quote Intensity
The results of table 5.4 highlight another important observation: although the hot potato and event-uncertainty hypotheses make opposite predictions regarding the relation between information and trading intensity, they are not necessarily competing hypotheses. That is, both effects could be operative: hot potato trading simply dominates when trading is most intense in this market. To examine whether there is independent support for event uncertainty, I exploit an "instrument" that is arguably more closely related to event uncertainty than inventory control. To understand this instrument, recognize that in Easley and O'Hara (1992) transaction intensity per se is the only dimension of trading intensity available for signaling the underlying state. The problem for our purposes is that transaction intensity is also the linchpin of the hot potato model. My data set, on the other hand, includes a second dimension of trading intensity: quoting intensity. The roughly 4: 1 ratio of not-dealt quotes to dealt quotes in table 5.1 above indicates that transactions alone may not be telling the full story. More important for discriminating event uncertainty from hot potato is the fact that quote requests per se typically signal heightened uncertainty and information gathering, whereas hot potato transactions minimize on quote requests in order to unload inventory rapidly. In short, quoting intensity provides another vehicle for Easley and O'Hara. Table 5.5 presents estimates of the information content of order flow, distinguishing between high and low quoting intensity as a measure of market pace. Once again I introduce dummy variables, in this case ht and it (see the equation heading the table). The dummy ht equals 1 if the total number of intervening quotes per minute is high, 0 otherwise; the dummy it equals 0 if the total number of intervening quotes per minute is high, 1 otherwise. The different definitions of a high number of intervening quotes appear in the far-left-hand column. These quotes are from the Dealing 2000-1 portion of the data set, described in section 5.2.1. These results provide support for the event-uncertainty hypothesis. The coefficient 131 reflecting high quoting intensity is significant, whereas the coefficient 13 ~ reflecting low quoting intensity is insignificant. A test of the restriction that 131 == 13 ~ is rejected at the 5 percent level in all three cases. To summarize, trades occurring when quoting intensity is high are significantly more informative than trades occurring when quoting intensity is low.
5.4
Conclusions
Our results suggest that, in foreign exchange, trading begets trading. The trading begotten is relatively uninformative, arising from repeated passage of inventory imbalances among dealers. Clearly, this could not arise under a specialist microstructure. A broad implication is that a microstructural understand-
199
Foreign Exchange Volume
Table 5.5
Is Order-Flow More Informative When Quoting Intensity Is High? Mit = 130 + f31 h tQjt + f3~ltQjt + f32(t-l + f34 D t + f3 s Dt- 1 + Eif ~l = ~;, ~l
Quoting intensity high if: 2: 3 intervening quotes per minute 2:
4 intervening quotes per minute
2:
5 intervening quotes per minute
(high)
2.16 (3.42) 2.41 (3.56) 2.72 (3.47)
~;
(low)
.87 (1.70) .84 (1.66) .89 (1.79)
Fraction High
P-value
301/842
.046
215/842
.026
144/842
.025
Note: t-statistics are given in parentheses. The coefficient ~l measures the information effect of those trades occurring when quoting intensity is high (h t = 1, It = 0), where high intensity is
defined in the first column by the total number of quotes-both made and received-since the previous incoming transaction. The coefficient ~; measures the information effect of those trades occurring when quoting intensity is low (h t = 0, It = 1), where low intensity is defined as not high. The "Fraction High" column presents the fraction of observations satisfying the corresponding definition of high-intensity quoting. The P-value column presents the significance level at which the null ~l = ~; can just be rejected. IlP it is the change in the transaction price (DM/$) from t 1 to t. Qjt is the dollar quantity transacted directly at dealer i's quoted prices, positive for buyerinitiated trades (i.e., effected at the offer) and negative for seller-initiated trades (at the bid). The units of Qjtare such that ~l = 1 implies a price effect ofDM 0.0001 for every $10 million. It is i's position at the end of period t. D t is an indicator variable with the value 1 if the trade is buyer initiated and the value -1 if seller initiated. Estimated using OLS, with heteroskedasticity- and autocorrelation-consistent (first-order) standard errors. Sample: 3-7 August 1992, 842 observations.
ing of this market requires much richer multiple-dealer theory than now exists (see, e.g., Ho and Stoll 1983). My principal empirical findings are the following: 1. Trades occurring when transaction intensity is high are significantly less informative than trades occurring when transaction intensity is low. 2. Clumped trades occurring in the same direction are significantly less informative than clumped trades occurring in the opposite direction. 3. Trades occurring when quoting intensity is high are significantly more informative than trades occurring when quoting intensity is low. I interpret the first two results as supportive of hot potato trading among dealers in foreign exchange. I interpret the third result as supportive of the Easley and O'Hara event-uncertainty hypothesis, although the vehicle differs from the transaction focus of their paper. Taken together, the results highlight the potential complementarity between these seemingly polar views. There is an important hardship in focusing on a dealership market like foreign exchange that warrants recognition. Empirical work on the specialist structure has the luxury of describing the behavior of a lone dealer. It is much more difficult to argue that by documenting the behavior of a single dealer in the foreign exchange market I have similarly captured the foreign exchange
200
Richard K. Lyons
market. The data required to generate a more complete picture are out of the question given current availability. Nevertheless, the dealer whom I have tracked is without a doubt one of the key players in this market, trading well over $1 billion per day and maintaining $10 million quote relationships with a number of other dealers. Is he representative of dealers in the core of the wholesale spot market? I would argue yes, at least with respect to the issues addressed here. But there is no doubt that different dealers have different trading styles.
Appendix Derivation of the Statistic Zjt in Equation (4) Beginning with equation (3), (3)
Qjt
=
8(I-Ljt - Pit)
=> .Qj/8 + Pit
+ Xjt
I-Ljt + Xj/8
=
=> Qj/8 + Pit = 'ASt + (1 - 'A)Cjt + Xj/8,
where 'A ==
=> Qjtl8 + Pit - 'ASt = (1 - 'A)(Vt + Wjt ) + Xj/8, (4) => Z.Jt
(j~/«(J'~
+
(J'~),
=
Vt
+ Wjt ,
since Cjt
== Qj/8 + Pit - 'ASt = V + W. + [1/8(1 - 'A)]X.. 1 _ 'A t Jt Jt
Derivation of the Price Representation in Equation (8) Beginning with equation (6),
(6) I can write (where I-Lit
Kk
==
(J'~kl[(J'~k
+
+
(j~],
k = s, I):
=
K~t
=
K~t + [~ ~ :k] [Q/fJ + Pit ~
(1 - Kk)Zjt
['A(1 - Kk)]S + [1 - Kk] [Q./8 1 _ 'A t 1 _ 'A jt
=
K
=
[Kk - 'A(11 --
""
~t + (1
KJ t
. [K SInce k
-
-
>tSt]
+ P.] It
Kk)]S + [1 - Kk] [Q./8 + P.] 'A t 1 - 'A jt It
- k)
[Q/8 + Pi,]. k =
'A( 1 - K k )] l-'A
K k] + [1-- -
l-'A
= 1.
s,
1,
201
Foreign Exchange Volume
Note also that 0 < A for both k == sand k == 1. Each of these properties follows from the definitions of K k and A and the fact that a~j == a~ + [8(1 - A)]-2ai. Substituting this expression for f.1 it into equation (6) yields
References Admati, A., and P. Pfleiderer. 1988. A theory of intraday patterns: Volume and price variability. Review of Financial Studies 1:3-40. Bollerslev, T., and I. Domowitz. 1993. Trading patterns and prices in the interbank foreign exchange market. Journal of Finance 48: 1421-44. Easley, D., and M. O'Hara. 1992. Time and the process of security price adjustment. Journal of Finance 47:577-605. Flood, M. 1992. Market structure and inefficiency in the foreign exchange market. Journal of International Money and Finance 13: 131-58. Glosten, L., and P. Milgrom. 1985. Bid, ask, and transaction prices in a specialist market with heterogeneously informed agents. Journal of Financial Economics 14:71-100. Goodhart, C. 1988. The foreign exchange market: A random walk with a dragging anchor. Economica 55:437-60. - - - . 1989. "News" and the foreign exchange market. Paper presented to the Manchester Statistical Society, 17 October. Hasbrouck, J. 1991. The summary informativeness of stock trades: An econometric analysis. Review of Financial Studies 4:571-95. Hausman, J., A. Lo, and C. MacKinlay. 1992. An ordered probit analysis of transaction stock prices. Journal of Financial Economics 31 :319-79. Ho, T., and H. Stoll. 1983. The dynamics of dealer markets under competition. Journal of Finance 38:1053-74. Lyons, R. 1995a. A simultaneous trade model of the foreign exchange hot potato. University of California, Berkeley, Business School. Typescript. - - - . 1995b. Tests of microstructural hypotheses in the foreign exchange market. Journal of Financial Economics 39:321~51. Madhavan, A., and S. Smidt. 1991. A Bayesian model of intraday specialist pricing. Journal of Financial Economics 30:99-134. Manaster, S., and S. Mann. 1993. Life in the pits: Competitive marketmaking and inventory control. University of Utah, May. Mimeo. New York Federal Reserve Bank. 1992. Summary of results of the U.S. foreign exchange market turnover survey conducted in April 1992. New York. Mimeo. O'Hara, M. 1995. Market microstructure theory. Cambridge, Mass.: Blackwell. Reuters. 1990. The Reuter Dealing 2000-1 Service: User guide, version 3. London: Reuters.
202
Richard K. Lyons
Comment
Mark D. Flood
A comment on Richard Lyons's paper must begin with mention of its data. Lyons has assembled a data set with a level of detail that is unusual in microstructural studies and unprecedented in studies of foreign exchange market microstructure. For the first time, we have an essentially complete and sufficiently long (one-week) time series of quoted spreads (both direct and brokered) and transaction prices for a foreign exchange marketmaker. It should be emphasized that even the untransacted prices in this data set are live quotesand not the indicative prices (such as those collected by Charles Goodhart) that heretofore represented the best intradaily data set available to researchers. Moreover, the data set also includes the quantity of all transactions and the marketmaker's inventory position, with everything time-stamped to the minute. We thus have contemporaneous measurement of all main aspects of a marketmaker's behavior and the major inputs to his or her decision-making process. At the risk of sounding ungrateful, let me point out the only two significant shortcomings of the data set. First, there is, as I understand it, no listing of intradaily news announcements to accompany the marketmaker's data. Such data would have been available, for example, from the Reuters financial newswire-indeed, they may still exist in a Reuters archive-and would have allowed analysis of the marketmaker's response to such events. Second, the data are limited to a single marketmaker. I must acknowledge that it is almost inconceivable that anyone could get access to such data for multiple marketmakers simultaneously. Nonetheless, this is a limitation for two reasons. First, as Lyons acknowledges, we cannot be sure that the marketmaker observed here is representative. It is reasonable to suppose that different marketmakers have different strengths, weaknesses, and constraints and that, therefore, they will have different trading strategies. Second, there are interesting characteristics of the microstructure, including especially the alleged hot potato phenomenon, that best reveal themselves in the interaction of marketmakers rather than the isolated behavior of an individual. I turn now to the theory that Lyons uses to motivate and derive the central empirical hypotheses of the paper, the hot potato and event-uncertainty hypotheses. I suggest an avenue for improving the model as a representation of a foreign exchange marketmaker, as distinguished from a stock exchange specialist. Let me emphasize that what follows is intended as a suggestion for future research rather than an indictment of the present paper. Lyons is aware of the issues raised here and addresses most of them in the paper or in the companion piece, Lyons (1995), which is recommended to readers of the present paper. I found that many of my questions about the latter were answered by reference to the former. Mark D. Flood is assistant professor of finance at Concordia University in Montreal.
203
Foreign Exchange Volume
The theoretical model used here is taken essentially unaltered from Madhavan and Smidt (1991). They are modeling a stock exchange specialist facing a "trader," potentially with inside information. We can reason that the Madhavan and Smidt analysis cannot be a fully accurate model of the foreign exchange market. I shall use Lyons's equation (6), which defines the marketmaker's transaction prices, to focus my explanation of why this is so:
or, substituting, Pit == (a:S t
+ a~Ci)/(a: + a~) + a (/it
- I;)
+ "{D it ·
Equation (6) divides the marketmaker's transaction price into three additive factors: (1) there is a baseline estimate, J.1it' of the intrinsic value of the foreign currency, stated as a convex combination of two signals, one public (S) and one private (Cit); (2) there is a technical "inventory-shading" adjustment, a (/it - I;), to this baseline estimate; and (3) there is a second technical adjustment for the bid-ask spread, "{D it . In this model, all informational innovations are impounded in the first term. This arrangement reflects the intellectual lineage of the theory. In traditional microstructural models going back at least as far as Stigler (1964) and his "jobber's tum" or Demsetz (1968), the (monopolistic) marketmaker-by definition one who stands ready to quote prices and transact on demandprovides liquidity services. The marketmaker, typically conceived as a stock exchange specialist, quotes a market-clearing price (or, under uncertainty, her best estimate of the market-clearing price) and is compensated through the bid-ask spread for her service: waiting around with a securities inventory and trading with all comers. Because she quotes a market-clearing price, she accumulates no inventory (on average). In the later "adverse selection" models, the marketmaker must also be compensated for risk bearing since some traders will come to the marketmaker with profitable insider information, a situation that the marketmaker cannot avoid and therefore must insure against via a wider bid-ask spread. The reason that this cannot accurately represent a foreign exchange marketmaker is that foreign exchange marketmakers cannot base their quoted prices on an estimate of the market-clearing price. Foreign exchange marketmakers are surrounded by competing marketmakers, all of whom have the resources to exploit arbitrage opportunities. This produces an imperative of arbitrage avoidance. Marketmakers who would quote off-market prices (i.e., a bid-ask spread that does not overlap with the spreads prevailing elsewhere in the market) are extremely likely to find themselves with a large inventory that could have been had at abetter price. Thus, marketmakers must attempt to keep their quotes consistent with those of all other marketmakers. For this reason, the determination of prices in equation (6) should be dominated by price information. It is instructive to consider the following counterar-
204
Richard K. Lyons
gument: that the model does not specify the exact nature of the signals (St and Cit) and that these therefore need not include nonprice information at all; therefore, the model indeed allows for marketmakers' behavior that is dominated by price information. There are two significant flaws with such an argument. First, even if J.1 it is determined only by price information, the arbitrage avoidance rule will still be violated if the inventory discrepancy ((t - I;) is sufficiently large. The problem is therefore with the functional form of equation (6) rather than simply the interpretation given to St and Cit. Lyons is aware of this concern: Lyons (1995) estimates the coefficient ex and calculates the size of an inventory discrepancy required for the inventory shading adjustment in equation (6) to overwhelm the bid-ask spread: roughly $40 million. In fact, the marketmaker whose behavior is measured here seldom has an inventory in excess of $40 million (see Lyons 1995, fig. 3). This fact reduces the status of my criticism here from an indictment to a quibble. Second, and more fundamentally, behavior that considers only price information under an arbitrage avoidance rule leaves the exchange rate indeterminate: any price consensus will avoid arbitrage. While this would be consistent with the herd behavior (e.g., speculative bubbles) that some researchers believe characterizes certain exchange rate episodes, a very heavy burden of proof must be placed on anyone who would argue that marketmaker behavior ignores nonprice information. Positing that marketmaker pricing is dominated by price information does not imply, of course, that marketmakers ignore or even discount nonprice information. The desired inventory position represents the other main element of the marketmaker's strategy. To the extent that the current market consensus price fails to reflect all the marketmaker's (nonprice) information, this discrepancy should be exploited through speculative position taking. Borrowing a bit of monetary policy jargon, there are two targets (arbitrage avoidance and speculative profits) and two policy instruments (price and inventory). 1* is thus a measure of the extent to which the marketmaker believes that the market price misestimates the value of the foreign currency. Unfortunately, in the Madhavan and Smidt model, 1* is a constant. Although Lyons offers a technique for making 1* depend on information (see his section 1.4), this approach requires that public information (St) be limited to price information. Moreover, this approach is not incorporated elsewhere in the paper. Representing a separate role for nonprice information ultimately requires that one distinguish between price and nonprice signals in the notation of the model. If we achieve this with superscripts, then equation (6) can be rewritten as
There are three differences in this proposed reformulation: (1) the baseline estimate, J.1it' is a function only of price signals, Sf and Cft; (2) inventory shad-
205
Foreign Exchange Volume
ing is allowed, but now as a nonlinear function of both f-Lit and the inventory discrepancy, to incorporate the arbitrage avoidance rule; and (3) desired inventory is a function of the nonprice signals, S7P and C7i, as well as the baseline estimate, f-Lit" This, of course, is a reformulation of a single equation in a larger model. While rederiving the model to address the concerns raised here is a nontrivial assignment, such a derivation would represent an important advance in our theoretical understanding of decentralized, multiple-dealer markets such as the foreign exchange market. References Demsetz, Harold. 1968. The cost of transacting. Quarterly Journal of Economics 82, no. 1 (February): 33-53. Lyons, Richard K. 1995. Tests of microstructural hypotheses in the foreign exchange market. Journal of Financial Economics 39:321-51. Madhavan, Ananth, and Seymour Smidt. 1991. A Bayesian model of intraday specialist pricing. Journal of Financial Economics 30, no. 1:99-134. Stigler, George J. 1964. Public regulation of the securities markets. Journal ofBusiness 37, no. 2 (April): 117-42.
Comment
Antonio Mello
This paper is a case study of the motives for trading foreign exchange currency. The author tests two hypotheses: either trading is generated by inventory reasons, and in that case it does not convey information when time between consecutive trades is short, or, alternatively, trading is generated by the arrival of new information and intense trading means that an information event has occurred. Using direct quotes and trades from a dealer, covering the five trading days of a particular week in the summer of 1992, the author concludes that both motives can explain trading. The strength and originality of the study results from the data set used. It shows how important it is, in doing work on high-frequency data, to use the correct transaction series. Indeed, data can significantly affect the results as well as our understanding of the economic phenomena. In this respect, the analysis of the behavior of a particular dealer is very informative and certainly improves our knowledge. However, having established that directly reported real-time transactions data are best to test a particular hypothesis (against another), one needs not only to spend more time with the same dealer-that is, having not just one week, but several weeks, and especially different event weeks (turbulent vs. calm periods)-but also to collect data from a panel of Antonio Mello is associate professor of finance at the University of Wisconsin-Madison and a research fellow of the Centre for Economic Policy Research.
206
Richard K. Lyons
different dealers, to control for differences in preferences, in size, in capital, and in information. The two views of trading intensity analyzed, the event-uncertainty view and the hot potato view, deserve some comment. First, I find it difficult to justify the hot potato view: either a dealer is at the optimal inventory level, or he is not. If he is, then he must be indifferent, on a risk-adjusted basis, to trade and not to trade, and the quotes from trading with a liquidity trader must reflect the fact that he must be compensated, on average, for deviating from an optimal inventory level. This makes the trade movement in one direction neither necessary nor optimal. So it does not when the dealer's inventory is not at the optimal level. Perhaps what is really happening is that the dealer is frequently trading to rebalance his optimal inventory level. This would be consistent with a model that accounts for changes in the desired level of inventory. If the relative price of currencies changes, then optimal inventory composition should also change, as the opportunity cost of holding different currencies changes. This should happen regardless of the length of period analyzed, although in practice the revisions of the desired positions should occur only at discrete and endogenous intervals. This more general formulation is certainly difficult to test, but it is also a more realistic one. It requires nonlinear estimation methods, and the interpretation of the results is surely more complex. Second, the tests are based on the sensitivity of price changes to the order flow, which can be interpreted as a test of market depth. In that case, as the time interval shortens, on average, one expects the price changes to be smaller, which is exactly the result the author obtains. Indeed, the feeling that I have is that the results seem to be highly dependent on the definition of short time, a metric that must be endogenous and dependent on the prevailing market conditions. Third, to test a particular theory, it is not sufficient to show whether a particular coefficient is significant and has the right sign. It is also necessary to look at other coefficients and to show that the estimated model displays good adherence to the data. Finally, in testing and interpreting the results, it is important to consider the fact that price improvement is discrete. Depending on the relevance of this matter, OLS may not be appropriate. Also, if prices change discretely, it may very well be the case that prices are revised only after IQ, and the interpretation of the results in the tables changes accordingly. Although there are some points that deserve attention and must be tightened, overall I think that this paper is a contribution in the right direction, and therefore it must be welcomed.
II
Speculation, Exchange Rate Crises, and Macroeconomic Fundamentals
6
Dynamic Hedging and the Interest Rate Defense Peter M. Garber and Michael G.Spencer
Hand in glove with the internationalization of portfolios and the interlinking of money markets across currencies has been the expanded use of methods to hedge currency risk. The rapid proliferation of hedging techniques and the reduction in communication and transactions costs have proceeded simultaneously with these trends. While basic hedging instruments such as forward exchange contracts have a long history, the use of newer instruments such as exchange-traded options and futures contracts and over-the-counter (OTe) options and currency swaps has grown dramatically in the past decade. In addition, option-pricing methods have been used in dynamic hedging strategies to construct tailor-made synthetic derivative products-a transplantation to currency markets of the portfolio insurance methods used to hedge equity market exposure. The crash of 1987 led to justifiable skepticism about the ability of mechanistic trading strategies like dynamic hedging actually to deliver the intended hedge protection when markets are illiquid. 1 In addition, these strategies have been criticized for their tendency to exacerbate price trends. Such criticisms carryover to the use of dynamic hedging in currency markets, although currency markets are usually among the most liquid of financial markets. In this paper, we examine the effect of dynamic hedging strategies on forPeter M. Garber is professor of economics at Brown University. Michael G. Spencer is an economist in the Research Department of the International Monetary Fund. The authors thank Philippe Jorion, Paolo Kind, Richard Lyons, John Montgomery, Victor Ng, David Ordoobadi, the editors, and other conference participants for helpful comments and for discussions on the use of dynamic hedging by portfolio managers. The conclusions of this paper are those of the authors and are not necessarily those of the International Monetary Fund. 1. See, e.g., the Brady Commission (1988) and SEC (1988) reports. Grossman (1988) forecast this problem. Gennotte and Leland (1990) model the relation between hedging operations and market liquidity and show how a relatively small volume of transactions initiated by hedgers can lead to a large price change.
209
210
Peter M. Garber and Michael G. Spencer
eign exchange markets during those crisis periods when even the exchange markets can become illiquid. Although we place some emphasis on the wellknown inability of these strategies to perform well for the hedger when a discontinuity in the exchange rate or an upsurge of volatility occurs, we are concerned primarily with the effect of hedging strategies on the efficacy of the classic central bank interest rate defense of a fixed exchange rate. It is typically believed that a central bank can defend an exchange rate if it is willing to raise short-term interest rates sufficiently high to squeeze holders of short positions in its currency. In the presence of dynamic hedging, however, mechanistic selling of the domestic currency may arise, in the end game of the interest rate defense, and this may overwhelm the credit lines available to the central bank for intervention in the exchange market before those squeezed by the interest rate increase start to buy. Thus, our ultimate focus is on market and central bank behavior in the crucial last moments of a fixed exchange rate, the boundary point toward which the collapsing system converges. The essay is organized as follows. In section 6.1, we outline the growth of the derivative markets in currency products and analyze the instruments available for balancing a currency position. We then describe the role of dynamic hedging in the currency markets. In section 6.2, we consider the hedging operations of nonbanks and the techniques in general use. We then analyze schematically the methods used by banks to hedge the currency exposures in their foreign exchange books. In section 6.3, we examine the mechanics of central bank currency intervention and the effect of interest rate defenses on market liquidity, particularly on the response of dynamic hedging programs to interest rate increases. We also consider how the interaction between the timing of different trading programs-dynamic hedging versus closing positions to avoid a squeeze-and the credit lines of the central bank may force the central bank to abandon a fixed exchange rate if it is driven either to the limit of its credit line or to its self-imposed position limit before buyers of the currency arrive.
6.1 6.1.1
The Role of Dynamic Hedging in Foreign Exchange Markets Markets for Foreign Exchange Products
The use of financial derivatives has grown rapidly in recent years. The notional value of outstanding exchange-traded and over-the-counter (OTe) financial derivative contracts-including futures, forwards, forward rate agreements, swaps, options, caps, floors, and collars-has grown from approximately $7.2 trillion at the end of 1989 to $17.6 trillion at the end of 1992. 2 2. These estimates are derived in General Accounting Office (1994). The notional value of a contract is the nominal amount used as a base to calculate a transfer of payments according to a contractual formula. For example, a simple interest rate swap may have a notional principal of $10
211
Dynamic Hedging
By expanding the opportunities for borrowers and lenders to change the risk characteristics-such as maturity or currency denomination-of their portfolios, the growth in derivatives markets has dramatically altered the nature of international finance and the behavior of market participants. The 1992 notional amounts are composed of $4.8 trillion in exchange-traded contracts, $4.7 trillion in swaps, and $8.1 trillion in aTe options and forward contracts. Foreign exchange derivatives are important components of these markets, particularly the aTe markets. While the notional principal of outstanding exchange-traded foreign exchange derivatives at the end of 1992 was only $105 billion, there were $860 billion in currency swaps and $5.5 trillion in foreign exchange forwards and aTe options outstanding. In contrast, the notional principal of outstanding interest rate products was $4.4 trillion in exchange-traded contracts, $3.9 trillion in swaps, $634 billion in aTe options, and $2 trillion in forward rate agreements. Stock index derivatives totaled $245 billion. The aTe markets in derivative products are concentrated in the hands of a small number of large banks and securities firms in the major financial centers. For example, bank holding companies with more than $10 billion in assets hold between 98 and 100 percent of all aTe derivative positions taken by U.S. banks. 3 aTe contracts are often designed specifically for the needs of particular end users and therefore have tailor-made features such as maturity, currency denomination, and notional principal and are frequently combined with other derivatives and sold as a package. Many aTe trades are interdealer trades in which dealers seek to balance their positions. Exchange-traded derivative products-futures and options-are standardized, retail-sized products. Although they are retail in nature, they are generally used by the dealers in aTe markets to balance positions when credit lines with other financial institutions are filled or when wholesale counterparties are hard to find. Because the exchange's clearinghouse is the counterparty to each contract, and because positions are usually well collateralized through margin requirements, evaluation of creditworthiness is less of an issue on organized exchanges than in the aTe market. 4 The most actively traded financial derivatives on organized exchanges are futures on interest rates, primarily U.S.
million. This notional value is not delivered as principal. Rather, the counterparties would deliver or receive the net between the fixed interest rate applied to $10 million and the floating rate amount, so the claims that the counterparties might have on each other are far smaller than the notional value. 3. Estimates reported in Board of Governors of the Federal Reserve System (1993). For discussions of the activities of banks in OTC derivatives markets, see also Bank of England (1993), Bank for International Settlements (1992), Deutsche Bundesbank (1993), Commodity Futures Trading Commission (1993), General Accounting Office (1994), Group of Thirty (1993), and Goldstein and Folkerts-Landau (1993). 4. OTC derivatives dominate exchange-traded products with limited liquidity such as longerdated contracts or options that are not at or near the money.
212
Peter M. Garber and Michael G. Spencer
Treasury bond rates, Eurodollars, French government bonds (OAT), German Bunds, and Japanese government bonds. The expansion in derivatives markets is reflected in the turnover statistics that have also increased substantially in recent years. Annual turnover of exchange-traded derivatives increased from an estimated 420 million contracts in 1989 to 774 million in 1993. 5 The estimated global· net foreign exchange market turnover, comprising transactions in spot and forward foreign exchange and currency swaps and OTC options, was $880 billion per day in April 1992, compared with $620 billion in April 1989 (Bank for International Settlements 1993). Spot trades accounted for 47 percent of reported turnover, while foreign exchange swaps resulted in 39 percent of turnover. Trading in currency futures averaged $9.5 billion per day, and trading in currency options averaged approximately $28 billion, about 82 percent of which was OTC. 6.1.2
Hedging Foreign Exchange Exposure
Open positions denominated in foreign currencies expose market participants to losses from exchange rate changes. Accounting for such risk is vital for portfolio managers with foreign currency exposure, corporates with foreign-currency-denominated assets or liabilities such as receivables or payables, or banks with currency exposure. Institutional investors play an important role in such investment. At the end of 1991, institutional investorsmutual funds, pension funds, and insurance companies-in OECD countries had total assets of approximately $11.7 trillion, compared to the assets of com-:: mercial banks, which totaled $19.6 trillion. The sizes of their exposures in absolute terms and even in relation to their total assets can be quite large. 6 For example, U.S. mutual funds and pension funds held $214 billion in foreign assets or 5 percent of their combined end-1991 assets of $4.1 trillion. In contrast, U.K. mutual funds and pension funds invested $151 billion abroad-23 percent of their total assets. Institutional investors in Germany, Japan, and the Netherlands also invest sizable proportions of their assets abroad. More significant, perhaps, there are few restrictions on the foreign investments of institutional investors in industrial countries, and the trend appears to be toward relaxing those constraints that do exist. Banks, in contrast, often have welldefined position limits-either statutory or self-imposed-on their foreign exchange exposures. The risks from such holdings are hedged or reduced by taking an offsetting position in the foreign currency-for example, a long position is hedged by shorting the currency in some fashion. This may consist of a spot sale with borrowing in the currency to cover settlement, a forward or future sale, or the 5. See Goldstein and Folkerts-Landau (1994). Intertemporal comparisons of trading volume are only suggestive; no adjustment is made for changes in the composition of trading activity across contracts of different sizes. 6. For a discussion of the foreign holdings of institutional investors in industrial countries, see Goldstein et al. (1993).
213
Dynamic Hedging
acquisition of a put option or sale of a call option on the currency. Access to these instruments differs across types of hedgers: exchange-traded futures or options are retail products with little credit risk but with relatively high margin requirements; OTC products provided by banks and nonbank dealers are typically offered in much larger notional values and require a credit line from the bank to the customer along with a bank's assessment of its exposure to a given client. Options generally provide a partial hedge. For example, a portfolio manager may seek to ensure a floor to the domestic currency value of the foreign currency component of its portfolio, but the portfolio remains subject to the risk of currency fluctuations while the portfolio value is above the floor.? Individual firms and portfolio managers ultimately must tum to banks to engage in foreign exchange hedging since banks are the principal dealers in the foreign exchange spot and derivatives markets. By taking the opposite side of a transaction undertaken by a customer, a bank will acquire foreign exchange exposure that it will then attempt to eliminate. For those exposures that do not net out in the course of a day's trading with other customers-for example, currency or value-date mismatches in forward contract long and short positions or different features of options contracts-the bank must actively seek coverage by initiating its own transactions in the same OTC and exchange-traded derivatives markets. 6.1.3
Mechanics of Option Pricing and Dynamic Hedging
Because option-pricing theory is at the heart of dynamic hedging, it is helpful at this point to review the basic option-pricing formula for foreign exchange-the Garman/Kohlhagen formula. 8 Although banks and other wholesale traders may use more sophisticated pricing methods that account for varying interest rates and exchange rate volatility, the Garman/Kohlhagen formula is in general operational use by pension fund and other portfolio managers, and it is pedagogically useful for illustrating the management of risk in a bank's foreign exchange book. 9 7. In addition, portfolios will be subject to basis risk when the security underlying the hedge instrument is not identical to the security whose return is being hedged so that the returns on the two securities are not perfectly correlated. A hedge constructed with a related, but not identical, instrument to the one whose value is being hedged is called a cross-hedge. 8. For the development of this formula, see Garman and Kohlhagen (1983) or Grabbe (1983). For pricing formulas taking account of stochastic volatility, see Chiang and Okunev (1993), Kroner and Sultan (1993), Melino and Turnbull (1990), Naik (1993), and Perraudin and Sorenson (1992). Dumas, Jennergren, and Naslund (1993) derive option-pricing formulas for currencies restricted by target zones as in the European exchange rate mechanism (ERM). However, the majority of options contracts are written for dollar transactions. In April 1992, 82 percent of net foreign exchange turnover involved the dollar on one side of the transaction; only 7 percent of total net turnover involved exchanges of one ERM currency for another (Bank for International Settlements 1993). In the OTt options market, 74 percent of transactions involved the dollar. 9. Most exchange-traded currency options, other than those traded on the Philadelphia Stock Exchange, are options on futures, for which the GarmaniKohlhagen formula for spot exchange options is inapplicable. However, since the OTC segment of the options market accounts for more than 85 percent of activity (see Bank for International Settlements 1993), the formula for options
214
Peter M. Garber and Michael G. Spencer
Suppose that a customer buys a European put option to deliver deutsche mark for dollars after T periods for an exercise price of $X per deutsche mark. The value of the put option, P t , is (1)
Pt
=
-[1 - N(d1)]exp[ -roM 71S
+ [1
- N(d2)]exp[ -r$T]X,
where r DM and r$ are the (constant) risk-free instantaneous deutsche mark and dollar interest rates, S is the current dollar/deutsche mark spot exchange rate, and X is the exercise or strike exchange rate of the option. 10 N(d 1) is the value of the normal distribution function evaluated at the argument d 1 = [In(SIX)
+ (r$ - rOM + (j2/2)T]/(j~,
where (j is the (constant) instantaneous standard deviation or volatility of the exchange rate S. Finally, d2 = d 1 - (Y~. The put pricing formula is determined by finding the short position in deutsche mark loans and the long position in dollar loans such that a portfolio with these positions and also short a put is riskless with respect to small exchange rate movements. Thus, an investor that wants to hedge its exposure to fluctuations in the dollar/deutsche mark exchange rate can either hedge a long deutsche mark position by buying a put option or use equation (1) to determine positions in deutsche mark and dollar loans that mimic the value of a put-that is, to create a synthetic put. The basic security in the first half of the formula is a loan promising to deliver one deutsche mark in Tperiods-this has a deutsche mark present value of exp[ - rDMl1 and a dollar value of exp[ - roM71S. The coefficient - [1 - N(d 1)] indicates that the mimicking portfolio should consist of a short position of a fraction of such a deutsche mark loan-that is, a short deutsche mark position. Similarly, the dollar position is long a fraction [1 N(d2 )] of a loan promising to pay X dollars in T periods with a present dollar value of exp[ -r$71X. However, since d 1 and d2 constantly move with the exchange rate, the interest rate differential, and the standard deviation projected for exchange rate movements, the positions must be adjusted constantlyhence the term dynamic hedging-to maintain the equivalence of the position to a put option.
on spot exchange rates is more relevant to our discussion. In any event, since currency futures contracts are very sensitive to changes in the interest rate differential, the delta of an option on a currency future is more sensitive to interest rate changes than is the delta of an option on the spot exchange rate, which would tend to strengthen our conclusion. 10. This equation is identical in form to the Merton adaptation of the Black-Scholes put formula for a stock that pays a continuous, constant dividend. This formula is constructed on the assumption that the percentage change in the price of the underlying security, in this case the dollar/ deutsche mark exchange rate, follows a Wiener process, that the instantaneous interest rates in both countries and the standard deviation of the percentage exchange rate change are fixed parameters for the life of the option. Such a simple formula does not exist for American put options; these must be evaluated by numerical methods (see Hull 1993).
215
Dynamic Hedging
The foreign exchange exposure of the agent that sells the put is to the possibility of having to buy deutsche mark at the exercise price at date T. Under the assumptions underlying the pricing formula, it is not necessary to hedge the total face value of the contract prior to the exercise date. How much of the face value to hedge, which in tum determines the hedge ratio, is provided by the option's delta, the change in the value of the option with respect to a movement in the exchange rate. From the pricing formula developed above, the delta of a currency put option is -- [1 -- N(d1)]exp( - rOM]). Thus, a rise in the dollar value of the deutsche mark makes it less likely that the option will be exercised and reduces the value of the put. The put delta takes values between - 1, for a deep in-the-money option that would almost certainly be exercised, to 0, for a deep out-of-the-money option that would never be exercised. The negative of delta, therefore, provides a proxy for the probability of exercise. Delta multiplied by the number of units of foreign currency provides an estimate of the expected foreign exchange that is sold short at any point in time to hedge against possible exercise of the option. A writer of a put option may hedge the option dynamically according to the prescriptions of the put pricing formula. First, it must establish the portfolio that mimics the value of the option: for example, by shorting [1 - N(d)l)]exp( - r OM7) deutsche mark spot for dollars and buying [1 - N(d2 )]exp[ - r$l1X in U.S. Treasury bills. As the exchange rate fluctuates, the now-hedged writer of the option must adjust the short deutsche mark and long dollar positions according to the formula to continue to mimic the option. Typically, the adjustments will not be continuous; instead, to avoid transactions costs, adjustments to the mimicking portfolio will be made as part of a regular rebalancing exercise. Among other assumptions, the put pricing formula is based on assuming that exchange rate volatility will remain constant during the life of the contract. Because volatility typically is not constant, the mimicking portfolio will never perfectly track the actual option's value-gains or losses relative to the initial option premium will always occur-and so the portfolio must constantly be adjusted to changes in volatilities as measured, frequently, by implied volatilities in options prices. If volatility jumps above the value implicit in the price of the actual put option, the writer of the put who also engages in dynamic hedging will take a loss, and the buyer of the put will gain. It is well known that strategies to create synthetic options to hedge actual options through the use of dynamic trading, designed to be delta neutral, can be used to take positions on volatility in underlying prices and in interest rates (see, e.g., Cookson 1993). The loss to the writer is immediately apparent if the portfolio is marked to market. A volatility increase will, ceteris paribus, increase the value of the actual option (a liability) and leave unchanged the value of the hedging portfolio (the supposedly balancing asset). Alternatively, if the option value is not marked to market, the loss will be booked through the dynamic adjustment of
216
Peter M. Garber and Michael G. Spencer
deutsche mark and dollar positions until the exercise date. According to the hedging strategy, a rise in the exchange rate will cause the writer of the put to reduce the short deutsche mark position: the writer of the option will buy deutsche mark when the deutsche mark appreciates and sell when it depreciates. This "buy dear-sell cheap" strategy generates a foreseeable loss to the writer of the put, for which it is compensated by the put premium. If volatility jumps, however, the premium will be insufficient to cover the now greaterthan-expected realized loss on these hedging trades.
6.2 6.2.1
Dynamic Hedging by Type of Institution Currency Risk Management by Fund Managers
Managers of pension funds, mutual funds, and bank trust accounts typically manage their currency risk by dynamic hedging operations-including the use of synthetic securities. For fixed-interest holdings of pension funds with obligations denominated in a given currency, the hedge reflects the desire by fund management to place a floor on the long-term value of foreign-currencydenominated holdings. For funds investing in foreign equitie's, the long-term reasons for establishing currency hedges is not as obvious because of the longrun tendency for exchange rates to conform with purchasing power parity. Nevertheless, in the short term-on a quarterly or an annual basis-fund managers' performance, and therefore their compensation, is often compared to a benchmark. Moreover, fund managers seek to protect short-term performance from significant declines to prevent an increase in redemptions. Similarly, for pension funds, underfunding of liabilities may force an injection of securities into the fund that tests the liquidity of the parent entity. For these reasons, fund managers are sensitive in the short term to exchange rate movements and will wish to hedge positions. A hedge can be established by simply shorting the currency through a forward or future sale in a static hedge, by replicating a put option synthetically, by using constant-proportion portfolio insurance, or by acquiring an actual put option, thereby shifting the dynamic hedging operation to the seller of the put. In the simplest hedging operation, fund directors may establish currency risk targets or limits to which management must adhere by following agreed hedging strategies. To place an absolute ceiling on losses from currency fluctuations, fund directors may mandate the acquisition of a put option to cover the entire foreign exchange position of the fund. If they are willing to bear more risk from volatility changes, fund directors may instruct management to replicate a put dynamically.ll This method has 11. Using real put contracts to hedge long positions is not entirely free of volatility risk, of course, since changes in volatility can result in losses when put contracts are rolled over if the maturity of the contracts is shorter than the horizon of the hedging operation.
217
Dynamic Hedging
become typical for fund management. As indicated above, this buy-high, selllow strategy ex post will have been less costly than an actual put if volatility declines and more costly if volatility increases. Finally, many portfolio managers follow a constant-percentage portfolio insurance strategy: this is a buyhigh, sell-low dynamic strategy that does not replicate a put option. 12 Rather, it is driven entirely by price movements. For example, one realization of this strategy may aim at outperforming a 50 percent hedged position and would begin with a 50 percent hedge. A 1 percent move in the exchange rate would trigger an x percent change in the hedge ratio. If the foreign currency appreciated by 10 percent, the hedge ratio would fall to 50 - lOx percent. Currency depreciations would be met with opposite adjustments in the hedge ratio. The strategy tends to work well when exchange rate changes come in trends but fails with a mere jump in volatility. 13 Dynamic strategies are often implemented through cross-hedges-that is, a hedge may be implemented through shorting a currency whose exchange rate is highly correlated with the currency in which the fund holds securities. The purpose is to take advantage of greater liquidity in the exchange market or an interest rate premium in the currency used for the cross-hedge. 6.2.2
Risk Management of Bank Foreign Exchange Books
Because of internal risk-control operations and regulation of foreign exchange risk, banks are active in using dynamic hedging techniques. Typically, they will hedge the net exposure to exchange rate changes acquired through transactions with clients, but they may leverage exchange risk when trading for proprietary accounts. Regulation on banks' net foreign exchange positions varies widely across industrial countries. 14 In some countries, such as the United States, banks' exposures and internal controls are monitored on a regular basis, although there are no specified limits. Elsewhere, as in, for example, Germany, Japan, and the United Kingdom, guidelines or stronger constraints limit open positions to a specified ratio to total capital. Banks' internal risk management controls include the separation of dealing operations-in which buy/sell orders are taken-and back-office activities where contracts are confirmed and settled, the imposition of position limits on the dealing book, and limits on the extension of credit to individual counterparties. The dominance of the major dealing banks in the markets for foreign ex12. This strategy is referred to by Leland, O'Brien and Rubinstein and Associates as a perpetual protection policy. 13. A constant-percentage portfolio insurance strategy has an advantage over an option replication strategy in that at the end of the period a renewal of the hedge does not require a large trading operation. For an option replication strategy, at expiration the portfolio is either 100 percent hedged or completely unhedged. Renewal of the strategy for another period then requires a large jump in the hedge ratio. 14. For a discussion on the regulatory and internal constraints on banks' foreign exchange trading, see Goldstein et al. (1993).
218
Peter M. Garber and Michael G. Spencer
change options complicates the control of foreign exchange risk and makes banks major users of dynamic hedging techniques. In its study of foreign exchange market transactions in April 1992, the Bank for International Settlements (1993) found that only 10 percent of options trading was conducted on organized exchanges. Moreover, two-thirds of the banks' options transactions, measured by notional principal, had other banks or dealers as counterparties and only one-third involved a nonbank counterparty. The high ratio of interdealer to customer business can be attributed in part to the dealers' hedging operations. A bank that writes an option becomes exposed to the possibility that the option will be exercised, and it will have to buy or sell foreign currency (depending on whether it has written a put or a call). The simplest hedge in this case would be to acquire a perfectly offsetting contract. For a bank that maintains a large options book, many of its options contracts will indeed offset each other. Because OTC options are by nature nonstandard, however, a bank must have a method to break down each option in its book into its implied foreign exchange position. It can then determine its global net position in each currency by adding its net position from trading in other foreign exchange products to its net position implied in its options book. The foreign exchange equivalent into which a bank will decompose its options will depend on the currency options pricing formula used by the bank, but it will usually be based on delta hedging methods. The bank calculates the delta for all the contracts it has written or bought and multiplies these by the face values of the contracts. These are then added up for each currency to estimate the expected net foreign currency delivery requirement. For European-style options, in which exercise is possible only at maturity, the hedge portfolio will include futures or forward contracts that offset these amounts, while, for American-style options, the hedge will include cash positions. Because the management of the foreign exchange book is global, the amounts required to hedge the options will be netted against spot and forward net positions. For example, suppose that the global position in the currency option book of a bank making a market in derivatives is short one OTC European deutsche mark put option that allows the holder to sell DM 1.00 for $X at time T and long one European put option to sell DM 1.00 for French francs at P. If the bank uses the Garman/Kohlhagen formula, its deutsche mark position from its options book is Option
DM Position
1. Short 1 put DM/$
[1 - N(d1)]exp[ -rDM I1
2. Long 1 put DM/Fr
-[1 - N(d;)]exp[ -rDMT*]
In these formulas, d 1 and d~ are defined as above with the appropriate volatilities and exercise prices substituted for each option. If the bank is also long
219
Dynamic Hedging
deutsche mark in its forward and spot trading, it can determine its global foreign exchange exposure in deutsche mark by adding these three quantities. The bank can then hedge the foreign exchange risk by taking the opposite position in the forward market. Because the implied delivery dates across its deutsche mark contracts may differ, this still leaves the bank with an interest rate risk that can be hedged through appropriate deutsche mark forwards or swaps.
6.3
Hedging in a Crisis
Dynamic hedging strategies are not an entirely new activity-stop-Ioss trading had always been triggered by price movements beyond a certain threshold. Dynamic hedging simply mechanizes this response. To the extent, however, that the technique has been adopted by large segments of the financial intermediation industry and has been implemented more rapidly than previous techniques, dynamic hedging strategies have added to trading volume and have accentuated price movements by contributing to momentary illiquidity. In this section, we consider how the widespread use of dynamic hedging techniques-notably, those involved in option replication-may interact with central bank exchange rate and liquidity policies to undermine a defense of a fixed exchange rate system. When a fixed exchange rate regime moves toward a crisis, speculation against the currency is generally channeled through forward sales of the currency to the banking system. Some margin is required by counterparty banks, but this can be leveraged up by a factor of ten or more by the speculator. In a crisis, these sales will generally not be matched by other customers' forward purchases of the currency. The central bank defending the currency may intervene with forward purchases, but the extent of such an operation is limited by the unwillingness of a central bank to risk large capital losses on negative net foreign exchange positions and by limits on credit lines to the central bank made available by the major dealing banks. 15 Once the central bank ceases to buy its currency in the forward market, banks must balance their forward purchases with spot sales of the currency (to balance the net currency position) and by currency swaps (to balance maturities). Likewise, during a crisis, the central bank will be the most important buyer on the spot market through its intervention to maintain the fixed exchange rate. At the same time, it provides its currency through the discount window to the banks who need to sell currency in order to match their forward and spot foreign exchange positions as discussed in the previous paragraph. By providing liquidity to banks through this kind of facility, the central bank is effectively financing the attack on its own reserves. To settle its spot transactions, the central bank must deliver its own foreign exchange reserves or draw down lines 15. The ability of the central bank to enter forward contracts with its own nationally chartered banks is circumscribed by credit line limits imposed by banks elsewhere on these banks.
220
Peter M. Garber and Michael G. Spencer
of credit from other central banks or multilateral entities. As its short foreign exchange position mounts during the intervention, the central bank must act by raising the discount rate. This increases the cost to speculators who speculate against the currency by borrowing from the central bank. The central bank also imposes a squeeze on short sellers by channeling available credit away from identified speculators. This final operation is the classic interest rate defense of a· fixed exchange rate. It works though a liquidity effect in the money market-domestic credit grows less rapidly than central bank net reserves decline, thereby producing a decline in the supply of the domestic settlement medium. If large short positions in the currency are due for settlement, holders of short positions may sell foreign exchange to the central bank rather than face the high interest costs of rolling over overnight loans in the weak currency. The costs to holders of short positions are further accentuated if in addition they are caught in a squeeze so that they have to pay more than the discount rate to obtain funds. The market's acquisition of foreign exchange from the central bank does not arise exclusively from forward sales by nonbank speculators. Speculators and hedgers may also buy put options on the weak currency from the banks. Again, in a crisis, the banking system will likely be unable to find nonbank sellers of puts to balance these position. 16 To hedge, the bank that writes the put may create a long position in a synthetic put by selling the weak currency forward, by selling on the futures market, or by selling spot and entering a swap contract. Any of these operations will trigger a spot sale of the weak currency to the central bank as described above. 6.3.1
The Effect of Interest Rate Changes on Dynamic Hedging
Once a central bank raises interest rates in defense of the fixed exchange rate, hedging operations may trigger further sales of the currency rather than the purchases anticipated from the squeeze. This result follows from the relation between interest rate movements and the hedging portfolio of formula (1).17 16. Even if a nonbank seller of puts exists somewhere in the financial system, the selling bank seeking cover may not find the nonbank seller of puts through the banking system. In a crisis, gross volumes of trading surge, thereby causing many banks to reach their credit ceilings with other banks. As the banking system becomes illiquid in this way, transactions that passed through the banking system on a credit basis now must seek a cash market. To hedge, the selling bank will place an order to buy a put onto the organized currency options market, where credit risk is not an issue, and will find the potential seller in this market. As the crisis progresses and interbank credit lines fill, volume will tend to move to the more secure organized exchanges. 17. As Bill Branson has reminded us, differentiating delta with respect to rOM yields an expression that is not easily signed. In numerical evaluations, delta is a downward-sloping convex function of rOM for most parameter values. However, for options with long maturities or very low implied volatilities or that are deep in the money, the relation may tum positive, although relatively flat, after sufficiently large increases in the foreign interest rate. For example, for the parameters used in fig. 6.1 below, if the contract maturity was six months, the slope would tum positive after rOM exceeds 39 percent. For very long options (e.g., one year) with very low volatilities, delta may be everywhere increasing in the foreign interest rate.
221
Dynamic Hedging
Intuitively, the interest rate differential between the two currencies reflects the expected rate of depreciation of the exchange rate plus a risk premium. Unless volatility increases or attitudes toward risk change, a rise in the differential between deutsche mark and dollar interest rates means that the deutsche mark is expected to depreciate more rapidly against the dollar-that is, the delta or hedge ratio increases. 18 With an unchanged current exchange rate, exercise price, and exchange rate volatility, the put option is much more likely to finish in the money when the interest rate jumps upward. That the option is more likely to be exercised means that it provides a higher effective hedge to a portfolio manager covering a deutsche mark portfolio. The manager of the bank's portfolio who uses a synthetic put in a dynamic hedging operation must likewise provide an increased hedge ratio in response to the greater probability that the option will be exercised. This means that he must short sell more deutsche mark so that his synthetic put continues to mimic an actual put. Taken to an extreme, if deutsche mark interest rates rise so high that, according to the underlying theory, it is almost certain that a put option will be exercised, the put then provides the equivalent of a 100 percent hedge ratio. The bank's portfolio manager using a synthetic put must similarly sell sufficient deutsche mark to cover his entire deutsche mark position to provide the same coverage as an actual put. How important will the dynamic hedging response be? Figure 6.1 provides some indication of this effect. This figure plots the put option delta against the foreign interest rate for a one-month, at-the-money put and for an in-themoney put of equal maturity where the spot exchange rate is assumed to be 1 percent below the strike price. Delta is a declining, convex function of the Such anomalies are likely to be unimportant for three reasons. First, Bank for International Settlements (1993) data on the maturity structure of forward contracts show that foreign exchange dealers' positions are strongly weighted toward the near term: 64 percent of contracts have maturities of less than a week, 99 percent are for less than a year, and ERM currencies have relatively short maturities compared to non-ERM currencies. In addition, futures and exchange-traded options transactions tend to be concentrated in contracts with maturities well below one year. Thus, the behavior of long-term options may not be very relevant. Second, central bank liquidity squeezes generally have their greatest effect on short-term interest rates, so increases in long-term rates are unlikely to be large enough to reverse the tendency toward an increase in the hedge ratio. In the United Kingdom, on 16 September 1994 six-month sterling deposit rates rose only twenty basis points, while one-week rates rose by thirty-five hundred basis points; during the ERM crises, six-month interest rate differentials against dollar LIBOR (London interbank offered rate) rates were highest for Sweden, and these reached only 23 percent. Finally, simulations using actual interest rates and historical volatilities for the currencies involved in the ERM crisis failed to yield any cases where the slope reversed over the range of observed interest rates. Since actual volatilities used in pricing options during a crisis can be expected to exceed historical volatilities, it is likely that the slope of the delta:rDM relation was if anything more negative than that implied by either these simulations or fig. 6.1. 18. A central bank squeeze generally operates through overnight interest rates, which are not the interest rates used to value longer-dated options. Nevertheless, in a squeeze, a jump in overnight rates will usually have a strong effect on one-month, three-month, and even one-year interest rates, which are relevant to option pricing. As the maturity of the option lengthens, a given movement in the relevant interest rate will have a stronger effect on the value of option and on the delta.
222
Peter M. Garber and Michael G. Spencer -0.45.---------
----.
-0.50
-0.55
t-the-money option
-0.60
-0.65
-0.70
-0.75
In-the-money option
-0. 80 t--~....,___r--.__,-r--..--.-__,____r-r___...,__~...,...,...__r_-..__~.....-I 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 Foreign i nleresl rale
Fig.6.1
Sensitivity of the put option delta to the foreign interest rate
Note: Assumptions: domestic rate of 3 percent, volatility of 15 percent, one-month option.
foreign interest rate. The response of dynamic hedging programs during the final days of a managed or fixed exchange rate regime can be inferred from such plots. In the days leading up to the collapse of an exchange rate band regime, the gradual depreciation in the spot exchange rate will have a significant effect on the hedge ratio, necessitating a gradual increase in the short foreign currency position. However, in the final hours or minutes of such a regime or of an absolutely fixed exchange rate, the use of large interest rate increases to defend the fixed exchange rate can result in increases in the hedge ratio of a similar magnitude. In the United Kingdom, for example, on 16 September 1992 the Bank of England increased the base lending rate twice, from 10 percent to 12 percent and then again to 15 percent (effective the next day).19 The one-month London interbank offer rate increased from 10.4 percent at the end of the previous day to 28.9 percent at the end of the sixteenth. Such an interest rate increase would result in a decrease in the delta (or an increase in the hedge ratio) of an at-themoney put of over 20 percent, from -0.54 to -0.66-a larger change than would have been obtained from a 1 percent depreciation at the initial interest rate. In the Swedish market, the increase in the marginal lending rate from 75 percent to 500 percent on 16 September led to an increase in the one-month 19. For descriptions of the European currency crisis of 1992-93, see Goldstein et al. (1993) and Group of Ten (1993).
223
Dynamic Hedging
STIBOR (Stockholm interbank offered rate) rate from 25 to 70 percent. An increase in rates of this magnitude would imply a 30 percent increase in the hedge ratio. For countries such as France and Italy where one-month interest rates were not as variable, the largest increases in delta that would be implied by the interest rate data are of an order of magnitude of 10 percent. In an exchange crisis, therefore, a large defensive rise in the interest rate aimed at imposing a squeeze on speculators will instantaneously trigger hedging programs to order sales of the weak currency.20 The experiment conducted using the contracts illustrated in figure 6.1 suggests that the selling triggered by dynamic hedging programs during an interest rate defense can be significant. Since delta approximates the proportion of the portfolio to be hedged by short positions in the foreign currency, these examples suggest that short positions might be increased by 20 percent or more of the portfolio value in response to an aggressive interest rate increase. The existence of a large amount of such programs in the market would undermine the use of an interest rate defense of a weak currency~the moment that a central bank raises interest rates, it might face an avalanche of sales of its currency rather than the purchases of the squeezed shorts that it had anticipated. 21 In effect, the hedging programs make the hedgers insensitive to the added costs of funding their weak currency sales. 20. Who is actually squeezed in such a defense? All borrowers in the weak currency whose debts are due for settlement or rollover soon (after two days) will find that their costs and risks have suddenly jumped as they now have to pay high and volatile yields to the money mal:ket scalpers that are unleashed by the squeeze. This group could conceivably include even those who have constructed synthetic puts if they have established their short currency position by borrowing on overnight rollover credit, as Richard Lyons has pointed out to us in his conference comments. Typically, however, a synthetic option is constructed by establishing a short forward position whose expiration date coincides with· the expiration date of the option. If the existing hedges were constructed well before the interest rate defense was launched and with a relatively long maturity, they would have locked in longer-term finance, and the position would be immune from a short squeeze. 21. Industry sources indicate that, indeed, when there is an increase in the interest rate spread with no movement in the exchange rate, the forward rate discount will trigger a selloff in the currency through dynamic hedging. During the European exchange rate mechanism crisis of September 1992, e.g., industry sources estimate that dynamic hedging sales to adjust positions because of increases in interest rate spreads, exchange rate movements, and increases in volatility accounted for 20-30 percent of the selling in the crisis. It was a major factor in the lira market one week after the first devaluation and also in the Swedish krona market in 1992. Up to 10 percent of the sales were due to increases in interest rate spreads. In the case of the United Kingdom, on 16 September 1992 the dramatic increase in forward discounts triggered sales of pounds. When interest rates rose and nothing happened to the exchange rates, the selling programs were turned on. The lack of movement (appreciation) in the exchange rate meant that the forward rate fell farther below the floor. Thus, the full force of programmed sales triggered by interest rate movements was not offset by exchange rate improvement. Another source of the sales volumes at this moment was the rising perceived volatility resulting from the suddenly larger movement of the forward rate below the floor. The effect of dynamic hedging sales may also have been a source of some of the selling pressure observed on 12 August 1994, when the Italian lira depreciated sharply after the Banca D'Italia raised the discount rate by fifty basis points, although the consensus view is that markets reacted to a belief that the interest rate increases were fiscally unsustainable.
224
Peter M. Garber and Michael G. Spencer
If the central bank has a credit line limit in foreign exchange or a selfimposed net reserve position limit, the upsurge of selling brought about by the interest rate increase might cause a sudden jump to its limit and force it to cease intervention in defense of the exchange rate. Whether this counterintuitive result occurs depends on the weight of these mechanistic traders relative to those caught in the short squeeze. In one scenario, the hedging operation may in any case far exceed the amount of the weak currency demanded by those caught in the squeeze. In this case, the timing of the hedging sales-the prearranged rule for awakening the selling programs-relative to the time at which those caught in the short squeeze appear on the market is immaterial to the survival of the fixed exchange rate. Dominance by the mechanistic hedges will defeat the interest rate defense. In the scenario in which the amounts of these opposite transactions are roughly balanced or even where those caught in the short squeeze dominate, the timing of transactions is key. If the selling programs switch on instantly but the buying operations to cover short positions occur with some lag, the central banks' net short foreign exchange limit may be exceeded prior to the appearance of the buyers of its currency, causing the abandonment of the fixed exchange rate. Buyers might have appeared by the end of the day to offset the sellers, but the initial selling may unnerve the central bank and force devaluation. The devaluation will ratify the actions both of the sellers and of those caught in the squeeze who hesitated. Sellers will have sold prior to the devaluation of the exchange rate, and those caught in the squeeze can buy back into the weak currency at a lower price. If the central bank simultaneously relaxes the high interest rates, overnight borrowing will cease to be a problem for those caught short, and the squeeze will be suspended.
6.4
Conclusion
In their effect on the viability of the interest rate defense of a fixed exchange rate, dynamic hedging programs can be interpreted as a new wrinkle on an old phenomenon. Skeptical participants in the foreign exchange market have sometimes interpreted a defensive increase in the interest rate as the last rearguard action preparatory to the abandonment of a fixed rate. In this light, the suddenly higher interest rate differential signals only the extent of the impending depreciation of the exchange rate and certainly not a drastic and extended tightening of liquidity in the weak currency's money markets. Interpreting the interest rate increase in this way dictates that. a speculative selling program should be begun. Dynamic hedging programs automatically place this interpretation on an interest rate increase; thus, they are a mechanization of the previously informal skepticism that occasionally arose about exchange rate defenses. To the extent that such programs are present in generating large selling volumes, they signal a major shift toward skepticism about the strength of
225
Dynamic Hedging
the central bank's adherence to the policy of defending the exchange rate, thereby undermining the efficacy of a previously useful defensive tool. The scenario that we depict here is a technical story about the character of minute-by-minute trading in the death throes of a fixed exchange rate. A dramatic interest rate increase in a last ditch defense triggers dramatic selling pressure. If this technical feature of the market is important in the last moments of a fixed exchange rate, it is necessary to implement a defense operation that takes it into account. For example, it is often argued that a resolute defense of a fixed exchange rate regime requires that at an early date interest rates be raised gradually, although ultimately to high levels. 22 Such a policy would trigger daily selling of the currency by dynamic hedgers, but not in quantities that would overwhelm the central bank's net reserve limits before the appearance as buyers by the end of the day of those caught short in the currency. Thus, raising rates gradually in an interest rate defense may immunize the central bank against being pushed intraday beyond its position limits.
References Bank of England. 1993. Derivatives: Report of an internal working group. London, April. Bank for International Settlements. 1992. Recent developments in international interbank relations. Basle, October. - - - . 1993. Central bank survey offoreign exchange market activity in April 1992. Basle, March. Board of Governors of the Federal Reserve System. Federal Deposit Insurance Corporation. Office of the Comptroller of the Currency. 1993. Derivative product activities of commercial banks. Joint Study Conducted in Response to Questions Posed by Senator Riegle on Derivative Products. Mimeo. Washington, D.C., 27 January. Brady Commission. 1988. Report of the Presidential Task Force on Market Mechanisms. Washington, D.C.: U.S. Government Printing Office. Chiang, Raymond, and John Okunev. 1993. An alternative formulation on the pricing of foreign currency options. Journal of Futures Markets 13, no. 8:903-7. Commodity Futures Trading Commission. 1993. aTC derivative markets and their regulation. Washington, D.C. Cookson, Richard. 1993. Moving in the right direction. Risk 6, no. 10:22-26. Deutsche Bundesbank. 1993. Off-balance-sheet activities of German banks. Monthly Report of the Deutsche Bundesbank 45, no. 10:45-67. Dumas, Bernard, L. Peter Jennergren, and Berti! Naslund. 1993. Realignment risk and currency option pricing in target zones. Working Paper no. 4458. Cambridge, Mass.: National Bureau of Economic Research. Garman, M., and S. Kohlhagen. 1983. Foreign currency option values. Journal ofInternational Money and Finance 2:231-37. 22. "Early" is relative to the time of outbreak of the next speculative attack. How to recognize when an attack will come in order to implement this early defense is problematic.
226
Peter M. Garber and Michael G. Spencer
General Accounting Office. 1994. Financial derivatives: Actions needed to protect the financial system. Washington, D.C.: U.S. Government Printing Office. Gennotte, Gerard, and Hayne Leland. 1990. Market liquidity, hedging and crashes. American Economic Review 80:999-1021. Goldstein, Morris, and David Folkerts-Landau. 1993. International capital markets, part II: Systematic issues in international finance. Washington, D.C.: International Monetary Fund. - - - . 1994. International capital markets developments, prospects and policy issues. Washington, D.C.: International Monetary Fund. Goldstein, Morris, David Folkerts-Landau, Peter Garber, Liliana Rojas-Suarez, and Michael Spencer. 1993. International capital markets, part I: Exchange rate management and international capital flows. Washington, D.C.: International Monetary Fund. Grabbe, O. 1983. The pricing of call and put options on foreign exchange. Journal of International Money and Finance 2:239-53. Grossman, Sanford J. 1988. An analysis of the implications for stock and futures price volatility of program trading and dynamic hedging strategies. Journal of Business 61, no. 3:275-98. Group of Ten. 1993. International capital movements and foreign exchange markets. Rome. Group of Thirty. 1993. Derivatives: Practices and principles. Washington, D.C. Hull, John C. 1993. Options, futures, and other derivative securities. 2d ed. Englewood Cliffs, N.J.: Prentice-Hall. Kroner, Kenneth, and Jahangir Sultan. 1993. Time-varying distributions and dynamic hedging with foreign currency futures. Journal ofFinancial and Quantitative Analysis 28, no. 4:535-50. Melino, Angelo, and Stuart Turnbull. 1990. Pricing foreign currency options with sto. chastic volatility. Journal of Econometrics 45:239-65. Naik, Vasanttilak. 1993. Option valuation and hedging strategies with jumps in the volatility of asset returns. Journal of Finance 98, no. 5:1969-83. Perraudin, William R., and Bent E. Sorenson. 1992. Foreign exchange option pricing in a continuous time arbitrage pricing model with stochastic volatility and jumps. Birkbeck College, University of London. Mimeo. Securities and Exchange Commission (SEC). 1988. The October 1987 market break. Washington, D.C.: U.S. Government Printing Office.
Comment
Richard K. Lyons
This comment is intended to clarify assumptions embedded in the option hedging analysis of the Garber and Spencer paper. In particular, the authors state (in section 6.3.1) that "a large defensive rise in the interest rate aimed at imposing a squeeze on speculators will instantaneously trigger hedging programs to order sales of the weak currency." Here, I demonstrate that this result is not unambiguous; rather, additional assumptions are required regarding the interest sensitivity of the underlying Richard K. Lyons is associate professor in the Haas School of Business at the University of California, Berkeley, and a faculty research fellow of the National Bureau of Economic Research.
227
Dynamic Hedging
portfolio's value relative to the hedge's value. To do so, I present a transparent example using a money market hedge, which is a special case of their option foreign exchange hedge.! The analysis of Garber and Spencer spans this case since-assuming frictionless markets-a money market hedge is equivalent to a forward hedge, which is in turn equivalent to a put option hedge with a strike price of infinity. Assumptions
1. Two-period investment horizon. 2. Equal borrowing and lending (investment) rates for the same maturity. 3. Foreign exchange hedge ratio is constant. Notation RpM = one-period nominal deutsche mark interest rate. R~M =
two-period nominal deutsche mark interest rate.
The Four Cases Case 1: (i) Investment: R~M (ii) Hedge borrowing: R~M (iii) t RfM => no change in deutsche mark borrowing for hedge => no spot sale of deutsche mark Case 2: (i) Investment: RfM (ii) Hedge borrowing: RfM (iii) t RpM => no change in deutsche mark borrowing for hedge => no spot sale of deutsche mark Case 3: (i) Investment: R~M (ii) Hedge borrowing: RfM (iii) t RpM => 1- in deutsche mark borrowing for hedge => spot purchase of deutsche mark Case 4: (i) Investment: RpM (ii) Hedge borrowing: R~M (iii) t RfM => t in deutsche mark borrowing for hedge => spot sale of deutsche mark
From these four cases it is clear that the Garber-Spencer effect on the spot market holds only in case 4. This may well be the relevant case for most investors, but that is an empirical matter. Note that if investors are using rolling hedges, as many do, then cases 2 and 3 might be more relevant. (Rolling hedges involve the rolling over of the hedge position because the hedging instrument has a shorter maturity than the cash flow being hedged.) 1. A money market hedge involves either borrowing or lending in the foreign currency to set up an offsetting foreign currency cash flow. Importantly, putting on (or. changing) a money market hedge will also involve a transaction in today's spot market.
7
Heterogeneous Behavior in Exchange Rate Crises Fabio C. Bagliano, Andrea Beltratti, and Giuseppe Bertola
The large changes experienced by European exchange rates certainly had large ex post wealth and welfare effects. It is not difficult to define and measure official reserve losses during the buildup of a currency crisis: as a loss of reserves entails an exchange of foreign-currency-denominated assets for domestic-currency-denominated ones, it is equally straightforward, in principle, to evaluate the accounting loss that ensues from eventual devaluation in a central bank's balance sheet that, through intervention, has become heavily exposed to exchange rate depreciation. Of course, exchange rate policy is primarily aimed at longer-run macroeconomic issues, and public-sector losses from devaluation are trivial from that perspective. From a microeconomic point of view, however, gains and losses in the central bank's balance sheets correspond to very real losses and gains in private-sector balance sheets: if the devaluation does occur after the loss of reserves, "speculators" earn capital gains; symmetrically, as long as the devaluation does not occur, high domestic interest rates hurt the private sector if monetary policy is tightened while yielding (accounting) profits for the central bank. At the empirical level, this perspective brings forth many difficult and relatively unexplored issues. In central banks' balance sheets, assets whose counterpart is in a resident agent's balance sheet ("domestic credit") are denominated in domestic currency, while "reserves," whose counterpart appears in foreign balance sheets, are denominated in foreign currency. Private-sector balFabio C. Bagliano is ricercatore at Universita di Torino. Andrea Beltratti is ricercatore at Universita di Torino. Giuseppe Bertola is professore associato at Univesita di Torino, a research fellow of the Centre for Economic Policy Research, and a faculty research fellow of the National Bureau of Economic Research. The authors are grateful to the discussants and to the conference participants for helpful comments. They also thank Chiara Bentivogli and Vincenzo Loi for kind assistance with Bank of Italy and Ufficio Italiano Cambi data and Onorato Castellino for useful comments on an early draft.
229
230
Fabio C. Bagliano, Andrea Beltratti, and Giuseppe Bertola
ance sheets, however, need not feature the same coincidence. Beyond standard balance-of-payments statistics based on an increasingly ill-defined distinction between "resident" and "nonresident" agents in a world of high capital mobility, only scant evidence is available on the currency and counterpart structure of various private agents' balance sheets. At the theoretical level, standard models of exchange rate crises are silent on many of the relevant issues: the counterpart of central bank reserve changes is taken to be portfolio reallocation by a single representative agent and is driven by current and/or expected monetary policy developments (for a recent survey, see Obstfeld [1994]). Introspection and data, however, suggest that different individuals' and institutions' financial positions are differently exposed to the risk of devaluations in reality. The widely cited "convergence play" in the period of relative stability leading up to the ERM (exchange rate mechanism) crises, namely, the fact that portfolio managers would try and take advantage of large interest rate differentials between "weak-" and "strong-" currency-denominated assets (see, e.g., IMF 1993, chap. 3), is itself evidence of heterogeneity in the financial market: whereas the attempt to take advantage of apparent arbitrage opportunities would lead to interest rate convergence in an equilibrium representative-agent model, interest rate differentials persisted in the precrisis ERM, to indicate that other agents ("speculators") were betting on "divergence." In this paper, we study interactions among optimizing agents and a central bank in an environment where a devaluation may occur with exogenously given probability. This has two advantages. On the one hand, exchange rate crises may be viewed as "controlled experiments" for the difficult task of modeling heterogeneous portfolio formation. On the other hand, focusing on devaluations makes it possible to frame the analysis in terms of two periods ("before" and "after" the devaluation) and to disregard ongoing dynamics to the extent possible. Our approach is intermediate between that of standard macroeconomic models, where exchange rate policy is viewed as a game between willful monetary authorities and a single "public" body, and that of more recent "microstructure" contributions concerned with individual traders' minute-byminute problems (see Lyons, chap. 5 in this volume, and the references therein). In our analysis, we do not focus on fundamentals as much as the former literature does: we shall not try to interpret the monetary authorities' behavior, and monetary and exchange rate policies are exogenous to our approach. In contrast to the "microstructure" approach, we explicitly model the behavior of the market's main actors, who provide inputs to the traders and intermediaries on which those complementary contributions focus. To keep our analysis as simple as possible, we model optimization in terms of a mean-variance return objective, abstracting from the portion of wealth allocation that is not directly relevant for our problem. We pay close attention to the structure of each agent's balance sheet, treat it as the outcome of individual maximization problems, and derive the structure of equilibrium returns on
231
Heterogeneous Behavior in Exchange Rate Crises
available assets. These features are essential to any framework intended to investigate devaluation gains and losses: as these depend on the structure of balance sheets, the model should be able to explain why agents other than the central bank (whose behavior is taken to be exogenous) were caught by the devaluation with a specific asset and liability structure. In our stylized model, devaluation expectations are uniform across agents. Of course, this is not uncontroversial (see Frankel and Rose 1994, sec. V, and the references therein). We prefer to allow for heterogeneity in more directly interpretable respects, namely, for differences in risk aversion, asset preference, and need for liquidity. Heterogeneous objectives lead different agents to take different positions in domestic and foreign assets and to react differently to changes in the perceived probability of devaluation. To the extent that our stylized agents may be taken to represent households, firms, financial institutions, and central banks, the model can be used to interpret certain characteristics of the data. The paper is organized as follows. Section 7.1 motivates our work with a review of readily available evidence on the Italian lira crisis in 1992. A first look at standard balance-of-payments statistics indicates that domestic banks, domestic nonbank investors, and foreign investors did contribute differently to official reserve losses. The theoretical analysis is organized in two main sections and several subsections. Section 7.2 sets up an accounting framework for the study of financial interactions among a central bank and a number of heterogeneous investors. In section 7.3, several elements of heterogeneity are considered and their implications in terms of portfolio choices and resulting gains and losses from devaluation evaluated. Section 7.4 goes back to the data, specifically to a variety of disaggregated statistics available in the Italian case. 1 We discuss the extent to which the peculiarities noted in section 7.1 may be rationalized by our theoretical considerations and the more disaggregated evidence. A concluding section summarizes the main findings and indicates directions for further research.
7.1
The Crisis of September 1992 in Italy
Our modeling perspective may be suitably applied to any period in which agents assign a positive (and nonnegligible) probability to the event of a change in the EMS parity grid. In this paper, we consider the lira exchange rate crisis of September 1992 and the summer months leading up to it. After some years of remarkable stability in the foreign exchange market (the most recent 1. We attempted to obtain similar statistics for other countries. Some Swedish data were kindly made available to us and may be considered in future work along the lines of this paper. Contacts with French and British central bank officials led to pleasant conversations, but no hard information. The officials we met readily acknowledged that the issues we focus on are of some importance and that balance-of-payment statistics throw very little light on them. Yet no better data appear to be available for those countries.
232
Fabio C. Bagliano, Andrea Beltratti, and Giuseppe Bertola
exchange rate crisis, leading to a devaluation of the lira, was in 1987), widespread fear of a realignment within the EMS (European Monetary System) originated from the negative and unexpected result of the Danish referendum on 2 June 1992. 2 The following four months witnessed sizable portfolio reallocations by various agents, with corresponding large changes in the reserve position of the central bank. This section provides a brief description of the events, highlighting the main facts that may motivate our theoretical analysis. The first thing to note is that June 1992 marks a change of some of the trends in capital movements characterizing the preceding months. In fact, the completion of the process of gradual capital movements liberalization in 1990 (together with the adoption of the "narrow band" within the ERM in January of the same year, supporting expectations of exchange rate stability) favored a protracted ·large outflow of domestic nonbanking capital (by more than L 48,000 billion in 1991), mainly due to portfolio reallocations of households and financial companies and a heavy inflow of foreign funds (by almost 37,000 billion). Notwithstanding the conspicuous deficit of the current account, amounting to 26,500 billion, the loss of foreign reserves of the Bank of Italy was limited to some 8,500 billion owing to the large inflow of foreign currencies through the banking sector. Between January and May 1992, the cumulative outflow of nonbanking domestic capital amounted to 39,100 billion and was offset by a net inflow of banking-sector capital of more than 35,000 billion. With a cumulative currentaccount deficit of 17,200 billion, foreign capital inflows of 17,500 billion (and small errors and omissions) limited the loss of central bank reserves to 6,300 billion. The Danish no vote in the referendum on the ratification of the Maastricht Treaty on 2 June marked the beginning of the EMS crisis: foreign exchange market participants, which had largely disregarded the fundamentals of ERM members, became more responsive to certain countries' macroeconomic disequilibria and their possibilities of converging according to schedule. The crisis led to the abandonment of the ERM by the lira and the British pound in mid-September. In the three months between June and August, there is evidence of differences in the behavior of the banking and nonbanking sectors of the Italian economy. In the face of a generalized rise in the perceived probability of a parity realignment within the EMS, different positions were taken by different agents. Figure 7.1 plots the monthly reserve loss of the Bank of Italy between February and December 1992, along with its determinants from balance-ofpayments statistics. The distance between the solid line and the thick dashed line measures domestic financial intermediaries' capital· flows: between June and August, this component sustained the financing needs of the Bank of Italy, 2. The data in table 5 in Eichengreen and Wyplosz (1993) indicate that almost 50 percent of surveyed dealers "first thought a change in ERM exchange rate was imminent" after the Danish referendum.
233
Heterogeneous Behavior in Exchange Rate Crises
15000.0 , . . . - - - - - - - - - - - - - - - - - - - - - - , 10000.0 5000.0 0.0
-+---=-"""=........;~----------_T___f:._:.__.:_:_--~
-5000.0 -10000.0 -15000.0 -20000.0 -25000.0
9201 9202 9203 9204 9205 9206 9207 9208 9209 9210 9211 9212 -
Fig. 7.1
Reserve change
Current Account
------
+ err.
and omiss.
+ Non-bank
CF
Italy's balance of payments, 1992
with a cumulative inflow of about 15,000 billion. Apart from the current account deficit (about 3,500 billion in cumulative terms), the bank's reserve losses were reflected in a private-sector capital outflow totaling 21,400 billion (or 33,000 billion including errors and omissions, which may at least in part represent transactions by families). 3 In early September, growing uncertainty in the run-up to the French referendum acted as a catalyst for the launching of speculative attacks, providing them with a fixed date. After heavy intervention by the Bundesbank and the Bank of Italy, and after an increase of 1.75 points in the discount rate, the lira was devalued by 6.76 percent on 13 September. After that, the Germans decreased the lombard by 0.25 and the discount by 0.5. A new attack was launched on 16 September; Britain left the ERM, Italy suspended the intervention limits, the peseta was devalued by 5 percent. From the balance-of-payments data of figure 7.1, which are available only on a monthly basis, we see that the crisis brought about a dramatic reserve outflow (minus 30,000 billion in September). Even more strikingly, this outflow was matched by banking-sector capital outflows by 26,000 billion: nonbanking capital flows were negligible over the month, even accounting for errors and omissions, while the current account deficit was similar to those of previous months at 4,800 billion. We feel that the remarkable apparent switch in domestic banks' portfolio 3. The balance-of-payments statistics do not register capital transactions unless they are carried out through authorized currency dealers, i.e., banks. Moreover, banks and individuals are not required to report operations valued at less than L 20 million.
234
Fabio C. Bagliano, Andrea Beltratti, and Giuseppe Bertola
choices deserves further attention at both the statistical and the theoretical levels. In interpreting the data displayed in the figure, however, one encounters various problems. First of all, the balance-of-payments data register end-ofperiod positions. The negligible change in private portfolio choices within the month of September might well hide offsetting movements before and after the devaluation. The lira/deutsche mark official parity was devalued by 7 percent on 13 September, and the exchange rate continued to depreciate after the abandonment of the ERM on 17 September. Speculation against the lira before each of those dates, and profit taking thereafter, would not be apparent in the available data. Second, the balance of payments registers only those transactions involving foreign counterparts. In a world of not only free capital mobility but also extensive financial deregulation, positions involving currency risk are routinely taken among agents based in the same country, while international transactions may well be denominated in either (or third) currencies. Data at higher-than-monthly frequencies are not available to us (nor indeed to the Bank of Italy itself). However, the exchange rate policy arm of the Bank of Italy (Ufficio Italiano Cambi, or UIC, which is in charge of such statistics) does maintain and publish extensive records of banks' foreign-currency positions with both foreign and domestic counterparts. We analyze such data in section 7.4 below, after a brief analysis of several theoretical issues. The very existence of banks, in fact, requires that economic agents be heterogeneous, for otherwise no transactions would occur except those involving the central bank. Accordingly, in the next section we formulate a simple model where agents may be heterogeneous under a variety of respects.
7.2
Theoretical Considerations
We begin by studying the balance sheets of the various agents. This is useful to provide an accounting framework that may connect the portfolio positions of different agents. We do not consider the complete structure of assets and liabilities: instead, we focus on the financial positions that are relevant to the problem at hand. In the simple structure with which we begin, there is one central bank and one private agent. The private agent obtains credit from the central bank in two different currencies and repays the debt at the end of the period. At time 0, the central bank's balance sheet reads (normalizing initial net worth to zero) (1)
On the left-hand side of (1), the central bank's assets include domestic credit do and foreign-exchange reserves, which amount to/o in foreign-currency terms and are converted into domestic currency at the initial exchange rate xo' On the right-hand side of (1), we have (domestic) high-powered money, mo' All terms are expressed in units of domestic currency. The distinction between "domestic credit" and "reserves" in (1) hinges on
235
Heterogeneous Behavior in Exchange Rate Crises
currency denomination, not on whether the counterpart of those assets is a resident of the domestic country or a foreign subject. In practice, however, domestic credit has resident counterpart and is in domestic currency; reserves have a nonresident counterpart and are in foreign currency. Throughout our discussion, we do not explicitly account for any agent's forward positions: there is no need to do so, in fact, as a forward position can be viewed as a combination of borrowing and lending in different currencies. The relevant portion of the public's balance sheet, also expressed in domestic currency, is a mirror image of the central bank's: (2)
Other assets and liabilities are also present but (by definition) net out to zero and will be irrelevant to our analysis. The economy's representative agent holds (domestic) money for transaction purposes and finances the long money position by shorting both domestic- and foreign-currency assets on the right-hand side. Below we discuss the criteria that are used to make optimal decisions. We consider only two periods or, better, an initial and a final position that might be thought of as embedded in an ongoing cash-in-advance sequence of models. Between time 0 and time 1, each (domestic-currency) unit of domestic credit yields a nominal interest i, while each (foreign-currency) unit of reserves yields nominal interest i*. For simplicity, money pays no interest but is held for transaction purposes. At time 1, the exchange rate is xl' and assets mature with interest. The exchange rate change (if any) and interest payments are reflected in two networth items shown in the balance sheets at time 1:
+ i) + 10(1 + i*)x 1 = rno + W;B,
(3)
do(l
(4)
rn o = do(l
+ i) + 10(1 + i*)x 1 + wi,
when expressed in domestic currency. There are now terms W;B and wi that account for a positive or negative net wealth at the end of period 1, as a consequence of movements of the exchange rate and of interest rate payments. Of course W;B = -wi: if there are only two agents, the gains of one are the losses of the other. We assume that the exchange rate at time 1 is determined on the basis of elements that are outside the control of the central bank, which does not have an objective function to maximize, and of the private agent. 7.2.1
Portfolio Choice
The central bank need not be concerned with its time 1 net worth, as m,onetary policy objectives probably dominate them. 4 Conversely, each individual member of the public gains utility from a low expected cost of debt and suffers disutility from a high variance of the same quantity. We specify the public's 4. The return on central bank assets affects the public sector's net worth and, eventually, the private sector's tax bill. Atomistic individual investors, however, have little reason to worry about that when deciding on their portfolio allocation.
236
Fabio C. Bagliano, Andrea Beltratti, and Giuseppe Bertola
objective function in terms of rates of return, and this can be interpreted as a second-order Taylor approximation of a constant relative risk-aversion utility function. 5 Let gross real returns be given by (5)
r == 1 + i -
(XI -
l)w
on domestic-currency denominated financial instruments and
(6)
r* == 1
+ i* + (x I
-
1)(1 - w)
on the asset denominated in foreign currency. To simplify notation, we normalize the initial exchange rate to unity: while Xo would in general be endogenous in a floating-rate situation, it is exogenously given in the precrisis EMS situation that we consider (if we abstract from within-band fluctuations), and it is harmless to normalize X o == 1. This specification also implies that the same correction is done to the returns of the two assets after these have been translated to either currency; in fact, one can think of the correction done to the return on foreign assets as the sum of the two components (XI - 1) and -w(x I - 1), the former necessary for going from foreign to domestic currency, the latter common to both assets and representing the net deduction that has to be made to domestic-currency net worth to express the individual's wealth in real terms. The parameter w determines how important exchange rate risk is for the investor's marginal utility of wealth. The representative investor derives utility from wealth expressed in domestic currency if w == 0; at the other extreme, only the foreign-currency value of time 1 net worth matters if w == 1; and 0 < w < 1 indexes intermediate cases. In standard finance-theoretic models with independently distributed asset returns and non-purchasing power parity (PPP) goods-price inflation, the weight w of the exchange rate in the investor's price index corresponds to his consumption-basket weight of goods priced in foreign currency (for a particularly clear exposition and references to earlier literature, see Dumas [1994]). This interpretation need not be appropriate in the extremely short-run perspective that we take in this paper: at the (say) weekly horizon when portfolio choices are made, the prices of both domestic and foreign goods are essentially fixed and independent of exchange rate developments. However, portfolio managers and investment returns are evaluated in terms of precise and presumably different currency denominations: German mutual funds aim to show good results in deutsche mark terms, and Italian banks aim to optimize net worth in lira terms. In general, this may be rationalized recognizing that asset returns, depreciation rates, and inflation rates are driven by underlying macro5. For a similar approximation, which usefully simplifies ,the algebra in a discrete-time framework, see Dornbusch (1983). For an extensive treatment of the relevant issues, see Adler and Dumas (1983). For a more rigorous discrete-time treatment, see Dumas (1994). Frankel (1986) and others have proposed macroeconomic and empirical applications of similar models.
237
Heterogeneous Behavior in Exchange Rate Crises
economic state variables. Thus, the relevance of short-term exchange rate developments to investment evaluation is mediated by their underlying relevance to the whole future path of utility-relevant prices and quantities. Formally, the adjustment represented by w refers to the wealth (rather than consumption) deflator that is appropriate when the underlying state variables' evolution results in an exchange rate equal to Xl at the end of the period. An Italian investor may be particularly concerned with lira returns and set w == 0 if the domestic currency is lira, not because he consumes only lira-price goods at time 1 (he may consume all sorts of goods, whose prices are anyway essentially constant at the weekly forecast horizon), but because the marginal value of time 1 wealth in his dynamic problem is high when lira assets do well (the lira appreciates relative to expectations), low when lira assets do poorly. However, the opposite might be true with a different assumption about the marginal utility of wealth in the two states, which would imply that a large w is attached to domestic investors. We specify a mean-variance objective function in rates-of-return terms and model the transaction services of high-powered money by an increasing and concave function T(·). Thus, the public's time 0 objective function reads (7)
max( T(mo)
+ mo{ £[1
~V [l
-
-
x.r -
- 'lI.r - (l - 'lI.)r*] (l - 'lI. )r*] } ),
where E denotes the expectations operator, V the variance operator, and X. is the share of initial debt denominated in domestic currency (there is no coefficient on the expectation, which just reflects normalization of T[·] and ",). Taking as given the rates of return and the distribution of the time 1 exchange rate, the public optimally chooses the amount of money held for transactions and portfolio composition. Our specification allows for a separation of the portfolio and of the initial debt problems: the investor chooses how much money to hold for transaction purposes, at which stage he needs to know how costly (in risk-adjusted terms) it is to finance money holdings; for each amount of money holdings, however, the portfolio shares are determined independently of the choice of initial debt. As we will see later, in general this does not imply that in equilibrium money demand is independent of the portfolio composition. To model the EMS precollapse situation, we let _ {I + Ll with probability p, 1 with probability (1 - p),
Xl -
where Ll is the (nonrandom) amount by which the domestic currency depreciates if the initial parity is abandoned. (We could let this be random, with messier algebra and little additional insight.)
238
Fabio C. Bagliano, Andrea Beltratti, and Giuseppe Bertola
The expected returns, variances, and covariances of domestic and foreign assets are easily computed:
Er Er*
= 1
=1+i-
+ i* +
pliw,
(1 - w)pli,
Vr
= (wli)2p(1
Vr*
- p),
w)Ii]2p(1 - p),
= [(1 -
Cov(r, r*) = -w(1 - w)li2p(1 - p). As for the role of portfolio composition, the expected overall opportunity cost of money holdings is, in gross terms, 1 + i* + A.(i - i* - pli) + (1 - w)pli
if the investor puts a share A. in domestic liabilities; the corresponding variance of realized returns is given by
and equals zero if A. = 1 - w (minimum variance portfolio). The derivative of the expected return with respect to A. is lip - i + i* and that of the return variance is 21i 2p( 1 - p )(A. + w - 1). Hence, to maximize the objective function (7), the portfolio share must satisfy the first-order condition
/1p - i + i* - 'Y [/1 2P (1 - p)(X- + w -
1)]
= 0,
from which we obtain A. = (1 - w) _
(8)
.
l -
('* l
+ PUA) .
)'li2p (1 - p)
If)' ~ 0, then A. diverges to plus or minus infinity according to the sign of the uncovered interest rate differential; if )' ~ 00, then A. ~ 1 - w. As for money demand, we have (9)
T' (rna)
+ [-i* - A.(i - i* - pli) -
~ {/1:P(l - p)[X- - (l -
(1 - w)pli] w)F} = 0,
or, considering (8), (10)
T
'( rna ) -_
.* + (1
Wl
-
)'
W l -
(i - i* - pli)2 . 2)'li2p (1 - p)
In the presence of risk aversion, the choice about the amount of initial debt depends on expected return and variance. Thus, the simple theoretical framework that we consider integrates money-demand and portfolio-choice aspects. Specifically, money demand turns out to depend, first, on a weighted average of the rates of return on assets denominated in the two currencies, with weights given by the asset-preference parameter w, and, second, on a downward adjust-
239
Heterogeneous Behavior in Exchange Rate Crises
ment of opportunity costs by a "speculative" term, which is an increasing function of the absolute deviation from uncovered interest parity (normalized by the exchange rate variance) and a decreasing function of the agent's degree of risk aversion: intuitively, an agent who is not infinitely risk averse will be able, by taking a "speculative" position, to reduce the financing cost of money holdings.
7.3
Market Equilibrium and Heterogeneity
While individuals are price takers and quantity setters, in equilibrium quantity and/or interest rates must adjust to their optimal choices. In a representative-agent framework of analysis, equation (8) would immediately determine the equilibrium risk premium as a function of the relative quantities of the two liabilities in the central bank's balance sheet; financial market equilibrium requires that
i - i* - pli
(11)
= 'Y1i2p(1
- p) [(1 - w) - ~].
The risk premium is equal to zero, of course, if asset supplies form a minimum variance portfolio for the representative investor; otherwise, the market must bear some risk and be compensated for it. If ~ > (1 - w), then i must be low, for any given pli, since investors want to be compensated with a lower cost of debt in order to keep in their portfolio a share of domestic liability that is larger than the one corresponding to their asset-preference parameter. If ~ < (1 - w), conversely, investors compete for domestic-currency debt and bid up the interest rate it pays. It is interesting to go beyond the standard representative-agent framework and try and rationalize the exchange rate crisis as the equilibrium outcome of interactions among heterogeneous agents. In reality, of course, there are very many potential sources of heterogeneity. The variables that we choose to emphasize in the model are the coefficient of risk aversion "I, the asset-preference parameter w, which the agent uses to transform nominal returns into real returns, and the objective function. In general, we may consider J investors with different coefficients w(j) and 'Y(j) and heterogeneous optimal portfolio shares given by 'A(j)
(12)
= (1 - w(j») _
i - (i* + lip). 'Y(j)p(1 - p)1i2
In market equilibrium, we must have d = o
L 'A(j)m(j) = L m(j)(1 J
j= 1
J
j= 1
w(j») -
•
('*
+
A
up p( 1 - P)1i2
l -
l
)
L ~(j) J
j= 1 'Y(j)'
from which an expression for the risk premium is readily obtained:
240
Fabio C. Bagliano, Andrea Beltratti, and Giuseppe Bertola
i - i* - dp
(13)
= p(1 - p)d 2
[
~J
m(j)(1 - w(j)) - d ]
£..;=1
.
Lf~l
0 .
mCJ ) 'Y(j)
The structure that we have described can be further enriched by modifying the objective function of some agents. In what follows, we consider the case of an intermediary, characterized by lack of demand for money owing to absence of transaction purposes. This intermediary has reasons to exist in equilibrium, as it helps close the gap between the supply structure of the central bank and the demand structure of the private agents. The balance sheet of such an intermediary is do + xofo = 00 + xo'o' where on the left we have the supply of the central bank and on the right we have the demand of the public. Net worth at time 1 is
w 1(I)
=
(00 - d o)(1 + i) + ('0 - 10)(1 + i*)x 1•
The portion of the (purely speculative) objective function regarding expected returns is E[(
°0 )r + (
00 + xo'o
xo'o 00 + xo'o
)r* (
do fox o + do
)r (
xofo toxo + do
)r*]
= E[A (l)(r - r*)],
where we use the fact that 00 + xo'o = fox o + do and the notation A(A)
°0 ,
=
A(L)
00 + xo'o
=
A(l) = A(A) - A(L).
do foxo + do'
By considering an objective function that also considers the variance of returns, we obtain a demand function:
A(l)
(14)
=
i - i* - dp 'Y(l)p(1 - p)d 2
where 'Y(I) is the coefficient of risk aversion of the intermediary. Equation (14) shows that demand of the intermediary completely ignores the minimum variance portfolio and requires a risk premium for intermediating between the structures of the balance sheets of the central bank and the public. In the absence of such a risk premium, A(l) = 0, or A(A) = A(L), and there is no economic role for the intermediary as the structure of his assets is identical with that of his liabilities. The equilibrium condition for a market with both an intermediary and many heterogeneous private agents is J
do =
L A(j)m(j) + ALm j=1
J
AA m =
L A(j)m(j) j=1
A(l)m.
241
Heterogeneous Behavior in Exchange Rate Crises
It follows that the equilibrium risk premium is now
(15)
i - i* - lip
== 1i2p (1 _ p)
(Lf=l 'A(j)"!(j) L~
'j=1
m(J)
dO). m
-~ +~(l) -
We see in (13) and in (15) that even the simple framework that we have been considering yields complicated equilibrium interactions between the various parameters. Moreover, money demand on the part of each agent is a function of the interest rates and of the various parameters, making it even more difficult to obtain sharp and general results. Suppose, for example, that one is interested in analyzing in the general case the reaction of the investors to a change in the probability of Qevaluation. The natural thing to do is to look at the demand function, equation (12), which, however, shows the importance of the risk premium. But equation (15) points out that the reaction of the equilibrium risk premium to a change in the probability of devaluation depends on the money demands of all the agents, in turn a function of the shares of the debts denominated in the two currencies. Lacking a closed-form solution, we set up two simple cases as candidates for understanding the patterns that we may see in the data. As we are interested in heterogeneity in the behavior of the agents as a reaction to changes in the probability of devaluation, we note that there are two factors in the structure of our model that may give rise to heterogeneity: on the one hand, changes in money demand (a scale effect); on the other, changes in the composition of portfolios (a share effect). In general, the two are interrelated, but, to build intuition, it is useful to consider each of them in isolation. 7.3.1
No Intermediary, No Risk Premium, Different Asset Preference
The first simple case that we consider is that where there are only two investors, which we regard as "domestic" (superscript D) and "foreign" (superscript F), respectively, and no intermediary. Further, we suppose that the structure of the supply of the central bank is such as to eliminate the risk premium, that is, Lf= 1(1 - w(j))m(j) == do- In this case, uncovered interest rate parity holds, that is, i == i* + pli, and the optimal share for each investor, from equation (12), equals the minimum variance portfolio. Also from equation (10), we see that money demand has a simple structure, T' (m(j)) == w(j)i* + (1 - w(j))i, for both the domestic and the foreign investors. What is the effect of an exogenous increase in p if W(F) ¥- W(D)? The scale effect can be seen clearly from equation (10), where the last fraction is equal to zero in light of uncovered interest parity: for given i*, an increase in p brings about an increase in the domestic interest rate, and this in turn decreases money demand. It follows that an increase in p decreases the stock of money. If the structure of the supply of the central bank does not change, then there is a decrease in the amounts of both domestic and foreign liabilities of the central
242
Fabio C. Bagliano, Andrea Beltratti, and Giuseppe Bertola
bank. The latter can be interpreted as a loss of reserves on the part of the central bank. The share effect is absent, as the desired share is equal to the one corresponding to the minimum variance portfolio. Note, however, that the scale effect by itself may yield a different reaction on the part of different investors in the holdings of debt denominated in the various currencies. From equation (12) is clear that, under the hypotheses of this particular case, the change in the holding of foreign debt is proportional to the change in money demand, jCj) == (1 - 'A cj)) rilcj) , where j == D, F, and where the dot over the variable indicates the (comparative statistics) change from one situation to another, in the current case from one with a given p to another with a larger p. The condition for having JCD) < jCF) is therefore in this case that the parameters WCD) and wCF ) are such as to maintain the following inequality: . CF)
'A CD ) < 1 - ~ (1 - 'A CF )). rilCD)
Of course, the size of the balance sheets (i.e., money demand) depends on financing costs and is heterogeneous across agents in general. If it were to happen that the change in money demand is the same for the two types of agents, the condition simplifies to 'A CD) < 'A CF) or, in terms of fundamental parameters, w CF ) < WCD). The natural question is therefore, Is it reasonable to assume that the inequality in fact represents the relative behavior of domestic versus foreign investors? As noted in section 7.2, this mayor may not be the case. What is important, however, is that the "scale" effect of money demand reduces the complicated general structure of the model to a simple mechanism that may be confronted with the character of available data. 7.3.2
One Intermediary and Different Degrees of Risk Aversion
In the second case, there is one agent (D) and one intermediary (F) with different attitudes toward risk: as we shall see, the intermediary's function is a meaningful one if this agent's risk aversion differs from that which would determine the equilibrium risk premium in his absence. As the intermediary does not hold any money, we have from equation (15) a relatively simple expression for the equilibrium risk premium:
i - i* - lip
= Ll2p (1
- p) (
1
~w:) i ~). ~CD)
~Cl)
Uncovered interest parity holds if 1 - WCD) == x., of course. In general, it is now possible to calculate the difference between the portfolio shares of the two agents:
243
Heterogeneous Behavior in Exchange Rate Crises
'A(D) -
'A(F)
=
-W(D) -
(1 -
W(D) -
x.)
"{(I) (
"{(I)
+
"{(D))
.
"{(D)
We find that a sufficient condition for 'A (D) < 'A (F) is (1 - W(D) - X.)('A (l) - 'A (D) > 0 and that 'A (D) is always less than 'A (F) when there is no risk premium. Thus, the model may rationalize heterogeneous portfolio composition (domestic- vs. foreign-currency-denominated assets) without relying on assetpreference parameters and without necessarily identifying the domestic investor as the one with a low w. Interestingly, differences in the change of the two agents' portfolio shares in response to sterilized intervention (an increase in x,) depend on differences in the coefficients of risk aversion. If the "foreign" intermediary is less risk averse than the private agent, then the difference decreases, and the domestic investor absorbs a larger part of the central bank's reserve loss.
7.4 Disaggregated Data We now return to the data. The nonbanking sector balance-of-payments data of figure 7.1 above, which still registeJ only international transactions, may be usefully disaggregated into domestic and foreign agents' capital inflows and outflows. Figure 7.2 shows the behavior of net total domestic and foreign capital movements (including loans, commercial credits, and direct and portfolio investments) from January to December 1992. Between January and May, the cumulative inflow of foreign capital reached L 17,500 billion. The following three months preceding the peak of the crisis see a marked slowdown in the net inflow of foreign capital: the factors behind this are the uncertainty on the political side, the negative effect of the insolvency of EFIM (a heavily indebted state-owned financial holding), and the general skepticism on the progress toward European integration following the result of the Danish referendum. 7.4.1
Foreign versus Domestic Investors
From our perspective, it is interesting to find that foreign investors as a group were "convergence players" during the buildup of the crisis: a net foreign-capital flow into Italy (and presumably into lira instruments) of some 3,000 billion during the June-August period contributed to limit the reserve losses of the Bank of Italy and increased the lira exposure of foreign residents. Conversely, domestic capital consistently flows out throughout the period, with an acceleration of previous trends to a cumulative figure of 25,000 billion, of which more than 20,000 is due to households' and firms' portfolio investments. Additional information on the likely heterogeneity of agents within the foreign and domestic nonbank sectors may be obtained by looking at the gross flows of foreign and domestic capital for portfolio investments. Figure 7.3 plots separately the amounts of portfolio investments and disinvestments by domes-
244
Fabio C. Bagliano, Andrea Beltratti, and Giuseppe Bertola
15000.0 - , - - - - - - - - - - - - - - - - - - - - - - - - - - - - ,
10000.0
5000.0
0.0
-+-------------::::-......:~--~---------l
-5000.0
-10000.0
-1 5000.0
--'---,------,---r---.---r--....---ir-------.-------r-----.------r---.----.J
9201 9202 9203 9204 9205 9206 9207 9208 9209 9210 9211 9212
-
Fig.7.2
Total Domestic
-
Total Foreign
Nonbank capital movements (net), 1992 60000.0 - r - - - - - - - - - - - - - - - - - - - - - - - - - , 55000.0 50000.0 45000.0 ~
40000.0
::i
..6
35000.0 30000.0 25000.0 20000.0 15000.0 ...L..--,------,r----,---..,---,----r--...--;-----,r----.----r--.---' 9201 9202 9203 9204 9205 9206 9207 9208 9209 9210 9211 ~212
-
Fig. 7.3
Domestic dis.(+) ....... Domestic inv.(-) -
Foreign inv.(+)
------- Foreign dis.(-)
Portfolio investments: domestic and foreign
tic and foreign agents in 1992. For both sectors, investments and disinvestments tend to move in the same direction. The moderate net inflow of foreign capital over the period is the result of large investments and disinvestments, reaching a peak of more than 55,000 billion in June and decreasing thereafter to a minimum of around 35,000 billion in August. To the extent that transac-
245
Heterogeneous Behavior in Exchange Rate Crises
tions of opposite sign are undertaken by the same agents within a month, such large gross figures may simply indicate high volume in the volatile April-July period, when many within-month round-trips were probably undertaken by active investors. 6 Such transactions were ex post inconsequential from the exchange rate point of view, given that no devaluation took place and exchange rate changes were extremely limited, but may have been profitable (or unprofitable) in light of the increasing and volatile interest rate differentials. Alternatively, high volume of both investments and disinvestments may reflect heterogeneity within the "foreign investor" aggregate-which unfortunately remains unobservable, as the data provide no disaggregation of foreign investors' positions with Italian counterparts. Domestic capital shows the same tendency, with investments and disinvestments rising until July to a peak of 45,000 and 35,000 billion, respectively. Once again, while part of this common movement in capital inflows and outflows may be the result of round-trip operations concluded within the month, the evidence may also suggest that the categories of foreign and domestic investors consist of classes of agents characterized by different behavior. 7.4.2
Domestic Banks and Investors
Within the banking-sector flows of figure 7.1 above, available statistics make it possible to analyze the net banking position in foreign currency, distinguishing between foreign and domestic counterparts, and to infer the extent to which banks maintained open positions in their own account, in addition to conduct.. ing foreign-currency operations motivated by the financing and investment needs of their domestic customers. In figure 7.4 we plot the changes (at constant exchange rates) in the foreign-currency position of resident banks from June to December 1992. We distinguish between changes due to spot operations with foreign counterparts, spot operations with domestic counterparts, and forward transactions. The total change in the spot position with foreign and domestic counterparts (the dashed line in fig. 7.4) shows that banks did not appreciably modify their net portfolio in the months preceding the crisis. Indeed, the inflow of foreign currency over June-August (by 14,700 billion, due to an increase in foreign-currency-denominated liabilities of 8,900 billion) was accompanied by a parallel change in the currency position with resident counterparts by some 10,600 billion. This is the result of an increase of banks' lending in foreign currency (mostly U.S. dollars, deutsche marks, and EeUs [European currency units]) to domestic agents (particularly nonfinancial firms) by 16,700 billion and an increase in liabilities (foreign-currency deposits by residents) by 6,100 billion. It is remarkable to find that, in June and July, domestic firms (and possibly households) borrowed a total of some 8,000 billion in foreign exchange from domestic banks. 6. Official sources do not indicate how frequently (daily, weekly, or even biweekly) "gross" portfolio transactions are collected and aggregated.
246
Fabio C. Bagliano, Andrea Beltratti, and Giuseppe Bertola
25000.0..,.----------------------,
....•.......•......... -5000.0
-10000.0
~~
_
.
-'----r----r----,----r------r----.,.-----,-----J
9206
9207
9208
9209
9210
.............. Spot Foreign count........ Spot For. and Dam. -
Fig. 7.4
9211
9212
Spot + Forward
Banks' changes in foreign-currency position
Once the change in the forward position is considered, it becomes evident that the banks' overall position changed only slightly over the whole JuneDecember period (the solid line in fig. 7.4). This is most clearly visible in September, when the spot position displayed a substantial change, resulting in a net outflow of foreign currency by 21,700 billion, only partially matched by a negative change in the foreign currency position toward domestic counterparts (due more to an increase in liabilities than to a reduction in assets). In this month, banks changed their net forward position by 16,800 billion, determined by an increase by 25,200 billion in forward debts mostly with nonresident counterparts. Indeed, throughout the June-September period, forward operations by banks were conducted mostly with nonresident agents, suggesting that, although domestic firms increased their borrowing in foreign currency from the banking system, they did not cover their foreign-currency position, relying on the stability of the lira exchange rate. As noted by Eichengreen, Rose, and Wyplosz (chap. 9 in this volume), Bank of Italy (1993), and IMF (1993), the mechanics of balance-of-payments crises fueled by foreign speculators entail lending by domestic institutions in domestic currency to foreign residents. Available data indicate that Italian banks did not engage in such activities, but other financial intermediaries (Istituti di Credito Speciale) did increase their foreign position in lira by some 7,000 billion between June and September. Inasmuch as it does not entail foreign exchange rate risk, such lending is not directly relevant to our analysis. It may indicate, however, that some foreign agents did take short positions against the lira, even though, on net, the foreign investor's aggregate was mildly supportive of the lira (see fig. 7.2 above).
247
Heterogeneous Behavior in Exchange Rate Crises
The distinction between households and firms within the domestic-investor aggregate is obviously very relevant to our perspective. The VIC kindly made some unpublished data available to us, which we plot in figures 7.5, 7.6, and 7.7 and on which we briefly comment below. In all these figures, and especially in figure 7.5 (financial companies), we again see high volatility, which is unlikely to be due to heterogeneity within such relatively narrow categories. We find in figure 7.6 that domestic nonfinancial firms increased assets with foreign counterparts by some 1,000 billion net over the period; these were presumably not the same firms that took positions of similar size and opposite sign with domestic banks, and it would be interesting to explore the source of such heterogeneous behavior. The behavior of households, in figure 7.7, is "speculative" but relatively smooth, with little acceleration of the steady capital flow resulting from the previous liberalization of cross-border financial transactions. 7.4.3
Gains and Losses from Devaluation
In September, the capital-flow trends were reversed: foreign capital flows show a negative, although small, figure of 500 billion, whereas domestic capital outflows came to a stop. However, the outflow of only 350 billion over the whole month is likely to be the aggregate result of two very different patterns before and after the devaluation and the suspension of the lira from the ERM. In fact, in the following quarter, until the end of 1992, large net inflows ~f domestic capital were registered, for a cumulative figure of 23,500 billion, as portfolios were reallocated to take in profits from devaluation. Meanwhile, foreign capital inflows resumed with a view to exploiting capital gains generated by the likely decreases of interest rates, contributing to rebuild the central bank's reserves by some 7,500 billion over the final quarter of 1992. The bulk of capital flows during the whole year, both domestic and foreign, are attributable to the portfolio investment component, to which households and financial companies are likely to contribute most. As already said, data for September must be interpreted with care since they possibly aggregate different behavior before and after devaluation. Moreover, a general reason for caution is the use, by both banks and nonbanking agents, of domestic currency swap (DCS) contracts to cover (or to take) a foreigncurrency position. Such operations have the same nature of forward transactions in foreign currency but only entail payments in lire, amounting to the difference between the spot exchange rate at the date of maturity of the operation and the rate agreed on at the contract date. Data on DCSs are available only for the banking sector from January 1993. 7 Our model of market equilibrium under heterogeneity points out many reasons why different agents may be differently exposed to the risk of devaluation. 7. Inspection of 1993 data indicates that DCSs were used by banks mainly to offset their overall spot + forward position, which was much larger than that shown in the figure. If banks' behavior was similar in the months of interest, when the spot + forward position of the banking sector was close to zero, the amount of offsetting DCSs may have been small.
248
Fabio C. Bagliano, Andrea Beltratti, and Giuseppe Bertola
30000.0 - - , - - - - - - - - - - - - - - - - - - - - - - - .
25000.0
20000.0 15000.0 10000.0
5000.0 0.0
-4-------------------.,.~----------j
-5000.0 -10000.0
-'---.,.---,...---,...---..,---.,---.,---,.---.----.r----.---,---l
9201 9202 9203 9204 9205 9206 9207 9208 9209 9210 9211
-
Fig. 7.5
Investments
Disinvestments -
_u
Balance
Portfolio transactions of financial companies 5000.0 - . - - - - - - - - - - - - - - - - - - - - - - - - - , 4000.0 ................
3000.0 2000.0
..............
.....
---_ .. ----_ ...
1000.0 0.0
-+---------------~---+----;
~
-1000.0 -2000.0 -3000.0 -4000.0
-5000.0 - ' - - - - , - - - . , - - - - - , - - - - , - - . , . - - . , . - - - - r - - - - r - - - , - - - , - - - r - - - ' 9201 9202 9203 9204 9205 9206 9207 9208 9209 9210 9211
-
Fig. 7.6
Investments
m
Disinvestments -
Balance
Portfolio transactions of nonfinancial private companies
Considering a theoretical model allows several intuitive insights. The size of interest differentials is endogenously determined by our model, so that ex ante market equilibrium implies indifference for every agent, given expectations and risk aversions. But, of course, heterogeneous portfolio composition (for whatever reasons) implies differential ex post gains and losses from devalua-
249
Heterogeneous Behavior in Exchange Rate Crises
1 0 0 0 0 . 0 . . . . . . - - - - - - - - - - - - - - - - - - , - : - , . ' - . ' . - .- - , 8000.0
6000.0 4000.0 2000.0 0.0
-+----------------~~=-----------1
-2000.0
-4000.0
9201
-
Fig. 7.7
9202
9203 9204 9205 9206
Investments
9207 9208
------- Disinvestments -
9209 9210 9211
Balance
Portfolio transactions of households
tion: if and when the devaluation is realized, "convergence players" lose, while "speculators" gain. Starting from the evidence discussed in previous subsections, we now try and quantify the gains and losses from devaluation for various (classes of) agents. We first consider the buildup of the crisis (from June to August 1992), when a generalized increase in the perceived probability of devaluation occurred, leading investors to reallocate their portfolios of lira and foreigncurrency assets and liabilities. Given the yield differential prevailing throughout the period between Italian and foreign short-term financial instruments, such reallocations entailed gains and losses on the portfolio returns. In the calculations, we used the spread between the Italian one-month interbank rate and an analogous rate for Germany as a "representative" yield differential. We also assumed that the exchange rate is fixed, given that within-band movements of the lira nominal exchange rate were very small on a monthly average basis. Cumulating the central bank reserve losses displayed in figure 7.1 above, and applying the interest rate differential shown in figure 7.8 to the resulting change in foreign-currency position between June and August, we find that the private sector as a whole suffered an ex post interest-differential loss ofL 325.9 billion. Applying the same procedure to the disaggregated capital flows of figures 7.2 and 7.4, we find that domestic investors suffered a L 276 billion loss on account of the interest-rate differential, while foreign investors gained (if only about L 20 billion) by playing on convergence during this period. Apart from the current account and the "errors and omissions" entry (which may reflect unrecorded capital flows of unknown origin), the last component of the
250
Fabio C. Bagliano, Andrea Beltratti, and Giuseppe Bertola
25.0.,.---------------------------,
20.0
III
C
·0 15.0 0.
~
10.0
5.0
--_ .. -----_ .. -_ .. ------ ------_ ... ---_ .. -----------------------------_ .. -- ..
--'---.---.___--.-_.,_-~----.----_._-,__-._____r-_.,_-____.__--'
9201
9202 9203 9204 9205 9206 9207 9208 9209 9210 9211
-
Fig.7.8
Italian rate
m
9212
German rate
Money market interest rates: Italy and Germany
balance of payments on a spot basis is that labeled "authorized foreigncurrency dealers" (essentially Italian banks). As noted above, this component appears very large in the "convergence" direction on the basis of balance-ofpayments conventions, resulting in a cumulative interest rate differential gain of some L 175 billion; however, offsetting transactions with domestic counterparts reduces the gain to only 67 billion on the banks' own account, while those domestic agents who were borrowing in foreign currency saved some L 106 billion in interest payments. Additional transactions on the forward market further reduced the interest rate gain of banks as well as their exposure to devaluation risk: the counterpart of such transactions is mostly nonresident in June, while the following months see the banks engaged in intermediation between resident and nonresident agents (nonresidents take long lira positions with banks on a forward basis, while residents' forward contracts "will deliver" foreign currency to banks). In September, the lira devaluation was realized: the 7 percent devaluation of the official ERM parity on 13 September was followed by further depreciation after the subsequent abandonment of the ERM. For the purpose of our rough calculations, we apply a 15 percent devaluation to the change in foreign exchange positions of the various categories of investors we consider. To the change in foreign-currency exposure between June and August, we should of course add the further investments or disinvestments in September before the devaluation date. Unfortunately, as already mentioned, time-disaggregated data are not available for September. Therefore, any calculation of devaluation gains and losses necessarily relies on hypotheses about the evolution of the
251
Heterogeneous Behavior in Exchange Rate Crises
agents' net position in September. In what follows, we make two simple (albeit extreme) assumptions as to the within-month behavior of domestic and foreign nonbank investors. First, we may take the net change for the month of September to have occurred before the devaluation: then, the cumulative central bank reserve loss would amount to L 52,743 billion between 1 June and the devaluation date, to imply an interest rate loss of L 744 billion and a capital gain of L 7,163 billion for private-sector counterparts. 8 On a disaggregated basis, the total change in domestic investors' foreign exchange positions between June and the end of September would have afforded a capital gain of about L 3,810 billion on devaluation, which of course dwarfs the L 516 billion June-September loss on interest rate differentials. Conversely, foreign investors would have suffered an overall loss of some L 350 billion, mainly due to the L 390 billion capital loss on devaluation. As to Italian banks, their spot position with foreign counterpart becomes much more positive in September, both because liabilities decrease and because assets increase (about half the net change is due to each). Their spot position in foreign currency with domestic counterpart decreases (- 5,600 billion); about half of this reflects a decline in foreign-currency loans to domestic customers, the rest an increase in foreign-currency liabilities (foreigncurrency bank accounts). Overall, calculations based on cumulative spot position changes from June to 30 September (evaluated at constant exchange rates) say that Italian banks should have gained about L 1,700 billion as a result of a 15 percent devaluation. As we know, however, their foreign-currency position was almost fully offset by counteracting transactions with domestic resident participants (whose total losses would be put by these calculations at L 563 billion on account of interest differential gains, + 152, and capital gain losses, -736) and on the forward market: the Bank of Italy did not take large forward positions (or, at least, none are apparent in available statistics), indicating that the sizable forward positions of the Italian banking sector must have had private counterparts. 9 Alternatively, we may try and estimate devaluation gains and losses allowing for a reversal of capital flows within the month of September. For example, we can impute all the gross outflows registered in September to the period before the devaluation. This admittedly extreme assumption would see domestic investors reducing their exposure to devaluation by some L 56,000 billion in the first half of the month: then, the cumulative change in their foreign-currency 8. Our model could be readily extended to allow for interactions between different central banks: in fact, the Bank of Italy's "reserve loss" largely consisted of an increase in official shortterm liabilities (L 20,000 billion were obtained from the ERM short-term loan facility, and L 2,300 billion were drawn [in deutsche marks] from a Bank for International Settlements credit line). Such positions, which on devaluation result in transfers between different countries' official institutions, are not directly relevant for the calculations that we perform here. 9. This is in interesting contrast with the Swedish data that we plan to analyze in future work. In Sweden, the private sector took positions against the krona mainly in the forward market, and the Swedish central bank absorbed such positions in the crisis period.
252
Fabio C. Bagliano, Andrea Beltratti, and Giuseppe Bertola
holdings since June would imply a capital gain of over L 12,000 billion from a 15 percent devaluation, and a large part of these would have been realized in the second half of the month, as the gross inflows of about L 57,000 billion would be the result of profit-taking portfolio reallocations under the assumption under consideration. The even larger gross flows in foreign investors' portfolios would similarly imply capital gains or losses on the order of L 10,000 billion.
7.5
Concluding Comments
This paper has tried to bring theoretical analysis and Italian data to bear on the redistributive effects of an exchange rate crisis. These take place certainly between central banks and the private sector as a whole and, to the extent that different agents' behavior is heterogeneous, also across different classes of private investors. While the theoretical model that we have set out is admittedly very simple, it suggests that economic theory may be able to rationalize differences in the behavior of "foreign" versus "domestic" agents, or of end users of liquidity versus financial intermediaries, without resorting to differences of opinion or information across such classes of investors. As we feel that it is conceptually unappealing to allow for heterogeneous expectations, we would prefer to try and rationalize what evidence is available, extending standard models to account for investor heterogeneity under different and more readily interpretable respects. Understanding the welfare effects of currency crises may be relevant to a better understanding of alternative exchange rate regimes. No exchange rate is truly and permanently fixed, and the welfare effects of exchange rate crises may be so large as to outweigh the gains from periods of relative exchange rate stability. Even disregarding such normative issues, however, further theoretical and empirical work on the microfoundations of transactions that result in reserve loss is certainly relevant to the mechanics of exchange rate.crises and of their resolution. Allowing for heterogeneous behavior across different classes of investor is likely to afford further insights in future research. Of course, not all dimensions of heterogeneity are necessarily relevant to the problem at hand: to the extent that different portfolio positions reflect unexplained "noise" and cancel each other out, they may be safely disregarded at the aggregate level. Useful theoretical models need to address heterogeneity not only across a central bank and a vaguely defined "market" but also across agents whose different objective functions and constraints are readily interpretable from an economic point of view and, one hopes, observable in the data. To an investment banker, an exchange rate crisis may be an opportunity to "ride the tide" and come out of the crisis with as large a portfolio value as might be obtained by moving from one currency to the other: such speculative objectives are sought by investors
253
Heterogeneous Behavior in Exchange Rate Crises
of this type, who face few institutional constraints on their portfolio policies. To a household that is simply concerned with the consumption of a basket of domestic good, the crisis need not be as exciting until its final effects on prices and/or wages unfold: even though a rational wealth maximizer should never give up the opportunity to increase wealth with a timely round-trip in foreign currency, suitable investment vehicles need not be readily available to households and other unsophisticated investors.
References Adler, Michael, and Bernard Dumas. 1983. International portfolio choice and corporation finance: A synthesis. Journal of Finance 38:925-84. Bank of Italy. 1993. Bollettino Economico, no. 20. Rome, February. Dornbusch, Rudiger. 1983. Exchange rate risk and the macroeconomics of exchange rate determination. In The internationalization offinancial markets and national economic policy, vol. 3, ed. R. Hawkins, R. Levich, and C. Wihlborg. Greenwich, Conn.: JAI. Dumas, Bernard. 1994. Partial-equilibrium vs. general-equilibrium models of international capital market equilibrium. In Handbook of international economics, ed. R. van der Ploeg. Oxford: Blackwell. Eichengreen, Barry, and Charles Wyplosz. 1993. The unstable EMS. Brookings Papers on Economic Activity, no. 1:51-143. Frankel, Jeffrey A. 1986. The implications of mean-variance optimization for four questions in international macroeconomics. Journal of International Money and Finance 5:S53-S75. Frankel, Jeffrey A., and Andrew K. Rose. 1994. An empirical characterization of nominal exchange rates. In The handbook of international economics, ed. G. Grossman and K. Rogoff. Amsterdam: North-Holland. International Monetary Fund (lMF). 1993. International capital markets, part I: Exchange rate management and international capital flows. Washington, D.C. Obstfeld, Maurice. 1994. The logic of currency crises. Working Paper no. 4640. Cambridge, Mass.: National Bureau of Economic Research.
Comment
Lorenzo Bini-Smaghi
The paper by Bagliano, Beltratti, and Bertola (henceforth BBB) can be broadly divided into two parts. In the first, the authors develop a theoretical model in which heterogeneous investor behavior is explained on the basis of differences in terms of agents' degree of risk aversion, asset preference, and objective
Lorenzo Bini-Smaghi is head of the Policy Division of the European Monetary Institute. The author thanks C. Bentivogli for useful suggestions. The author is entirely responsible for the opinions expressed.
254
Fabio C. Bagliano, Andrea Beltratti, and Giuseppe Bertola
function, while assuming, instead, for the purpose of the analysis, homogeneous expectations. In the second part, BBB examine Italian balance-ofpayments data to assess possible heterogeneity in agents' behavior and estimate the effect of the September 1992 devaluation of the lira in terms of gains and losses for the central bank, residents, and nonresidents. My comments concentrate on the second part. They mainly address the question of using balance-of-payments data for the analysis of heterogeneous behavior. It is suggested that great caution is required in the use of such data because of their limited quality and significance and that their usefulness for the analysis may be seriously impaired. Balance-of-payments data may seem an obvIous choice for the analysis of heterogeneous behavior since they report transactions between buyers and sellers of financial assets. However, not much further information on these agents is provided, except that they are located in different countries, which makes it very difficult to assess heterogeneity on the basis of possible differences in risk aversion, asset preference, or objective function, the main features of BBB's theoretical model. Furthermore, balance-of-payments data have several shortcomings. First, their statistical accuracy is rather limited, even in industrial countries. Looking in particular at Italian data, it is noticeable that in the period June-August 1992 the "errors and omissions" were recorded at L 11 trillion, about half the size of the change in official reserves and nearly three times the current account balance. More important, international banking statistics, which are generally considered to be more reliable, show for Italian banks an enormous discrepancy between assets and liabilities: L 65 trillion in September 1992 (20 more than in January 1992), which is more than half their net external asset position. Second, balance-of-payments statistics do not provide full information on the currency denomination of assets and liabilities sold or acquired by residents and nonresidents in a given period. It is therefore very difficult to draw inferences concerning the effect of balance-of-payments transactions on currency diversification or to calculate who lost and who gained from a devaluation. This shortcoming might have become particularly relevant for Italy in view of the rapid development of the Eurolira market, which led to a substantial increase in lira-denominated flows of funds between the domestic market and the Euromarket. Third, balance-of-payments statistics do not record a wide set of transactions undertaken by residents and nonresidents aimed at hedging the risk incurred in foreign exchange operations, for instance, domestic currency swaps. The data for the latter were not recorded by Italian banking statistics before the end of 1992. However, the data available thereafter show that the amount of such transactions has been comparable to the currency exposure of banks, suggesting that not taking this type of information into account may lead to biased conclusions.
255
Heterogeneous Behavior in Exchange Rate Crises
Let me tum now to some of the findings of the paper and indicate how they may be affected by the above comments. In section 7.4.1, BBB note that, during the buildup of the crisis (JuneAugust 1992), Italy recorded an inflow of foreign capital while residents were increasing their outflows, which would suggest that, while the former were "convergence players," the latter were instead diversifying their portfolio away from lira-denominated assets. This conclusion may not be fully justified. First, the acquisition of foreign assets by residents and nonresidents should be viewed as part of their respective portfolio diversification strategies; account should also be taken of investment in domestic assets. In the case of Italy, residents' acquisition of foreign assets must also be viewed in terms of the portfolio adjustment that took place after the liberalization of capital movements in 1990. The developments mentioned by BBB could therefore be consistent with homogeneous behavior if residents and nonresidents increased their respective holdings of foreign assets as part of a similar portfolio diversification process. Second, balance-of-payments data do not indicate whether the foreign inflows of capital recorded in the summer of 1992 implied an increase in nonresidents' long lira positions or whether they were instead covered for the exchange rate risk. Similarly, the outflow of funds by Italian residents may to a large extent have been accompanied by borrowing in foreign currency, as banking statistics indicate, or other forms of hedging against exchange rate risk. The larger the hedging component, which is not reported in balance-ofpayments data, the more residents and nonresidents' behavior may in fact have been rather homogeneous while appearing different. A definite answer requires more information than is provided in figure 7.2. Another statement in section 7.4.3 suggests that after the devaluation Italian residents sold back foreign assets "to take in profits from devaluation" while nonresidents were resuming inflows "with a view to exploiting capital gains generated by the likely decreases of interest rates, contributing to rebuild the central bank's reserves by some 7,500 billion over the final quarter of 1992." This apparent difference in behavior may not correspond to reality as the increase in official reserves was obtained mainly through swaps with commercial banks of funds mostly borrowed abroad by the latter. More generally, the reflow of capital by residents seems to have been accompanied by hedging of the foreign exchange risk through operations that, as BBB suggest, may have taken the form of domestic currency swaps. Indeed, the developments in the exchange rate of the lira after the devaluation of September 1992 confirm that the reflow of portfolio capital was largely hedged by both residents and nonresidents. This would suggest that, even after the devaluation of the lira, residents and nonresidents displayed rather similar investment behavior. In section 7.4.3, BBB seem to determine who lost and who gained from the lira devaluation of September 1992. This exercise is undoubtedly very difficult,
256
Fabio C. Bagliano, Andrea Beltratti, and Giuseppe Bertola
Table 7e.1
Italian Foreign Assets and Liabilities (end·1991) Proportion:
Assets: Loans Commercial creditsa Bank assets Official reserves Other Liabilities: Loans Commercial creditsa Bank liabilities Other liabilities Of which foreign official debtb
Total (trillion lire)
Of Total
In Lire
488.3 24.3 52.3 124.7 95.9 191.1 608.7 124.2 31.1 275.9 177.4 72.75
100.0 5.0 10.7 25.5 19.6 39.2 100.0 20.4 5.1 45.3 29.2 12.0
16.0 55.5 39.0 10.9 12.6 24.2 46.5 34.0 13.1 82.3 35.5
Source: Banca d'Italia, Bollettino economico, no. 20 (p. 39), and estimates based on VIC (Vfficio Italiano Cambi) data. aEstimates. bTreasury bills and proportion of official loans held by nonresidents.
especially in view of the comments made above, that is, that balance-ofpayments data do not indicate the extent to which foreign exchange risk is hedged. Furthermore, BBB conduct this exercise on the basis of flow data, that is, the changes in the portfolios of residents and nonresidents, rather than on the stocks of assets and liabilities underlying the net external position of Italy. The data on stocks at the end of 1991 show that Italian residents held L 488.3 trillion of foreign assets, of which 95.9 trillion was official reserves, while foreign liabilities amounted to 608.7 trillion. The data can be updated to September 1992, using balance-of-payments data, in particular to take account of the decrease in official reserves (by 59 trillion) and the increase in private residents' foreign assets. However, no estimate is yet available of the currency composition of the flows during these nine months. Conducting the exercise on the basis of end-1991 data (table 7C.l) shows that, excluding official reserves, the private sector held L 329.6 trillion of assets denominated in foreign currency, as against 310.8 trillion of liabilities. Thus, the net foreign currency position of the Italian private sector was slightly positive, by L 18.8 trillion. Its net position in lire was negative, by 235.2 trillion. A devaluation of the lira, with unchanged end-1991 stocks, improved the already positive net foreign asset position of Italian residents, measured in lire; conversely, it worsened the net position of nonresidents vis-a.-vis Italy, measured in foreign currency. The above considerations suggest that the gains and losses calculated on existing stocks are much larger than those measured by BBB on the basis of 1992 flows. An appropriate calculation of the effects of the devaluation would
257
Heterogeneous Behavior in Exchange Rate Crises
therefore need to be conducted on the basis of an integrated stock-flow analysis. Given the shortcomings of balance-of-payments data, one possibility for further analysis of heterogeneity would be to use survey data. The Group of Ten conducted an assessment of the causes of the 1992 crisis on the basis of a survey of the behavior of financial intermediaries in the major countries. In Italy's case, some interesting features emerge from the survey: • Although market participants developed a common assessment of the causes of the September 1992 crisis and behaved in a similar way, nonresident financial institutions appear to have been more active than the resident ones. • When the crisis became more acute, the behavior of the various operators became more similar, as "all participants were on the same side of the market." • The lira exchange market appears to be structured in a somewhat peculiar way, with two tiers: the first made up of a handful of large institutional investors that sometimes take very large positions and appear to have an information advantage and to playa role in forming the views of other participants; the second tier is made up of minor Italian marketmakers and final customers, who are largely price takers and generally more cautious. The survey analysis points to heterogeneous behavior in the foreign exchange market, although in terms not of residents and nonresidents but of marketmakers and others, leaders and followers. To a certain extent, the distinction may overlap with that of balance-of-payments since most marketmakers are foreign financial institutions, but the analytic and policy implications may be different. These, however, are only suggestions for further analysis of a subject that appears to be of particular interest for the understanding of the functioning of foreign exchange markets. It is a merit of BBB 's paper to have contributed and stimulated the discussion on this issue.
Comment
Richard K. Lyons
The paper's main objectives are (i) to demonstrate empirically that position taking in 1992 differed across groups and (ii) to rationalize those differences without resorting to differential information. My comments pertain more to the latter. I will begin with some broader observations and then move to the model's specifics. Richard K. Lyons is associate professor in the Hass School of Business at the University of California, Berkeley, and a faculty research fellow of the National Bureau of Economic Research.
258
Fabio C. Bagliano, Andrea Beltratti, and Giuseppe Bertola
First, I want to link this paper to the topic of the conference, namely, microstructure. Microstructure, broadly defined, encompasses at least three dimensions of the market: (1) physical structure (e.g., number of dealers), (2) informational structure (e.g., transparency of order flow), and (3) agent structure (e.g., discretionary vs. nondiscretionary liquidity traders). The contact point between microstructure and the Bagliano, Beltratti, and Bertola (henceforth BBB) paper is squarely in the third dimension-agent structure-the first two dimensions playing no role in their model. In BBB, what distinguishes a class of agents is the direct linkage to the domestic money supply: nominal domestic balances enter utility directly, together with the cost of providing domestic balances. Next, I want to comment on the association between crises and microstructure, an association that runs through much of part II of this volume. This focus might lead one to believe that the impetus for foreign exchange microstructure comes largely from EMS (European Monetary System) crises. My view is different. In my view, the impetus for microstructure comes primarily from the floating experience: (i) extraordinary volume, both levels and growth; (ii) increased volatility under floating without increased volatility of fundamentals; (iii) the 1980s dollar cycle; and (iv) forward discount bias. (Microstructural tools are well suited to address the first two; the last two are tougher.) To help judge the impetus from EMS crises, I pose the following question: If the floating dollar had never occurred and all we had were the experiences of the EMS, would we be pursuing microstructure at this juncture? I suspect not. I turn now to the BBB model. Some models are easier to take to the data than others. This model is one of the others. The result is a virtual decoupling of their two central objectives: empirical demonstration of differential position taking and theoretical rationalization without differential information. Although the authors are careful not to oversell the paper's connection of those objectives, the marriage is uneasy nonetheless. That said, the model is insightful in its integration of money demand with portfoljo choice. This is achieved by specifying the (domestic) public's balance sheet as equal to the flip side of the central bank's balance sheet. A cost of this approach is that the portfolio decision is therefore rather limited; but this cost is small in relation to the elegance of the approach. In the model, the central bank is passive. Its balance sheet is determined by the portfolio choice of the public. In choosing its portfolio, the public trades off the utility derived from (nominal) monetary assets against the disutility from the liabilities that finance the monetary assets. The disutility from the liabilities has two sources: (i) debt servicing costs and (ii) variability in debt servicing costs. Crisis (devaluation) influences the public's trade-off through both these sources since the cost of debt is a function of the stochastic exchange rate. From this structure, the authors are able to derive equilibrium portfolio implications. By introducing heterogeneity in risk aversion, liquidity
259
Heterogeneous Behavior in Exchange Rate Crises
preference, or asset preference, the authors can rationalize differential position taking without resorting to differential information. There are a number of nagging issues within the original version that the authors should clarify. I will touch on three in particular. First, where do the nominal interest rates corne from in this model? These are central to determining the cost of the public's liabilities. Although they appear exogenous through the derivation of market equilibrium, the authors' example has the domestic rate moving with a change in the probability of devaluation. The second nagging issue is the evolution of the central bank's balance sheet. What does it look like at t == I? Where does the return on its assets enter? Why doesn't the (domestic) public internalize this? Clarification along these lines will help support the sharp dichotomy between central bank and public. The third nagging issue is their first simple case. In particular, the experiment considers an exogenous increase in the probability of devaluation p under heterogeneous asset preferences, domestic versus foreign. It is not clear why the derivative of equation (11) has the sign it does. The authors should clarify the assumptions required (note that p[I - p] is maximized when p == 1/2). Further, if the interest rate moves, what determines how much it moves? In its current state, their first case has too much in the background to provide the clarity it might have. I close with two perspective comments. First, the concept of a completely passive central bank in the face of a crisis makes it especially difficult to harmonize the two parts of the paper-empirical and theoretical. This contributes substantially to the uneasy marriage noted above. Finally, the authors' distaste for heterogeneous expectations motivates them to look toward heterogeneity in risk aversion, liquidity preference, and asset preference. I do not share their distaste: an increasing body of empirical work supports the presence of heterogeneity in expectations. They describe their mechanism as "more readily interpretable." But how much can be said empirically about heterogeneity in risk aversion or liquidity preference? Is heterogeneous asset preference truly in the set of deep parameters, or is it something to be derived? How much do we learn about differential position taking from a model that starts with differential asset preference?
8
Exchange Rate Economics: What's Wrong with the Conventional Macro Approach? Robert P. Flood and Mark P. Taylor
To include a paper entitled as this one is in a volume devoted to an examination of the importance of market microstructure in foreign exchange markets is, to utilize a well-used phrase, to preach to the converted. The poor empirical performance of the major exchange rate models over the recent float is, moreover, extremely well documented (Frankel and Rose 1994; MacDonald and Taylor 1992; Taylor 1994). Nevertheless, it is important to have a formal statement of the theory and evidence relating to exchange rate models based on macroeconomic fundamentals as a ground-clearing exercise since only by stating what is wrong with the conventional macro approach can we hope to design models that fill the gaps left by the macro-based models. Thus, section 8.1 of this paper is devoted to a discussion of the theory and empirical evidence relating to the major macroeconomic exchange rate models developed during the last twenty years or so, including the flexible-price monetary model, the stickyprice, overshooting monetary model, the portfolio balance model, and the equilibrium model. In section 8.2, we provide a brief discussion of the theory and evidence relating to the speculative efficiency of foreign exchange markets. Beyond this, however, we want to demonstrate that, while the macro fundamentals are clearly not capable of explaining all-or even most-of the variation in short-term nominal exchange rate movements, the research program of the last twenty years has nevertheless not been entirely fruitless. Our aim is thus to examine the macro fundamentals as a means of "setting the parameters" within which microstructural models might be constructed. Thus, in section Robert P. Flood is a senior economist in the Research Department of the International Monetary Fund. Mark P. Taylor is professor of economics at the University of Liverpool. The authors are grateful for comments on earlier versions of the paper from Jeffrey Frankel, Andrew Rose, Lars Svensson, and other conference participants. The views represented in the paper are those of the authors and are not necessarily those of the International Monetary Fund or of its member authorities.
261
262
Robert P. Flood and Mark P. Taylor
8.3, we invert the question posed in the title of this paper and ask, What's right with the conventional macro approach? Using data on twenty-one industrialized countries for the floating-rate period, we show that, while the macro fundamentals may be a poor guide to variations in short-run exchange rate movements (where the short run is defined as one year or less), they may nevertheless have considerable explanatory power over longer horizons. A final section concludes the discussion and tries to give an answer to the question that the title of this paper poses.
8.1
Theory and Evidence on Exchange Rate Models Based on Macro Fundamentals
In this section, we review briefly the theory and evidence pertaining to the four major exchange rate models based on conventional macro fundamentals: the monetary model, the sticky-price monetary model, the equilibrium model, and the portfolio balance model. 1 The monetary model is the simplest of the four and assumes that all goods are perfect substitutes, as are all interestbearing assets, and that all markets clear continuously. The other three models relax, in various ways, some of the strong assumptions made in the monetary model and in some cases make explicit previously unarticulated assumptions. The sticky-price model makes two big changes from the monetary model: it adds multiple goods and allows slow adjustment of nominal goods prices. The equilibrium model also allows multiple goods, but it models asset preferences as depending on the covariation of real asset returns with the marginal utility of consumption for some assets and as determined by unmodeled constraints for other assets. Also, by paying explicit attention to individual and economywide constraints, the equilibrium model is intended to clarify the full effects of various policy options. Typically, equilibrium models require continuous goods market and asset market clearing. Portfolio balance models are distinguished by their preferred specification of asset demands and are eclectic with respect to goods market specifications. In portfolio balance models, different interest-bearing assets are not perfect substitutes so that uncovered interest rate parity does not hold and asset demands may be modeled along the lines suggested by Tobin (1969). Evidently, these four classes of model are not mutually exclusive. They share many common structural elements, including the property that expectations of the future are potentially crucial for current decisions, and, more important, they all share the property that current and expected future macro fundamentals are always at the heart of exchange rate determination. Prior to setting out the theory and evidence relating to these four models in detail, however, we consider what is probably the simplest and in many ways 1. Sections 8.1 and 8.2 draw on MacDonald and Taylor (1992) and Taylor (1994).
263
Exchange Rate Economics: What's Wrong with the Macro Approach?
the most fundamental link between the exchange rate and macroeconomic fundamentals: purchasing power parity. 8.1.1
Purchasing Power Parity
Purchasing power parity (PPP) is one of the simplest macro fundamental exchange rate models that one can imagine. Absolute purchasing power parity implies that the exchange rate is equal to the ratio of the two relevant national price levels; relative purchasing power parity posits that changes in the exchange rate are equal to changes in relative national prices: (1)
and (2)
where St denotes the logarithm of the spot exchange rate (domestic price of foreign currency), and Pt and P; denote the logarithms of suitably normalized national price levels for the domestic and foreign economies, respectively. The deviation from purchasing power parity is commonly referred to as the real exchange rate, defined here in logarithmic form: (3)
The professional consensus on the validity of purchasing power parity has shifted radically over the past two decades or so. Prior to the recent float, the consensus appeared to support the existence of a fairly stable real exchange rate-that is to say, the variance of ~t or ~~ was thought of as small relative to the variance in relative prices or relative inflation rates. The prevailing orthodoxy of the early 1970s, largely associated with the monetary approach to the exchange rate, assumed the much stronger proposition of continuous purchasing power parity-that is, that the variance of ~t or ~~ was identically equal to zero (see, e.g., Frenkel 1976; and Frenkel and Johnson 1978). Proponents of early monetary exchange rate models, moreover, argued that, while the exchange rate may apparently diverge from PPP when conventional price indices are used, the condition would be seen to hold if one could observe the "true" price indices that are relevant for deflating national monies, so that observed variation in ~t or ~~ was really due to variation in measurement errors. In the mid- to late 1970s, in the light of the very high variability of real exchange rates, this extreme position was largely abandoned as the variability in observed deviations from PPP became so large that it became clear that they could not be due to measurement errors alone. Subsequently, studies published mostly in the 1980s, which could not reject the hypothesis of random-walk behavior in real exchange rates-that ~t or 'TT t followed a random walk (e.g., Adler and Lehmann 1983)-reduced further the confidence in purchasing
264
Robert P. Flood and Mark P. Taylor
power parity and led to the rather widespread belief that PPP was of little use empirically and that real exchange rate movements were highly persistent. More recently, in an extension of this literature, researchers have tested for more general mean reversion or stationarity of the real exchange (where the alternative hypothesis is a more general unit root process rather than specifically a random walk) have interpreted the null hypothesis of stationarity as equivalent to the existence of long-run purchasing power parity. Relatedly, researchers have also allowed the slope coefficients on domestic and foreign prices to differ from unity by testing for cointegration between the nominal exchange rate and domestic and foreign prices. Early cointegration studies generally reported a failure of significant mean reversion of the exchange rate toward purchasing power parity for the recent floating experience (Taylor 1988; Mark 1990) but were supportive of reversion toward purchasing power parity notably for the interwar float (Taylor and McMahon 1988) and for the exchange rates of high-inflation countries (McNown and Wallace 1989). Very recent applied work on long-run purchasing power parity among the major industrialized economies has, however, been more favorable toward the longrun purchasing power parity hypothesis for the recent float (e.g., Cheung and Lai 1993; Lothian and Taylor 1994). A number of authors have argued that the data period for the recent float alone may simply be too short to provide any reasonable degree of test power in the normal statistical tests for stationarity of the real exchange rate (e.g., Frankel 1990). 8.1.2
The Flexible-Price Monetary Model
The monetary approach to the exchange rate, which emerged as the dominant exchange rate model at the start of the recent float in the early 1970s (e.g., Frenkel 1976), starts from the definition of the exchange rate as the relative price of two monies and attempts to model that relative price in terms of the relative supply of and demand for those monies. Assuming stable, log-linear money demand functions at home and abroad (all variables except interest rates expressed in logarithms), the demand for money, m, is assumed to depend linearly on real income, y, the price level, p, and the level of the nominal interest rate, i (foreign variables are denoted by an asterisk). Assuming continuous purchasing power parity (~t == 0 in [1]), and substituting out for relative prices, it is straightforward to derive the fundamental equation of the flexible-price monetary model:
(4) where K and e are the income elasticity and interest rate semielasticity of the demand for money (here assumed equal at home and abroad for expository purposes). By invoking the uncovered interest parity condition, we can substitute LlS~+l for (it - i;) in order to emphasize the forward-looking nature of the model:
265
Exchange Rate Economics: What's Wrong with the Macro Approach?
(5)
St
==
where
The rational expectations solution to (5) is (7)
S,
= (1
+ e)-If. (_e_)i E[
1
+
e
where E[· I il t ] denotes the mathematical expectation conditioned on the information set available at time t, il t •2 It is well known from the rational expectations literature, however, that equation (7) is only one solution to (5) from a potentially infinite set involving rational bubbles. Note that the flexible-price monetary model is really just a purchasing power parity model of the exchange rate, where the proximate force driving relative prices is assumed to be relative excess demand for money. The very high volatility of real exchange rates during the 1970s float, conspicuously refuting the assumption of continuous purchasing power parity, led to the development of two further models, the sticky-price monetary model and the so-called equilibrium model. Both of these can be viewed as extensions or modifications in some way of the flexible-price monetary model. Before examining the empirical evidence on the flexible-price monetary model, we give a brief exposition of the sticky-price monetary model. 8.1.3
The Sticky-Price Monetary Model
Sticky-price monetary models, due originally to Dornbusch (1976), allow short-term overshooting of the nominal and real exchange rates above their long-run equilibrium levels. This results as the "jump variables" in the system (exchange rates and interest rates) compensate for stickiness in other variables-notably goods prices. The essential characteristics of the sticky-price monetary model can be seen in a three-equation structural model in continuous time, holding foreign variables and domestic income constant (these are simplifying rather than necessary assumptions): (8)
(9) (10)
s == m == p
p == -y[a +
i - i*,
+ Ky -
J.L(S - p) -
ei,
\iii -
y].
Equation (8) is the uncovered interest parity condition expressed in continuous time and utilizing certainty equivalence because of the linearity of the model. 2. The other three models that we will present below may similarly be represented in terms of the entire expected future. We present that solution only for the model at hand and mention the succeeding models' deviations from the monetary model.
266
Robert P. Flood and Mark P. Taylor
Equation (9) is a domestic money-market equilibrium condition, and equation (10) is a Phillips curve relation, relating domestic price movements to excess aggregate demand, where aggregate demand has an autonomous component, a component depending on international competitiveness, and a component that is interest rate sensitive. 3 If we use a bar to denote a variable in long-run (noninflationary) equilibrium, we can reduce this system to a two-equation differential equation system: 4 (11)
0 1/8] [Ps] = ['Yf-L -"1 (f-L + '1'/8)
[s - Ps] . p -
The coefficient matrix in (11) has a negative determinant, so the system has a unique convergent saddle path. The qualitative solution to (11) is shown in figure 8.1, where the saddle path slopes down left to right in (s, p )-space. Monetary shocks will lead to overshooting in the model as the long-run equilibrium relative price level-and hence the saddle path-shifts, causing a discontinuous shift in the exchange rate onto the new saddle path, with prices initially constant. This is then followed by a slower movement of prices and the exchange rate toward the new equilibrium level. Now consider the effects of a real shock to tastes for the domestic good as opposed to the foreign good. Say, for example, that there is a permanent shift in demand toward the home good, which would be represented by an increase in cx. In terms of figure 8.1, the effect of the shift is to displace the equilibrium horizontally with the exchange rate fully and immediately making the adjustment to bring relative international prices into equilibrium. Unless the demand shift decays, there is no tendency for this disturbance's real effects to decrease over time. A useful way to rewrite the model in discrete time is (12)
where
and where rt is the real interest rate and Zt reflects real goods market influences on the nominal exchange rate for this case. It is immediately clear that shocks 3. For the sake of brevity, we are not distinguishing very carefully between the domestic price level, which includes domestic currency prices of imported goods, and the domestic currency price of domestic goods. If imported goods prices become sticky once they have been domestically priced, then the distinction is unimportant. As in the Dornbusch (1976) presentation, we allow goods demand to depend apparently on the nominal rather than the real interest rate. We think of such a presentation in Dornbusch's terms as a semi-reduced form in which the inflation terms have been aggregated on the equation's left-hand side. Such complications have been clarified in the literature (Flood 1981), but none of them alter very much the basic results of the Dornbusch model. 4. Note that the level of the money stock is exogenous to the model, assumed under the control of the authorities. Thus, for any given level of the money stock, the perfect foresight equilibrium involves assuming m = Iii.
267
Exchange Rate Economics: What's Wrong with the Macro Approach?
p
p=o
L p I-------+-------:wr------+------ ; = 0
Fig.8.1
The saddle path for the sticky-price monetary model
and influence Zt make the sticky-price monetary model's predictions diverge from those of the flexible-price monetary model, at least in the short run. To the extent that shocks to Zt are transitory, however, the differences between the monetary and the sticky-price models are transitory also. 8.1.4
Empirical Evidence on the Flexible-Price and Sticky-Price Monetary Models
Initial support for the flexible-price monetary model was provided by Frenkel (1976), who utilized data for the German marklU.S. dOlIar exchange rate during the German hyperinflation of the 1920s. The subsequent accumulation of data for the 1970s float allowed estimation of the model for the major exchange rates during the recent float, and initial studies were also broadly supportive of the flexible-price monetary model (e.g., Bilson 1978; and Dornbusch 1976). Beyond the late 1970s, however, the flexible-price monetary model (or its real interest differential variant) ceases to provide a good explanation of variations in exchange rate data: the estimated equations break down, providing poor fits, exhibiting incorrectly signed coefficients, and failing general 'equation diagnostics (e.g., Frankel 1993). The evidence for the sticky-price monetary model is also weak when the data period is extended beyond the late 1970s (Backus 1984). Another implication of the sticky-price monetary model is proportional variation between the real exchange rate and the real interest rate differential. This follows from the basic assumptions of the overshooting model: slowly adjusting prices and uncovered interest rate parity. A number of studies have failed to find strong evidence of this relation, notably Meese and Rogoff (1988), who could not find cointegration between real exchange rates and real interest rate differentials.
268
Robert P. Flood and Mark P. Taylor
More recently, MacDonald and Taylor (1993, 1994) apply multivariate cointegration analysis and dynamic modeling techniques to a number of exchange rates and find some evidence to support the monetary model as a long-run equilibrium toward which the exchange rate converges, while allowing for complicated short-run dynamics. Since all the monetary models collapse to an equilibrium condition of the form (6) in the long run, these tests have no power to discriminate between the alternative varieties. The usefulness of the cointegration approach suggested by these studies should, moreover, be taken as at most tentative: their robustness across different data periods and exchange rates has yet to be demonstrated. 8.1.5
The Portfolio Balance Model
The key distinguishing feature of the portfolio balance model is the assumed imperfect substitutability between domestic and foreign assets. 5 Consider a simple model in which the net financial wealth of the private sector (W) is divided into three components: money (M), domestically issued bonds (B), and foreign bonds denominated in foreign currency and held by domestic residents (B*). B can be thought of as a government debt held by the domestic private sector; B* is the level of net claims on foreigners held by the private sector. Since, under a free float, a current account surplus on the balance of payments must be exactly matched by a capital account deficit (i.e., capital outflow and hence an increase in net foreign indebtedness to the domestic economy), the current account must give the rate of accumulation of B* over time. With foreign and domestic interest rates given by i and i* as before, we can write down our definition of wealth and simple domestic demand functions for its components as follows: 6 (14) (15) (16) (17) (18)
W=M+ B
+ SB*,
se)w, M < 0, M < 0, B = B(i, i* + se)w, B > 0, B < 0, SB* = B*(i, i* + se)w, B < 0, B > 0, M
= M(i,
i*· +
1
2
1
2
1
B* = T(S/P)
2
+ i*B*, T 1 > 0,
where se denotes the expected rate of depreciation of the domestic currency. Relation (14) is an identity defining wealth and (15), (16), and (17) are standard asset demand functions. 7 Equation (18) gives the rate of change of B*, the 5. A comprehensive treatment of the portfolio balance model is given in Branson and Henderson (1985). 6. X k denotes the partial derivative of X(·) with respect to the kth argument. The shift to uppercase letters here indicates that variables are in levels rather than logarithms. As throughout, interest rates are in percentage terms. 7. Note that, as is standard in most expositions of the portfolio balance model, the scale variable is the level of wealth, W, and the demand functions are homogeneous in wealth; this allows them to be written in nominal terms (assuming homogeneity in prices and real wealth, prices cancel out).
269
Exchange Rate Economics: What's Wrong with the Macro Approach?
capital account, as equal to the current account, which is in turn equal to the sum of the trade balance, T(·), and net debt service receipts, i*B*. The trade balance depends positively on the level of the real exchange rate (a devaluation improves the trade balance). The exchange rate is then determined by solving equations (14)-( 18) for given levels of M, B, and B*, normally assuming rational expectations. Disturbances to these stocks will result in movements in S in both the short run (solve [14]-[18] allowing the left-hand side of [18] to be nonzero) and the long run (impose the constraint that all asset levels are constant). More structure can be put on the model by assuming that the asset demand functions are determined by agents optimizing a function of the mean and variance of their end-of-period wealth. 8.1.6
Empirical Evidence on the Portfolio Balance Model
Log-linear versions of reduced-form portfolio balance exchange rate equations, using cumulated current accounts for the stock of foreign assets, have, however, been estimated for many of the major exchange rates for the 1970s float, with poor results: estimated coefficients are often insignificant, and there is a persistent problem of residual autocorrelation (e.g., Branson, Halttunen, and Masson 1977; Frankel 1993; see also Lewis 1988). The imperfect substitutability of domestic and foreign assets that is assumed in the portfolio balance model is equivalent to assuming that there is a risk premium separating expected depreciation and the domestic-foreign interest differential, and in the portfolio balance model this risk premium will be a function of relative domestic and foreign debt outstanding. An alternative, indirect method of testing the portfolio balance model, therefore, is to test for empirical relations of this kind. Investigations of this kind have usually reported statistically insignificant relations (see Frankel 1982; and Rogoff 1984). In a recent study of the effectiveness of exchange rate intervention for dollar/mark and dollar/Swiss franc during the 1980s, Dominguez and Frankel (1993) measure the risk premium using survey data and show that the resulting measure can in fact be explained by an empirical model that is consistent with the portfolio balance model, with the additional assumption of mean-variance optimization on the part of investors. In some ways, the relative success of the Dominguez and Frankel (1993) study is consistent with the recent empirical literature on foreign exchange market efficiency, .discussed below, which suggests the existence of significant foreign exchange risk premia and nonrational expectations. 8.1.7
The Equilibrium Model
Equilibrium exchange rate models of the type developed originally by Stockman (1980) and Lucas (1982) analyze the general equilibrium of a twocountry model by maximizing the expected present value of a representative agent's utility, subject to budget constraints and cash-in-advance constraints (by convention, agents are required to hold local currency, the accepted me-
270
Robert P. Flood and Mark P. Taylor
domestic goods
yin
y*/n
foreign goods
Fig. 8.2 A simple equilibrium exchange rate model
dium of exchange, with which to purchase goods).8 In an important sense, equilibrium models are an extension or a generalization of the flexible-price monetary model to allow for multiple traded goods and real shocks across countries. A simple equilibrium model can be sketched as follows. Consider a twocountry, two-good world in which prices are flexible and markets are in equilibrium, as In the flexible-price monetary model, but in which, in contrast to the monetary model, agents distinguish between domestic and foreign goods in terms of well-defined preferences. Further, for simplicity, assume that all agents, domestic or foreign, have identical preferences. Then, given domestic and foreign output of y and y*, respectively, the equilibrium relative price of foreign output, say II, must be the slope of a representative agent's indifference ~urve at the point (y*ln, yin) in foreign-domestic output per capita space (where nl2 is the number of individuals in each economy), as in figure 8.2. But the relative price of foreign output is the real exchange rate, which is defined in logarithmic form (11") by (3). Now consider log-linear domestic and foreign money demand functions for the representative agent: (19) (20) 8. For a more extensive, largely nontechnical exposition, see Stockman (1987). This literature is an offshoot of the real business cycle literature.
271
Exchange Rate Economics: What's Wrong with the Macro Approach?
domestic goods
y*/n
foreign goods
Fig. 8.3 The real exchange rate effect of a preference shift toward domestic goods in the equilibrium model
which, in an optimizing framework, can be interpreted as linearizations of expressions derived from maximizing the representative agent's utility function subject to a cash-in-advance constraint, assuming that government policy (and other influences on the constancy of parameters) remains constant. Combining these with the definition of the real exchange rate, (3), and solving for the nominal exchange rate,we can derive (21) Equation (21) represents a very simple formulation of nominal exchange rate determination in the equilibrium model. At first sight, it appears to be a simple modification of the monetary model. Indeed, relative monetary expansion leads to a depreciation of the domestic currency as in the simple monetary model. However, as an example of a situation that it would be impossible to analyze in a flexible-price monetary model, consider an exogenous shift in preferences away from foreign goods toward domestic goods, represented as a flattening of indifference curves as in figure 8.3 (from II to 12 ), With per capita outputs fixed, this implies a fall in the relative price of foreign output (or, conversely, a rise in the relative price of domestic output)- II falls (from III to Il 2 in fig. 8.3).9 Assuming unchanged monetary policies, this movement in the 9. Note that uppercase pi denotes the real exchange rate, lowercase pi the logarithm of the real exchange rate.
272
Robert P. Flood and Mark P. Taylor
real exchange rate will, however, be brought about entirely (and swiftly) by a movement in the nominal exchange rate without any movement in national price levels. Thus, demand shifts are capable of explaining the observed volatility of nominal exchange rates in excess of volatility in relative prices in equilibrium models. The fall in s, in this case matching the fall in 1T, will be observed as a decline in domestic competitiveness. 10 8.1.8
Empirical Evidence on the Equilibrium Model
Although there is, in principle, no reason why a linearized version of the equilibrium model should not be estimated, advocates of this approach have preferred to point to the "consistency" of the model with the observed behavior of exchange rates. Well-known, stylized facts of the recent float include the high volatility and correlation of real and nominal exchange rates and the absence of strong mean-reverting properties in either series. As we noted above, equilibrium models are capable of explaining the variability of nominal exchange rates in excess of relative price variability (and hence the variability of real exchange rates), but so is the sticky-price monetary model. Some authors have argued, however, that the difficulty that researchers have experienced in rejecting the hypothesis of nonstationarity in the real exchange rate is evidence against the sticky-price model and in favor of equilibrium models since the former class of models requires some sort of long-run convergence of the real exchange rate toward PPP, while an equilibrium model characterized by random-walk innovations to taste and technology would generate a nonstationary real exchange rate. Explaining the persistence in both real and nominal exchange rates over the recent float within the framework of the sticky-price model, it is argued, involves assuming either implausibly sluggish price adjustment or else that movements in nominal exchange rates are due largely to permanent real disturbances (e.g., Stockman 1987). This line of argument overlooks, however, the fact that relative shocks to tastes and technology between countries are more likely to be mean reverting (e.g., because of technology transfer). Moreover, as Frankel (1990) argues forcibly, noncontradiction is not the same as confirmation: simply being consistent with the facts is not enough to demonstrate the empirical validity of a theory. One testable implication of the simplest equilibrium models is the neutrality of the exchange rate with respect to the exchange rate regime: since the real exchange rate is determined by real variables such as tastes and technology, its 10. In the simple equilibrium model that we have sketched here, we have implicitly made a host of simplifying assumptions. Chief among these is the assumption that individuals in either economy hold exactly the same fractions of their wealth in any firm, domestic or foreign. If this assumption is violated, then supply and demand shifts will alter the relative distribution of wealth between domestic and foreign residents as, e.g., one country becomes relatively more productive. This, in turn, will affect the equilibrium level of the exchange rate (Stockman 1987).
273
Exchange Rate Economics: What's Wrong with the Macro Approach?
behavior ought to be independent of whether the nominal exchange rate is pegged or allowed to float freely. Using data for a large number of countries and time periods, however, researchers have invariably found that real exchange rates are significantly more volatile under floating nominal rate regimes (Stockman 1983; Mussa 1986; Baxter and Stockman 1989). Although this evidence does, indeed, constitute a rejection of the simplest equilibrium models, it is possible that the evidence is to some extent confounded by the endogeneity of the choice of exchange rate regime-that is, countries experiencing greater real disturbances are more likely to choose flexible exchange rate systems. Moreover, Stockman (1983) also shows that the assumptions necessary for regime neutrality are in fact quite restrictive in a fully specified equilibrium model and include Ricardian equivalence, no wealth-distribution effects of nominal price changes, no real effects of inflation, no real effects of changes in the level of the money supply, complete asset markets, completely flexible prices, and identical sets of government policies under different exchange rate systems. Since it is unlikely that all these conditions will be met in practice, Stockman argues that only the simplest class of equilibrium models should be rejected and that equilibrium models should be developed that relax some or all of these assumptions. Moreover, Stockman (1988) argues that, because of the increased likelihood of countries with fixed exchange rates introducing controls on trade or capital flows, a disturbance that would tend to raise the relative price of foreign goods (e.g., a preference shift toward foreign goods) will raise the probability that the domestic country wil~, at some future point, impose capital or trade restrictions that will raise tfie future relative world price of domestic goods. With intertemporal substitution, this induces a higher world demand for domestic goods now, serving to offset partly the direct effect of the disturbance, which was to raise the relative price of the foreign good, and hence to reduce the resulting movement in the real exchange rate. Thus, countries with pegged exchange rates will experience lower volatility in the real exchange rate than countries with flexible exchange rates. This discussion makes clear that the equilibrium model is not so much a model as a way of viewing the world in strictly equilibrium terms. In particular, it is not clear exactly what the proponents of this approach would accept as a decisive rejection of the model. 8.1.9
Forecasting with Macro-Based Exchange Rate Models
In a landmark paper, Meese and Rogoff (1983a) compare the out-of-sample forecasts produced by various macro-based exchange rate models with forecasts produced by a random-walk model, by the forward exchange rate, by a univariate regression of the spot rate, and by a vector autoregression. They use rolling regressions to generate a succession of out-of-sample forecasts for each model and for various time horizons. The conclusion that emerges from this
274
Robert P. Flood and Mark P. Taylor
study is that, on a comparison of root mean square errors (RMSEs), none of the asset market exchange rate models outperforms the simple random walk. 11 Further work by the same authors (Meese and Rogoff 1983b) suggested that the estimated models may have been affected by simultaneity bias. Imposing coefficient constraints taken from the empirical literature on money demand, Meese and Rogoff find that, although the coefficient-constrained asset reduced forms still fail to outperform the random-walk model for most horizons up to a year, combinations of parameter constraints can be found such that the models do outperform the random-walk model for horizons beyond twelve months. Even at these longer horizons, however, the models are unstable in the sense that the minimum RMSE models have different coefficient values at different horizons. Although beating the random walk still remains the standard metric in which to judge empirical exchange rate models, researchers have found that one key to improving forecast performance based on economic fundamentals lies in the introduction of equation dynamics. This has been done in various ways: by using dynamic forecasting equations for the forcing variables in the forwardlooking, rational expectations version of the flexible-price monetary model, by incorporating dynamic partial adjustment terms into the estimating equation, by using time-varying parameter estimation techniques, and-most recentlyby using dynamic error correction forms (Throop 1993; MacDonald and Taylor 1993, 1994; Mark 1992).
8.2 Speculative Efficiency In tandem with work on macro-based exchange rate models, there has developed a whole body of literature on the speculative efficiency of foreign exchange markets. This literature is important in the present context for at least three reasons. First, efficiency conditions such as uncovered interest parity are often used as building blocks in constructing macro-based exchange rate models. Second, the empirical literature of the efficiency of the foreign exchange market has thrown up a number of stylized facts (indeed, anomalies) that provide a challenge for models of foreign exchange market microstructure to explain. Third, and most important, a standard route by which market microstructure has traditionally been dismissed is via the assumption of market efficiency. If, indeed, smart speculators always tend to dominate in the foreign exchange market, as Friedman's classic apologia of floating exchange rates argues (Friedman 1953), then the aggregate behavior of foreign exchange markets can in fact be summarized by a handful of parity conditions that characterize the mar11. Meese and Rogoff (1983a) compare random-walk forecasts with those produced by the flexible-price monetary model, Frankel's (1979) real interest rate differential variant of the monetary model, and a synthesis of the monetary and portfolio balance models suggested by Hooper and Morton (1982).
275
Exchange Rate Economics: What's Wrong with the Macro Approach?
ket equilibrium. For t1?-e purposes of market efficiency, the most important parity condition that researchers have considered is that of uncovered interest rate parity. If the risk-neutral efficient markets hypothesis holds, then the expected foreign exchange gain from holding one currency rather than another (the expected exchange rate change) must be just offset by the opportunity cost of holding funds in this currency rather than the other (the interest rate differential). This is the uncovered interest rate parity condition: (22) Researchers have most often tested uncovered interest rate parity by regression-based analyses of spot and forward exchange rates. Assuming covered interest parity, the interest rate differential should be just equal to the forward premium. Under rational expectations, the expected change in the exchange rate should differ from the actual change only by a rational expectations forecast error. Hence, the uncovered interest rate parity condition (26) can be tested by estimating a regression equation of the form (23) where f~k) is the logarithm of the forward rate for maturity k periods ahead. .If agents are risk neutral and have rational expectations, we should expect the slope parameter, (3, to be equal to one and the disturbance term l1t+k-the rational expectations forecast error under the null hypothesis-to be uncorrelated with information available at time t. Empirical studies of (23), for a large variety of currencies and time periods, for the recent floating experience generally report results that are unfavorable to the efficient markets hypothesis under risk neutrality (e.g., Fama 1984). Indeed, it is a stylized fact that estimates of (3, using exchange rates against the dollar, are generally closer to minus unity than plus unity (Froot and Thaler 1990).12 The rejection of the simple, risk-neutral efficient markets hypothesis may be due to the risk aversion of market participants or to a departure from the pure rational expectations hypothesis, or both. If foreign exchange market participants are risk averse, the uncovered interest parity condition (22) may be distorted by a risk premium. If the risk premium is time varying and correlated with the forward premium or the interest rate differential, this would confound efficiency tests of the kind outlined above (Fama 1984). This reasoning has led to a search for stable empirical models of the risk premium on the assumption of rational expectations. Because of the theoretical relation between risk and the second moments of asset price distributions, researchers have often tested 12. Alternatively, researchers have imposed J3 = 1 in equations such as (23) and tested the orthogonality of the error term, 'llt+k' with respect to information available at time t. Such tests have also generally rejected the simple speculative efficiency hypothesis (see, e.g., Hodrick 1987).
276
Robert P. Flood and Mark P. Taylor
for a risk premium as a function of the variance of forecast errors or of exchange rate movements (Frankel 1982; Domowitz and Hakkio 1985; Giovannini and Jorion 1989). In common with other empirical risk premium models, such as latent variables formulations (Hansen and Hodrick 1983), such models have generally met with mixed and somewhat limited success and have not been found to be robust when applied to different data sets and sample periods. In sum, it appears that, for credible degrees of risk aversion, empirical risk premium models have so far been unable to explain to any significant degree the variation in the excess return from forward market speculation. An alternative explanation of the rejection of the simple efficient markets hypothesis is that there is a failure, in some sense, of the expectations component of the joint hypothesis. Examples in this group are the "peso problem" suggested by Kenneth Rogoff (1979), rational bubbles, learning about regime shifts (Lewis 1989), or inefficient information processing (as suggested, e.g., by John Bilson [1981]). The peso problem refers to the situation where agents attach a small probability to a large change in the economic fundamentals, which does not occur in sample. This will tend to produce a skew in the distribution of forecast errors even if agents' expectations are rational and thus may generate apparent evidence of nonzero excess returns from forward speculation. In common with peso problems, the presence of rational bubbles may also show up as nonzero excess returns even when agents are risk neutral. Similarly, when agents are learning about their environment, they may be unable fully to exploit arbitrage opportunities that are apparent in the data ex post. A problem with admitting peso problems, bubbles, or learning into the class of explanations of the forward discount bias is that, as noted above, a very large number of econometric studies-encompassing an even larger range of exchange rates and sample periods-have found that the direction of bias is the same, that is, that the estimated uncovered interest rate parity slope parameter, r3 in (23), is generally negative and closer to minus unity than plus unity. A problem with much of the empirical work on the possible rationalizations of the rejection of the simple, risk-neutral efficient markets hypothesis is that, in testing one leg of the joint hypothesis, researchers have typically assumed that the other leg is true. For instance, the search for a stable empirical risk premium model has generally been conditioned on the assumption of rational expectations. Thus, some researchers have employed survey data on exchange rate expectations to conduct tests of each component of the joint hypothesis (Froot and Frankel 1989). In general, the overall conclusion that emerges from survey data studies appears to be that both risk aversion and departures from rational expectations are responsible for rejection of the simple efficient markets hypothesis. 13 13. For surveys of the literature on foreign exchange risk premia, see Frankel (1988) and Lewis (1994).
277
8.3
Exchange Rate Economics: What's Wrong with the Macro Approach?
What's Right with the Conventional Macro Approach?
In this section, we present some further empirical evidence on the relation between exchange rate movements and macroeconomic fundamentals by examining the purchasing power parity and uncovered interest rate parity conditions' using panel data for twenty-two countries. 14 8.3.1
Purchasing Power Parity
Consider again the simple purchasing power parity equations (1) and (2), reproduced here with time-series disturbances ~t and e t and slope parameters J3 and "/:
= J3(Pt - p;) + ~t' Llst = ~(LlPt - Llp;) + ~~.
(24)
St
(25)
Estimation of these relations (with a constant term added) for dollar/sterling and dollar/mark, using annual data for the period 1973-92, yields: 15 dollar/sterling
(26) (27)
St = 0.671 + 0.734(Pt - p;) + ~t' R2 = 0.31, DW = 0.51, ADF = 2.25; Llst =
-0.039 - 0.235(LlPt - Llp;) (0.044) (1.266) R2 = 0.001, DW = 1.29;
+ ~~,
dollar/mark
(28) (29)
St = -0.535 + 0.606(Pt - p;) + ~t' R2 = 0.30, DW = 0.85, ADF = 1.13; Llst =
0.009 + 0.408(LlPt - Llp;) + ~~, (0.046) (0.861) R2 = 0.008, DW = 1.89.
14. We are, of course, not the first researchers to examine the PPP relation using panel data. Officer (1980), e.g., finds a broadly proportional relation between the rate of depreciation and relative inflation for a similar group of countries to those examined below, for the period 1913-75. MacDonald (1988) uses a variety of estimation techniques on pooled annual time-series crosssectional data for the G-5 countries plus Switzerland and provides evidence supportive of relative PPP. 15. The data used in this section are annual observations taken from the International Monetary Fund's International Financial Statistics CD-rom database: exchange rates are usually line ae, prices are CPI (line 64), output is real GDP (line 99b. r), and interest rates are mainly three-month Treasury bill rates (line 60c). Figures in parentheses are heteroskedastic-consistent standard errors, R2 denotes the coefficient of determination, DW the Durbin-Watson statistic, and ADF the augmented Dickey-Fuller statistic applied to the estimated residuals. The individual series were found to be approximately I( 1), so the estimated standard errors are not reported for the levels regressions.
278
Robert P. Flood and Mark P. Taylor
80 70
o
60
o
50 40
o o
0
o
0
0
30 20 10
o
00
o
-10 -20 - 30 ~-.,.....,.--.---,.-----.--...-----r---r---r--.---,.---+-30 -20 -10 0 10 20 30 40 50 60 70 80
Fig. 8.4 Scatter plot of annual exchange rate change against annual inflation differential
These equations are typical of single-equation results reported in the literature for the recent floating-rate period: there is no apparent sign of cointegration of the exchange rate and relative prices, and relative inflation explains less than 1 percent of the time-series variation in the nominal exchange rate. In figure 8.4, we have plotted the annual exchange rate change against the annual inflation differential for twenty-one industrialized countries against the United States for the period 1973-92. 16 This scatter plot is vaguely suggestive of a positive linear relation between the rate of depreciation and the inflation differential, especially for large inflation differentials. In figure 8.5, we also plot the annual rate of depreciation against the United States, but this time using. five-year averages (four for each country, corresponding to the periods 1973-77, 1978-82, and 1983-87, and 1988-92). Figure 8.5 appears to reveal a stronger medium-term relation between the exchange rate and relative inflation, in the sense that the scatter is much closer to the forty-five-degree line, albeit with one or two outliers. When the data are averaged over periods of ten or twenty years (figs. 8.6 and 8.7, respectively), the proportionality between average relative inflation and average depreciation becomes even more marked, as the scatter more or less collapses onto the forty-five-degree ray. These visual impressions are largely confirmed by regression analysis, the full results of which are given in the appendix. By simply pooling the twenty years of data on annual changes, for example, we obtained the following pooled estimate: 16. The countries besides the United States included in the sample are Australia, Austria, Belgium, Canada, Denmark, Finland, France, Germany, Greece, Iceland, Ireland, Italy, Japan, the Netherlands, New Zealand, Norway, Portugal, Spain, Sweden, Switzerland, and the United Kingdom. All data were taken from the IFS database.
80 70 60 50 0
40 30 0
20
0
0
10 0 -10
QJ
-20 - 30 -r--r----,---,----,---,--,....--.--,..---,...---r-_+_ -30 -20 -10 0 10 20 30 40 50 60 70 80
Fig. 8.5 Scatter plot of five-year average exchange rate change against five-year average inflation differential
80 70 60 50 40 0
30 20 10
0
00 0
0 -10 -20 - 30
.....j<'--,.------r---,------,.---.......--.-----...---.---....-----...--!-
-30 -20 -10
0
10
20
30
40
50
60
70
80
Fig. 8.6 Scatter plot of ten-year average exchange rate change against ten-year average inflation differential
80 70 60 50 40 30 20 10
o
o -10 -20 - 30 - r - r - - - - - r - , - - r - - - - , - - . . . , . - - - . , - - , - - - , - - . - - . - - 1 -30 -20 -10 0 10 20 30 40 50 60 70 80
Fig. 8.7 Scatter plot of twenty-year average exchange rate change against twenty-year average inflation differential
280 (30)
Robert P. Flood and Mark P. Taylor
0.022 -0.109(LlPt - Llp;) + ~;, (0.007) (0.108) R2 == 0.003, pooled data, annual changes, whole sample, N == 420. Llst ==
Clearly, there appears to be little gained from pooling the annual changes, with less than 1 percent of the pooled variation in the rate of depreciation being explained by relative inflation. However, in a pooled regression consisting of the four five-year average annual rates of depreciation for each country against the corresponding five-year average annual inflation differential, we obtained 0.356 + 0.807 (LlPt - Llp;) + ~;, (0.742) (0.159) R2 == 0.432, pooled data, five-year-average changes, whole sample, N == 84.
(31)
Llst ==
Equation (31) explains over 40 percent of the pooled variation in the rate of depreciation, the slope coefficient is strongly significantly different from zero but insignificantly different from unity, and the intercept term is insignificantly different from zero. Moreover, ordering the data in increasing order of absolute magnitude of the difference in five-year average annual inflation rates, we obtained for the fourth quartile -0.764 + 0.960(LlPt - Llp;) + ~;, (2.077) (0.227) R2 == 0.695, pooled data, five-year-average changes, fourth quartile, N == 21.
(32)
Llst ==
Nearly 70 percent of the pooled variation in the rate of depreciation is explained, and the point estimate of the slope coefficient is very close to unity. The results for the fourth quartile using five-year averages are broadly echoed for the whole-sample estimates using ten- and twenty-year averages. For the ten-year averages, for example, we obtained (33)
R2 ==
0.165 + 0.967 (LlPt - Llp;) + ~;, (0.722) (0.129) 0.693, pooled data, ten-year-average changes, whole sample, N Llst ==
==
21.
The strength of the relative PPP relation in the time-averaged data is quite intriguing, and we provide an interpretation of these results below. Before doing so, however, we present the results of analyzing the uncovered interest parity relation using panel data. 17 8.3.2
Uncovered Interest Rate Parity
Although it is not unusual to use low-frequency data when examining purchasing power parity, researchers are usually much more fastidious in their treatment of data for efficiency conditions such as uncovered interest rate par17. Although, as noted above, we are not the first researchers to examine PPP using panel data, we are unaware of any previous study that applies time-averaged panel data to study the uncovered interest rate parity relation.
281
Exchange Rate Economics: What's Wrong with the Macro Approach?
ity and covered interest rate parity (see, e.g., Hodrick 1987; Taylor 1988, 1989). This is for good reasons-namely, that tiny mismatches or imperfections in the data may be mistaken for arbitrage opportunities that, although small in size, would nevertheless be highly profitable for a foreign exchange market participant moving around very large sums of money. In the present context, however, we are concerned with looking at uncovered interest rate parity in very broad terms, to see whether interest differentials have any relation with exchange rate movements. 18 Clearly, our results should be taken as indicative only. Again using annual data for the same group of twenty-one countries plus the United States for the floating-rate period, 1973-92, we constructed the annual interest rate differential, using three-year government bond rates, and the three-year rate of exchange rate depreciation against the U.S. dollar. 19 For the pooled data, we obtained the following results: (34)
=
2.599 +0.596(it - i;) + ilt' (1.656) (0.195) R2 = 0.066, pooled data, annual data, whole sample, N = 420, X2 (1) = 4.72, (0.038) Ll 3 st+ 3
where figures in parentheses below coefficient estimates are method-ofmoments corrected standard errors, and X2 ( 1) is a Wald test statistic for the null hypothesis that the slope coefficient is unity (figures below it in parentheses denote the marginal significance level). Equation (34) is typical in the sense that the R2 is low and the slope coefficient significantly different from unity at the 5 percent level. It is, however, untypical in that the slope coefficient is positive. To investigate the effects of the size of the interest differential on the performance of the uncovered interest parity condition, we ordered the data by the size of the absolute interest differential and reestimated the relation for each of the four quartiles. For the first quartile (i.e., smallest values of lit - i;l), we obtained (35)
4.159 - 2.999(it - i;) + ilt' (3.243) (1.343) R2 = 0.036, pooled data, first quartile, N = 105, X2 (1) = 8.851, (0.003) Ll 3st+ 3 =
18. The view that forward premia or interest differentials have little relation with exchange rate movements appears to be a widespread conclusion of the literature: "Forward premia contain little information regarding subsequent exchange rate changes. As emphasized by Dornbusch (1980), Mussa (1979), and Frenkel (1981), exchange rate changes over the recent period of floating seem to have been largely unanticipated" (Cumby and Obstfeld 1984, 139). 19. Because the interest rates are expressed in annualized terms, it was appropriate to multiply them by three in order to put them into three-year terms. We used ordinary least squares with a method-of-moments correction to the estimated covariance matrix to allow for heteroskedasticity and moving-average errors.
282
Robert P. Flood and Mark P. Taylor
which exhibits the characteristic negative slope coefficient and a lower R 2 • For the second quartile, the result was
R2
=
=
5.208 - 0.798(it - i;) + ilt (1.732) (0.475) 0.024, pooled data, second quartile, N = 105, X2(1) = 14.345, (0.000)
d 3st+ 3
(36)
which shows a very typical slope coefficient estimate that is closer to minus unity than plus unity and an even smaller R2. For the third quartile, we obtained
(37)
=
-1.405 + 0.609 (it - i;) + ilt' (2.841) (0.329) R2 = 0.046, pooled data, third quartile, N = 105, X2(1) = 1.409, (0.235)
d 3st+ 3
while, for the fourth quartile (Le., the largest values of lit estimate was
(38)
i;D, the resulting
11.121 + 0.520(it - i;) + ilt (7.460) (0.273) R2 = 0.079, pooled data, fourth quartile, N = 105, X2(1) = 3.087. (0.079)
Ll 3 st+ 3
=
It is interesting to note that, for the third and fourth quartiles, the slope coefficient is insignificantly different from unity at the 5 percent level. However, although there is some improvement in the goodness of fit, the R2'S are still quite low, and the slope coefficients are not in fact significantly different from zero at the 5 percent level. As in the analysis of PPP, the next step was to average the data temporally. Using pooled five-year averages for the data for the whole sample, we obtained the following estimate:
(39)
0.632 + 0.751(it - i;) + ilt' (2.020) (0.171) R2 = 0.190, pooled data, five-year averages, whole sample, N = 84, X2 (1) = 2.124. (0.145)
Ll 3st+ 3
=
Equation (39) is quite impressive in the sense that the goodness of fit has risen dramatically from the unaveraged cases-by a factor of twenty-five or more. Moreover, the slope coefficient is now strongly significantly different from zero and insignificantly different from zero at the 5 percent level. These effects with respect to the goodness of fit and the slope coefficient are even more
283
Exchange Rate Economics: What's Wrong with the Macro Approach?
marked when we consider the results of running the same regressions using data that have been time averaged over ten or twenty years: ~3St+3
(40)
0.706 + 1.075(it - i;) + ilt' (1.942) (0.111) R2 == 0.400, pooled data, ten-year averages, whole sample, N == 42, X2 (1) == 0.461; (0.497)
(41)
-2.721 + 1.481(it - i;) + ilt' (1.556) (0.293) R2 == 0.772, pooled data, twenty-year averages, whole sample, N == 42, X2 (1) == 2.697. (0.100) ~3St+3
==
==
Given the plethora of results on uncovered interest rate parity that have found little or no empirical connection between interest rates and the exchange rate or, if anything, have found negative covariation between the interest differential or forward premium and the rate of depreciation, these results are very striking and quite intriguing. Moreover, as we noted above, we have used a ready-made data set from the IFS (International Financial Statistics, International Monetary Fund) and have not spent a great deal of time worrying about aligning maturity dates, checking that the instruments are identical in all relevant respects across countries, and so on. These shortcomings in the data set would be most relevant if we did not find evidence of uncovered interest rate parity in the averaged data; as it is, they serve only to underscore the strength of the relation that we appear to be unearthing. 8.3.3
Interpretation
In this section, we have shown that very simple macro fundamentals-relative inflation or relative interest rates-have poor explanatory power with respect to variations in exchange rate movements even over the one-year horizon. Taking five-, ten-, and twenty-year averages, however, we found that a strong proportionality between average exchange rate depreciation and average movements in the fundamentals begins to emerge. The analysis is thus indicative of something that Rick and lIsa knew long ago: the fundamental things apply as time goes by. Moreover, the analysis also suggests that the variation of deviations from the fundamentals appears to be inversely related to the size of the movements in the underlying fundamentals. Our interpretation of these results is that, while the nominal exchange rate is extremely hard to distinguish from a random walk even at the one-year horizon, a simple macro fundamentals-based model outperforms the random walk at horizons of five years or longer. To see this clearly, note that the n-year average annual change in the exchange rate is defined as
284
Robert P. Flood and Mark P. Taylor n
(43)
(lin) L~St+i == (lln)(St+n - St)· i=O
Thus, finding that the n-year average exchange rate change is explained by macro fundamentals is equivalent to finding that the n-year-ahead randomwalk forecast is beaten in sample by a simple macro fundamentals model. The results also tell us something about the nature of deviations from the fundamentals. The change in the exchange rate can be defined as the sum of a component that is explained by movements in the macro fundamentals (such as relative inflation or relative nominal interest rates, or whatever macro variables that, in turn, determine them), F t , and a component that is unexplained by the macro fundamentals, Ut : (43)
When ~St is averaged over periods of five years or more, the estimated slope coefficient from a regression of ~St onto F t tends toward unity, and the R2 rises dramatically, a result that holds for both the PPP and the VIP (uncovered interest rate parity) analyses. This implies that the variance of the movement in the exchange rate unexplained by the macro fundamentals declines dramatically over periods of five years or more, which must mean that the year-by-year timeseries errors cancel out approximately: n
(44)
(lln)L Ut+i ~ 0,
for n > 5 years.
i=O
Thus, there appears to be little effect from omitting Ut from the regressions using averaged data. However, the fact that the estimated slope coefficient in the regressions using annual, unaveraged data is quite different from the estimates obtained using the averaged data suggests that the unexplained compo.. nent is correlated with the component of the change explained by the macro fundamentals at the annual level: (45)
One interpretation of our results is that averaging Ut provides a filter that vastly reduces the variance of the disturbance term and that much of this reduction is in terms of reduced covariation of the residual with the fundamental. This covariance, of course, would also be reduced if we had truly exogenous fundamentals or instruments for the fundamentals. Thus, our analysis suggests the following. First, short-run deviations of the exchange rate, from the path consistent with the macro fundamentals alone, are responsible for the greater proportion of the short-run variation in nominal exchange rates. Second, these deviations apparently cancel out over periods of five years· or more. Third, and perhaps most puzzling, the deviations appear to be correlated with the fundamentals themselves in the annual data.
285
8.4
Exchange Rate Economics: What's Wrong with the Macro Approach?
Conclusion: What's Wrong with the Conyentional Macro Approach?
The empirical work summarized in sections 8.1 and 8.2 suggests that, for industrial countries during "normal times" (i.e., when they are not experiencing economic pathologies such as a hyperinflation), conventional macro fundamental models of exchange rate behavior are incapable of explaining the greater proportion of the variation in nominal exchange rate movements. It is apparent that there are important influences, not on the list of standard macro fundamentals, that affect short-run exchange rate behavior, and standard macro-based models perform poorly when subjected to standard time-series econometric testing--.-typically providing poor in-sample fits and miserable postsample predictive performance. Hence, there seems to be little professional disagreement with the view that, as a guide to the short-run behavior of the major exchange rates, exchange rate models based on macro fundamentals have largely failed. The macro-based models have, to some extent, been rehabilitated in studies that have used cointegration or error-correction-type models to forecast the exchange rate. These studies, however, in common with recent cointegration studies on exchange rates and purchasing power parity, provide evidence for the view that it is the longer-run or low-frequency movements in exchange rates that are correlated with the traditional macro fundamentals, while the shorter run movements are poorly understood or, to use the applied researcher's euphemism, "noisy." Some generic evidence on the relevance of economic fundamentals for short-run exchange rate behavior ~s provided in a recent study by Flood and Rose (1993). Observing the increased volatility of exchange rates under floating as opposed to fixed exchange rate regimes, these authors argue that any tentatively adequate exchange rate model should have fundamentals that are also much more volatile during floating-rate regimes. In fact, they find little shift in the volatility of economic fundamentals suggested by flexible-price or sticky-price monetary models across different nominal exchange rate regimes for a number of OECD exchange rates. Similar evidence is reported by Baxter and Stockman (1989). More generally, a number of studies have noted that, under the recent float, nominal exchange rates have shown much greater variability than important macroeconomic fundamentals such as price levels and real incomes (e.g., Dornbusch and Frankel 1988; Frankel and Froot 1990; Marston 1989).20 Again, this suggests that there are speculative forces at work in 20. Some analyses suggest that exchange rate volatility can change dramatically across regimes even though the volatility of the macro fundamentals does not. This point was crucial, e.g., in the Dornbusch (1976) overshooting model and in the Krugman (1991) target zone model. In both these models, and in rational expectations models in general, the Lucas critique applies, and the form of the reduced-form relation between the exchange rate and the fundamentals is not invariant
286
Robert P. Flood and Mark P. Taylor
the foreign exchange market that are not reflected in the usual menu of macroeconomic fundamentals: given the exhaustive interrogation of the macro fundamentals in this respect over the last twenty years, it would seem that our understanding of the short-run behavior of exchange rates is unlikely to be further enhanced by further examination of the macro fundamentals. And it is in this context that new work on the microstructure of the foreign exchange market seems both warranted and promising. The results of the research program into the speculative efficiency of the foreign exchange market also have important implications for the new research program into the microstructure of foreign exchange markets as well as the more conventional, macro-based approach. In particular, the rejection of the simple speculative efficiency hypothesis as applied to the foreign exchange market and the stylized empirical fact of a negative covariation between the rate of depreciation and the forward premium challenges conventional, macrobased approaches to the foreign exchange market since it suggests that one cannot take for granted many of the efficiency conditions that are typically subsumed in macro-based exchange rate models. Even under the assumption of risk aversion alongside that of rational expectations, the stylized fact of the so-called negative discount bias is very hard to explain (Fama 1984). Moreover, the evidence, from studies employing survey data, that foreign exchange market participants are neither risk averse nor conform to the rational expectations hypothesis suggests that the heterogeneity of agents' expectations across the foreign exchange market-itself highlighted in some recent survey data studies-may itself be an important feature determining short-run exchange rate behavior. The processes by which information is obtained and disseminated throughout markets is not amenable to analysis within a standard macro approach but is clearly of major importance given heterogeneity of agents' expectations and information sets (see, e.g., Lyons 1995). Information processing may also be at the root of the contagion in volatility across foreign exchange markets that has been documented. 21 Moreover, the finding that a high proportion of foreign exchange market participants deliberately use analytic techniques that ignore macro fundamentals (i.e., "technical" or "chartist" analysis), especially over shorter horizons (Taylor and Allen 1992), underscores the importance of allowing for the interaction of diverse forces in the short-run determination of exchange rates (Goodhart 1988; Frankel and Froot 1990). to the policy regime. To circumvent this well-known critique as well as the possibility that the presence of exchange rate bubbles may be regime dependent, Flood and Rose (1993) studied exchange rate and fundamentals volatility in structural equations rather than reduced-form expressions so that their conclusions are close to immune to the Lucas critique and bubbles issue. We say "close to immune" because it could be that exchange rate policy change results in structural coefficient drift, but this would be a thin reed on which to base exchange rate models. 21. So-called meteor showers (Engle, Ito, and Lin 1990).
287
Exchange Rate Economics: What's Wrong with the Macro Approach?
The additional empirical results reported in this paper underscore the view that empirical work on macro-based exchange rage models has been hampered by "contamination" of the data with a high degree of short-run noise. Industrial countries are different from each other, but the differences in exchange rate fundamentals change very slowly through time, while exchange rate time series show comparatively huge variation. To investigate this question, we work with a panel of data on twenty-two industrialized countries (including the United States) during the recent floating exchange rate era. In our investigation, we suppress noisy time-series variation in a series of steps. We first study two very simple exchange rate relations, relative purchasing power parity and uncovered interest rate parity, with annual data, then with fiveyear averaged, ten-year averaged, and, finally, twenty-year averaged data. Formally, as we temporally average the data, we also temporally average the time-series disturbances until, eventually, they disappear, leaving us with a cross-sectional model apparently purged of temporal noise. We find that the simple fundamentals models work extremely well in the pure cross section, with inflation or interest rate differentials explaining a very high proportion of the cross-sectional variation in exchange rate movements. We have, therefore, provided some additional evidence that the macro fundamentals should not be dismissed entirely. It is clear that the macro fundamentals in an important sense "set the parameters" within which exchange rates move but that these parameters are very broad indeed over the short run. Developing microstructural models of short-run exchange rate movements within these wide parameters is the challenge that researchers in this field now face.
288
Robert P. Flood and Mark P. Taylor
Appendix Detailed PPP Regression Results
Annual Changes
Table SA.I
4s t
= ex + y(4pt -
Sample
N
Estimates Method
Whole
420
OLS
Whole
420
WLS
Q1
105
OLS
Q1
105
WLS
Q2
105
OLS
Q2
105
WLS
Q3
105
OLS
Q3
105
WLS
Q4
105
OLS
Q4
105
WLS
4p~)
+
eX
.022 (.007) .026 (.008) .012 (.015) .015 (.015) .030 (.013) .027 (.014) .029 (.014) .028 (.015) .018 (.018) .027 (.020)
Et
'Y -.109 (.108) -.112 (.048) 1.254 (1.979) 1.534 (1.691) .820 (.706) .523 (.730) -.259 (.385) -.310 (.391) -.102 (.121) -.114 (.092)
F(",
= 1)
152.012 (.000) 529.266 (.000) .017 (.900) .100 (.750) .065 (.800) .426 (.520) 10.683 (.001) 11.175 (.001) 83.408 (.000) 146.526 (.000)
R2 .003 .011 .004 .011 .011 .001 .00:4 .003 .006 .014
Note: OLS stands for ordinary least squares and WLS for weighted least squares, using the absolute values of the regressor as weights. Qi denotes results for the ith quartile of the pooled sample when it is ordered according to the absolute magnitude of the regressor (so that Q4 contains the largest absolute values of the regressor). Figures in parentheses below coefficient estimates are estimated standard errors (and are heteroskedasticity robust for the OLS results); those below test statistics are marginal significance levels.
289
Exchange Rate Economics: What's Wrong with the Macro Approach?
Table 8A.2
Five-Year Averages
4s t = a +
Y(~t - ~~)
Estimates Method
Sample
N
Whole
84
OLS
Whole
84
WLS
Q1
21
OLS
Q1
21
WLS
Q2
21
OLS
Q2
21
WLS
Q3
21
OLS
Q3
21
WLS
Q4
21
OLS
Q4
21
WLS
+
Et
eX
~
F(", = 1)
R2
.356 (.742) -.108 (.984) .592 (1.109) -.474 (1.165) -.344 (1.264) -.523 (1.344) .510 (1.287) 1.221 (1.327) -.764 (2.077) -2.050 (2.684)
.807 (.159) -.955 (.064) 1.927 (1.539) 1.908 (1.298) .202 (.751) .327 (.769) -.223 (.364) -.202 (.344) .960 (.227) 1.059 (.143)
1.490 (.220) .463 (.490) .362 (.540) .489 (.490) 1.128 (.290) .767 (.390) 11.262 (.001) 12.188 (.002) .031 (.860) .171 (.680)
.432 .812 .065 .100 .004 .012 .019 .039 .695 .837
Note: OLS stands for ordinary least squares and WLS for weighted least squares, using the absolute values of the regressor as weights. Qi denotes results for the ith quartile of the pooled sample when it is ordered according to the absolute magnitude of the regressor (so that Q4 contains the largest absolute values of the regressor). Figures in parentheses below coefficient estimates are estimated standard errors (and are heteroskedasticity robust for the OLS results); those below test statistics are marginal significance levels.
290
Robert P. Flood and Mark P. Taylor
Ten-Year Averages
Table 8A.3
4s t = a + 'Y(4.pt -
4.p~)
Estimates Method
Sample
N
Whole
42
OLS
Whole
42
WLS
HI
21
OLS
HI
21
WLS
H2
21
OLS
H2
21
WLS
+
Et
&
'Y
F("! = 1)
R2
.165 (.722) -.086 (.855) -.333 (.775) -.020 (.916) .887 (1.258) -.125 (1.349)
.967 (.129) 1.058 (.059) .866 (.721) .880 (.619) .924 (.162) 1.060 (.085)
.062 (.800) .977 (.330) .035 (.850) .038 (.850) .221 (.643) .495 (.490)
.693 .931 .082 .095 .753 .933
Note: OLS stands for ordinary least squares and WLS for weighted least squares, using the absolute values of the regressor as weights. HI denotes results for the first half of the pooled sample when it is ordered according to the absolute magnitude of the regressor and H2 the second half of the ordered sample (so that H2 contains the largest absolute values of the regressor). Figures in parentheses below coefficient estimates are estimated standard errors (and are heteroskedasticity robust for the OLS results); those below test statistics are marginal significance levels.
Table 8A.4
Twenty-Year Averages
4s t = a + 'Y(4.pt Sample
N
Estimates Method
Whole
21
OLS
Whole
21
WLS
HI
10
OLS
HI
10
WLS
H2
11
OLS
H2
11
WLS
4.p~)
+
Et
&
'Y
F(,,! = 1)
R2
.224 (.911) .152 (.824) -.779 (.278) -.806 (.265) 1.402 (1.851) .580 (1.423)
.905 (.116) .957 (.062) 1.353 (.177) 1.323 (.164) .819 (.179) .935 (.097)
.673 (.410) .476 (.500) 4.003 (.050) 3.881 (.080) 1.013 (.310) .455 (.517)
.735 .957 .864 .890 .686 .952
Note: OLS stands for ordinary least squares and WLS for weighted least squares, using the absolute values of the regressor as weights. HI denotes results for the first half of the pooled sample when it is ordered according to the absolute magnitude of the regressor and H2 the second half of the ordered sample (so that H2 contains the largest absolute values of the regressor). Figures in parentheses below coefficient estimates are estimated standard errors (and are heteroskedasticity robust for the OLS results); those below test statistics are marginal significance levels.
291
Exchange Rate Economics: What's Wrong with the Macro Approach?
References Adler, Michael, and Bruce Lehmann. 1983. Deviations from purchasing power parity in the long run. Journal of Finance 38, no. 4: 1471-78. Backus, David K. 1984. Empirical models of the exchange rate: Separating the wheat from the chaff. Canadian Journal of Economics 17:824-46. Baxter, Marianne, and Alan C. Stockman. 1989. Business cycles and the exchange rate regime: Some international evidence. Journal of Monetary Economics 23, no. 3:377-400. Bilson, John E O. 1978. Rational expectations and the exchange rate. In The economics of exchange rates, ed. Jacob A. Frenkel and Harry G. Johnson. Reading, Mass. Addison-Wesley. - - - . 1981. The speculative efficiency hypothesis. Journal of Business 54:435-51. Branson, William H., Hannu Halttunen, and Paul R. Masson. 1977. Exchange rates in the short run: The dollar-deutschemark rate. European Economic Review 10, no. 4:303-24. Branson, William H., and Dale W. Henderson. 1985. The specification and influence of asset markets. In Handbook of international economics, ed. Ronald W. Jones and Peter B. Kenen. Amsterdam: North-Holland. Cheung, Yin-Wong, and Kon S. Lai. 1993. Long-run purchasing power parity during the recent float. Journal of International Economics 34, no. 1: 181-92. Cumby, Robert E., and Maurice Obstfeld. 1984. International interest rate and price level linkages under flexible exchange rates: A review of recent evidence. In Exchange rate theory and practice, ed. John E O. Bilson and Richard C. Marston. Chicago: University of Chicago Press. Dominguez, Kathryn M., and Jeffrey A. Frankel. 1993. Does foreign exchange intervention matter? The portfolio effect. American Economic Review 83, no. 4: 1356-69. Domowitz, Ian, and Craig Hakkio. 1985. Conditional variance and the risk premium in the foreign exchange market. Journal of International Economics 19, no. 3:47-66. Dornbusch, Rudiger. 1976. Expectations and exchange rate dynamics. Journal ofPolitical Economy 84, no. 6: 1161-76. - - - . 1980. Exchange rate economics: Where do we stand? Brookings Papers on Economic Activity 143-85. Dornbusch, Rudiger, and Jeffrey Frankel. 1988. The flexible exchange rate system: Experience and alternatives. In International Trade and Finance in a Polycentric World, ed. Silvio Borner. New York: St. Martin's Press. Engle, Robert E, Takatoshi Ito, and Wen-Ling Lin. 1990. Meteor showers or heat waves? Heteroskedastic intra-daily volatility in the foreign exchange market. Econometrica 58, no. 3:525-42. Fama, Eugene E 1984. Forward and spot exchange rates. Journal ofMonetary Economics 14, no. 3:319-38. Flood, Robert P. 1981. Explanations of exchange rate volatility and other empirical regularities in some popular models of the foreign exchange market. CarnegieRochester Conference Series on Public Policy 15 (Autumn): 219-49. Flood, Robert P., and Andrew K. Rose. 1993. Fixing exchange rates: A virtual quest for fundamentals. Discussion Paper no. 838. London: Centre for Economic Policy Research. Frankel, Jeffrey A. 1979. On the mark: A theory of floating exchange rates based on real interest differentials. American Economic Review 69:610-22. - - - . 1982. In search of the exchange risk premium: A six-currency test assuming mean-variance optimization. Journal of International Money and Finance 1, no. 3:255-74.
292
Robert P. Flood and Mark P. Taylor
- - - . 1988. Recent estimates of time-variation in the conditional variance and in the exchange risk premium. Journal of International Money and Finance 7, no. 1:115-25. - - - . 1990. Zen and the art of modern macroeconomics: The search for perfect nothingness. In Monetary policy for a volatile global economy, ed. William Haraf and Thomas Willett. Washington, D.C.: American Enterprise Institute. - - - . 1993. Monetary and portfolio balance models of the detennination of exchange rates. In On exchange rates. Cambridge, Mass.: MIT Press. Frankel, Jeffrey A., and Kenneth A. Froot. 1990. Chartists, fundamentalist8 and the demand for dollars. In Private behaviour and government policy in interdependent economies, ed. Anthony S. Courakis and Mark P. Taylor. Oxford: Oxford University Press. Frankel, Jeffrey A., and Andrew K. Rose. 1994. An empirical characterization of nominal exchange rates. Department of Economics, University of California, Berkeley. Mimeo. Frenkel, Jacob A. 1976. A monetary approach to the exchange rate: Doctrinal aspects and empirical evidence. Scandinavian Journal of Economics 78, no. 1:200-224. - - - . 1981. Flexible exchange rates, prices and the role of "news." Journal ofPolitical Economy 89:665-705. Frenkel, Jacob A., and Harry G. Johnson, eds. 1978. The economics of exchange rates. Reading, Mass. : Addison-Wesley. Friedman, Milton. 1953. The case for flexible exchange rates. In Essays in positive economics, ed. Milton Friedman. Chicago: University of Chicago Press. Froot, Kenneth A., and Jeffrey A. Frankel. 1989. Forward discount bias: Is it an exchange risk premium? Quarterly Journal of Economics 104, no. 1:139-61. Froot, Kenneth A., and Richard H. Thaler. 1990. Anomalies: Foreign exchange. Journal of Economic Perspectives 4, no. 2: 179-92. Giovannini, Alberto, and Philippe Jorion. 1989. The time variation of risk and return in the foreign exchange and stock markets. Journal of Finance 44, no. 2:307-25. Goodhart, Charles A. E. 1988. The foreign exchange market: A random walk with a dragging anchor. Economica 55, no. 4:437-60. Hansen, Lars P., and Robert 1. Hodrick. 1983. Risk averse speculation in the forward foreign exchange market: An econometric analysis of linear models. In Exchange rates and international macroeconomics, ed. Jacob A. Frenkel. Chicago: University of Chicago Press. Hodrick, Robert J. 1987. The empirical evidence on the efficiency of forward and futures foreign exchange markets. London: Harwood. Hooper, Peter, and John Morton. 1982. Fluctuations in the dollar: A model of nominal and real exchange rate determination. Journal of International Money and Finance 1, no. 1:39-56. Krugman, Paul R. 1991. Target zones and exchange rate dynamics. Quarterly Journal of Economics 106, no. 3:669-82. Lewis, Karen K. 1988. Testing and portfolio balance model: A multilateral approach. Journal of International Economics 24: 133-50. - - - . 1989. Changing beliefs and systematic rational forecast errors with evidence from foreign exchange. American Economic Review 79, no. 4:621-36. - - - . 1994. International financial markets. Wharton School, University of Pennsylvania. Mimeo. Lothian, James R., and Mark P. Taylor. 1994. Real exchange rate behavior: The recent float from the perspective of the past two centuries. Washington, D.C.: International Monetary Fund. Lucas, Robert E. 1982. Interest rates and currency prices in a two-country world. Journal of Monetary Economics 10, no. 4:335-59.
293
Exchange Rate Economics: What's Wrong with the Macro Approach?
Lyons, Richard K. 1995. Tests of microstructural hypotheses in the foreign exchange market. Journal of Financial Economics 39:321-51. MacDonald, Ronald. 1988. Purchasing power parity: Some "long-run" evidence from the recent float. Economist 136, no. 2:239-51. MacDonald, Ronald, and Mark P. Taylor. 1992. Exchange rate economics: A survey. International Monetary Fund Staff Papers 39, no. 1: 1-57. - - - . 1993. The monetary approach to the exchange rate: Rational expectations, long-run equilibrium, and forecasting. International Monetary Fund StaffPapers 40, no. 1:89-107. - - - . 1994. The Monetary Model of the exchange rate: Long-run relationships, short-run dynamics, and how to beat a random walk. Journal ofInternational Money and Finance. Mark, Nelson C. 1990. Real and nominal exchange rates in the long run. Journal of International Economics 28, no. 1: 115-36. - - - . 1992. Exchange rates and fundamentals: Evidence on long-horizon predictability and overshooting. Working paper. Ohio State University, August. Marston, Richard C. 1989. Real and nominal exchange rate variability. Empirica 16, no. 2:147-60. McNown, Robert, and Myles Wallace. 1989. National price levels, purchasing power parity and cointegratiori: A test of four high-inflation economies. Journal ofInternational Money and Finance 8, no. 4:533-45. Meese, Richard A., and Kenneth Rogoff. 1983a. Empirical exchange rate models of the seventies: Do they fit out of sample. Journal of International Economics 14, no. 2:3-24. - - - . 1983b. The out-of-sample failure of empirical exchange rate models: Sampling error or misspecification? In Exchange rates and international macroeconomics, ed. Jacob A. Frenkel. Chicago: University of Chicago Press. - - - . 1988. Was it real? The exchange rate-interest differential relationship over the modern floating-rate period. Journal of Finance 43, no. 4:933-48. Mussa, Michael. 1979. Empirical regularities in the behavior of exchange rates and theories of the foreign exchange market. Carnegie-Rochester Conference Series on Public Policy 11:9-57. - - - . 1986. Nominal exchange rate regimes and the behavior of real exchange rates: Evidence and implications. Carnegie-Rochester Conference Series on Public Policy 25: 117-213. Officer, Lawrence H. 1980. Effective exchange rates and price ratios over the long-run: A test of the purchasing power parity theory. Canadian Journal of Economics 8:206-30. Rogoff, Kenneth. 1979. Expectations and exchange rate volatility. Ph.D. diss., Massachusetts Institute of Technology. - - - . 1984. On the effects of sterilized intervention: An analysis of weekly data. Journal of Monetary Economics 14: 133-50. Stockman, Alan C. 1980. A theory of exchange rate determination. Journal ofPolitical Economy 88, no. 4:673-98. - - - . 1983. Real exchange rates under alternative nominal exchange rate systems. Journal of International Money and Finance 2: 147-66. - - - . 1987. The equilibrium approach to exchange rates. Federal Reserve Bank of Richmond Economic Review 73, no. 2:12-31. - - - . 1988. Real exchange rate variability under pegged and nominal floating exchange rate systems: An equilibrium theory. Working Paper no. 2565. Cambridge, Mass.: National Bureau of Economic Research. Taylor, Mark P. 1988. An empirical examination of long-run purchasing power parity using cointegration techniques. Applied Economics 20, no. 10: 1369-82.
294
Robert P. Flood and Mark P. Taylor
- - - . 1989. Covered interest arbitrage and market turbulence. Economic Journal 99, no. 2:376-91. - - - . 1994. The economics of exchange rates. Washington, D.C.: International Monetary Fund. Mimeo. Taylor, Mark P., and Helen L. Allen. 1992. The use of technical analysis in the foreign exchange market. Journal of International Money and Finance 11, no. 3:304-14. Taylor, Mark P., and Patrick C. McMahon. 1988. Long-run purchasing power parity in the 1920s. European Economic Review 32, no. 1:179-97. Throop, Adrian W. 1993. A generalized uncovered interest parity model of exchange rates. Federal Reserve Bank of San Francisco Economic Review, no. 2:3-16. Tobin, James. 1969. A general equilibrium approach to monetary theory. Journal of Money, Credit and Banking 1, no. 1:15-29.
Comment
Andrew K. Rose
There are basically two parts to this paper. The first (and larger) part is a survey of the conventional international macroeconomic literature on floating exchange rates. This work, closely related to a series of papers by Ronald MacDonald and Taylor, is a well-balanced, thorough survey of the evidence on exchange rate determination. The authors survey four classes of theoretical models of exchange rate determination and show that the empirical performance of these models is quite poor. I find the work eminently reasonable and have very little to add. The second part of the paper is. relatively short but innovative and interesting. The authors assemble a panel of data on exchange rates and monetary fundamentals for over twenty countries in the period since the collapse of the Bretton Woods regime. The authors show that pooling the annual observations provides only very weak evidence of tendencies to purchasing power parity (PPP). However, moving to coarser time frequencies seems to give much stronger evidence of PPP. For instance, using five-year averages gives a higher regression coefficient in a regression of the exchange rate change on the inflation differential than a similar regression estimated with annual differences (the goodness of fit improves as well). Averaging the data over ten years improves matters even more. This is true despite the fact that time-averaged data have fewer outliers than finer-frequency data (as is obvious from the dimensions of figs. 8.5-8.7). It is also true despite the fact that the sample span of the data (in terms of both country and time coverage) stays constant when moving between frequencies. In this, as in many contexts, less is more; throwing away short-run variation in exchange and inflation rates leads to a fit more consistent with one's theoretical priors. It is interesting that throwing away Andrew K. Rose is associate professor and chair of economics at the Haas School of Business, University of California, Berkeley, a research fellow of the Centre for Economic Policy Research, and a research associate of the National Bureau of Economic Research.
295
Exchange Rate Economics: What's Wrong with the Macro Approach?
within-country time-series varIatIon (and increasing the proportion of between-country variation) not only does not hurt the regression analysis but positively helps matters. I find this result quite interesting. It clearly implies (as the authors point out) that there is a nontrivial correlation between the regression disturbance and the inflation differential that leads to a biased coefficient at the annual frequency. But what is the source of this correlation? Perhaps the authors have found out only that (price index) measurement error is fatal in such simple regressions at short- and medium-run frequencies. On the other hand, this kind of regression has been criticized by a number of economists in that it is not based on any underlying theoretical model of either price or exchange rate adjustment so that the regression coefficients are not structural parameters; in addition, there is no obvious alternative hypothesis posed. Exploring this result further, and trying to pin down the source of the correlation, remains an interesting and important task for future research. However, unlike many comparable tasks in international macroeconomics, it is a manifestly promising avenue of research since the results are consistent with the prior beliefs of many economists (including me), namely, that theories like PPP work well in the long run but are barely detectable in the short run. Thus, this paper has performed two valuable tasks. First, it has demonstrated the lack of empirical success of conventional macroeconomic exchange rate models. Second, it has produced an interesting set of results on model fit in both the short and the long run. However, my one criticism of it is that the authors could be more forceful in pinning down the specific failures of the macroeconomic approach to exchange rates that can potentially be answered with microstructure analysis. We have the opportunity to set the agenda for microstructural work; it is important to lend focus to this emerging literature at an early stage. Here is my wish list of important topics in which we should be interested (many are related): 1. First is the role of noise trading and bubbles. Can microstructural work explain the evidence of apparent excessive volatility in floating foreign exchange markets (derived from news regressions, the poor fit of macro equations, and potentially the literature on deviations from uncovered interest parity)? Can noise trading explain the high volume on foreign exchange markets? How about the apparent short-run near irrelevance of macroeconomic factors? 2. What leads to the heterogeneous beliefs manifestly apparent in foreign exchange markets? Is the source of the heterogeneity disagreements about macroeconomic phenomena, models of the economy, or something completely different? 3. How does intervention fit in? Why is it taken so seriously by both market participants and central banks when the macroeconomic presumption is that sterilized intervention is almost irrelevant? Is there a microstructural reason why intervention should be kept secret?
296
Robert P. Flood and Mark P. Taylor
4. Can microstructure analysis shed light on how the credibility of a foreign exchange authority (trying to limit exchange rate fluctuations) is determined? We know that macroeconomic phenomena are not all that important. 5. Can microstructure analysis explain the apparent preference of central banks for fixed or managed exchange rates? In general, how are issues of regime choice affected by micro features? 6. What is the behavior of market participants like in times of "unusual" activity, especially speculative attacks and hyperinflations? For instance, are micro phenomena relevant in precipitating or exacerbating speculative attacks? 7. Can microstructure analysis shed light on the apparent evidence of "contagion/infection" effects witnessed during the 1992-93 exchange rate mechanism crisis? In general, what are the spillovers between different foreign exchange markets? Are there externalities between these and other asset markets? 8. Do the microstructural features of foreign exchange markets have important consequences for hedging? If so, are there important effects on international trade flows? 9. Can microstructure explain the well-established large and persistent deviations from uncovered interest parity? In summary, the paper by Flood and Taylor is a valuable contribution in two respects: it provides a clean survey of the floating exchange rate literature, and it provides some intriguing new evidence on short-and long-run exchange rate behavior. The authors have convincingly demonstrated that macroeconomic models have not provided a satisfactory answer to the key question of exchange rate determination. However, it is still my belief that the macroeconomic analysis was after the right question, and I would urge us all to keep our collective eyes on the prize: Why do floating exchange rates fluctuate so much? Why do fixed exchange rates both persist the way they do and collapse? And what are the consequences for all this for international trade, macroeconomic policy, and foreign exchange market policies?
Comment
Lars E. O. Svensson
Flood and Taylor's paper is a fine survey of theories of exchange rate determination and empirical tests of these theories. It also presents interesting results in favor of long·run purchasing power parity. One of the main points of the survey is to emphasize the by now well-known Lars E. O. Svensson is professor of international economics at the Institute for International Economic Studies at Stockholm University, a research fellow of the Centre for Economic Policy Research, and a research associate of the National Bureau of Economic Research. Comments by Jeff Frankel are gratefully acknowledged.
297
Exchange Rate Economics: What's Wrong with the Macro Approach?
situation that existing theories cannot predict and explain short-run exchange rate movements during floating exchange rate regimes and, more specifically, that existing theories cannot outbeat a random walk. This unfortunate situation refers to floating exchange rates. I would like to add that, for fixed exchange rate regimes, the situation is different, in a particular aspect. Real-world fixed exchange rates usually have bands within which exchange rates fluctuate; the Bretton Woods system had a bandwidth of ± 1 percentage point; previously ERM (exchange rate mechanism), the European exchange rate cooperation, had a bandwidth of either ±2.25 or ±6 percentage points, whereas after the July 1993 crisis the bandwidth was increased to ± 15 percentage points (with the exception of the Dutch gulden/deutsche mark exchange rate). It turns out that it is much easier to explain and predict exchange rate movements within such bands than to predict freely floating exchange rates. Table 8C.l (from Svensson [1993, table 3]; see that paper for details) shows typical regression results for fixed exchange rate regimes with bands. The table shows, for six ERM currencies, regressions of the rate of depreciation within the band during the next three months (relative to the deutsche mark) on the current exchange rate (x), the currency's three-month Euro interest rate (i), and the deutsche mark three-month Euro rate (i*). The intercepts are allowed to differ across periods between realignment dates. If exchange rates within bands were martingales, the R2'S in these equations would be close to zero, and the coefficients of exchange rates and interest rates would not be significantly different from zero. Instead, we see that the R2'S lie between 0.26 and 0.62 and that all coefficients for the exchange rate are negative and significantly differ.ent from zero (Newey-West standard errors are reported within parentheses). When the currency is weak, it will on average appreciate. This is of course just mean reversion within the band, which perhaps is not so surprising. However, mean reversion is not the end of the story. We also see that the coefficient for the own-currency Euro interest rate is negative and significant in four cases out of six, whereas the coefficient for the foreign interest rate is positive and significant in two cases. This sign pattern is consistent with central banks engaging in interest rate smoothing. When the domestic-currency interest rate (or the interest rate differential relative to the deutsche mark) is high, the currency will in the future on average appreciate relative to the deutsche mark. This expected appreciation will prevent the domestic-currency interest rate from being even higher. Put differently, by inducing an expected appreciation of the currency, the central bank prevents the domestic interest rate from being even higher. This association between high interest rate differentials and future expected appreciation within the band may appear to be counter to uncovered interest parity, which predicts an association between high interest rate differentials and future expected depreciation. However, the latter association is between high interest rate differentials and expected total depreciation, including re-
298
Robert P. Flood and Mark P. Taylor
Table 8e.1
Exchange Rate Depreciation within the Band: Three Months (1) BF/DM
Intercepts: 13 March 1979 24 September 1979
13.13 (1.81 ) 16.98 (2.66)
30 November 1979
(2) DKIDM
18.56 (2.78)
(3)
(4)
FF/DM
IL/DM
2.42 (1.88) -2.66 (3.12)
1.47 (5.07) 8.79 (8.29)
2.86 (4.33)
7.95 (10.82) 12.03 (10.19)
15.28 (4.02)
22 February 1982 14 June 1982 21 March 1983
21.70 (2.51) 24.54 (2.35) 20.16 (2.53) 13.60 (1.97)
I
16.08 (3.90) 17.84 (3.71) 16.76 (4.16) 9.80 (2.55)
I 1.06 (3.72) -.88 (2.48)
22 July 1985 7 April 1986
I
12.54 (1.52)
4 August 1986 12 January 1987 8 January 19909 April 1992 Coefficients:
x
Diagnostics: N
Rl CT
I
9.40 (1.50)
I -1.25 (.26) -1.33 (.21) .37 (.12) 2,743 .62 1.8
I
13.80 (2.07)
I
10.83 (2.44)
I -2.12 (.28) -1.04 (.23) .22 (.18) 2,686 .56 2.6
(6)
NGIDM
7.85 (1.07) -.08 (1.09)
I
23 March 1981 5 October 1981
(5) IPIDM
I
2.83 (1.67)
I
.54 (1.97)
I -1.90 (.36) -.11 (.22) .37 (.20) 2,802 .43 2.7
I
8.76 (7.66) 4.44 (6.30) 5.99 (4.71) 2.81 (4.39)
I
6.39 (3.84) 3.30 (7.51) -1.42 (.39) -.29 (.22) .02 (.78) 2,624 .26 4.3
18.02 (4.19)
I 16.21 (3.38) 9.56 (2.84)
1.13 (.67)
I
12.46 (2.23) 13.00 (2.16) 7.30 (2.08)
I -1.75 (.42) -.80 (.21) .18 (.15) 2,211
-2.74 (.55) -.94 (.31) .84 (.27) 3,039
.55 2.3
.52 1.6
Note: OLS with Newey-West standard errors within parentheses (7 lags, 7 = 65). Regressand is (X'+T - x,)/7dt (%/year), tdt = 63/261; regressors are x,(%), i~ (%/year), and i;T (%/year), where x is the is In(BF/DM), ... , In(NG/DM), i~ is the 7 maturity BF, ... , NG Euro interest rate, and 7 maturity DM Euro interest rate. A vertical bar for a realignment date indicates that the corresponding currency was not realigned and that the estimate straight above applies. Interest rates for IP were not available before October 1981. The second regime for Danish krone/deutsche mark is too short to be estimated. For details, see Svensson (1993). BF = Belgian franc; DM = deutsche mark; DK = Danish krone; FF = French franc; IL = lira; IP = Irish pound; NG = Dutch gulden.
t;T
299
Exchange Rate Economics: What's Wrong with the Macro Approach?
alignments, whereas the former is between high interest rate differentials and expected depreciation relative to central parity. For a sample period during which a realignment is expected with some probability but does not occur owing to the shortness of the sample (the peso problem), high interest rate differentials will be associated with future realized total appreciation, even if expectations are rational. If expectations are not rational but exaggerate the probability of a realignment, for instance, in a situation in which a central bank is struggling to establish the credibility of a fixed exchange rate regime, high interest rate differentials will be associated with future total realized appreciations also for longer sample periods. Thus, central banks' interest rate smoothing within exchange rate bands may contribute to the often observed empirical rejection of uncovered interest parity. Additional evidence for interest rate smoothing within exchange rate bands is presented in figure 8C.1 (from Svensson [1994, fig. 2d]; see that paper for details). The top panel's thin curve shows a time series of deviations from central parity of the Swedish krona (relative to the currency basket that the krona was pegged to prior to May 1992 when the pegging to the ECU [European currency unit] began). The horizontal dotted lines show the bandwidth, ± 1.5 percentage points. The thin curves in the second and third panels from the top show the Swedish krona and currency basket one-month interest rates, respectively. The curve in the fourth panel from the top shows estimates of the market's realignment expectations (expected rates of realignment). Whereas the thin curves show the actual historical development of these series, the thick curves in the first and second panels from the top show the result of a simulation where the central bank faces historical disturbances but pursues an optimal policy according to an objective function that puts certain weight on both exchange rate and interest rate smoothing. Let us look at just one incident. In the fall of 1990, expectations of a devaluation rose dramatically. In the fourth panel from the top, we see that the expected rate of realignment peaked in the fall of 1990. Everything else equal, this would show up in a one-to-one increase in the domestic-currency interest rate. Instead, we see in the second panel from the top that the domestic interest rate increased by much less (the actual increase and the optimal increase coincide in this incident). As can be seen in the top panel, Sveriges Riksbank allowed the krona to depreciate within the band (somewhat less than the optimal policy). This created an expected appreciation for the krona that dampened the effect of the increased realignment expectations on the domestic interest rate. The resulting expected rate of appreciation of the krona is displayed in the fifth panel from the top. Hence, in contrast to what is the case for floating exchange rates, for exchange rates within bands exchange rate movements can be both predicted and to some extent explained, for instance, in terms of central banks smoothing interest rates. Of course, it would be better to be able to predict and explain movements of floating exchange rates and in general to explain exchange rate
300
Robert P. Flood and Mark P. Taylor
~._
_
_
_
x
_
..
..-I
3
3 N N
-----------------~--~~---------------
------3
3
3
Fig. 8e.1
Evidence of interest rate smoothing
Note: x is percentage deviation of exchange rate from central parity; iT and i*T are domestic and foreign one-month interest rates, respectively; gT is expected rate of realignment; 'T equals four weeks, dt = one week. Thin curve: actual. Thick curve: optimal. Dashed line: mean. For details, see Svensson (1994).
levels in terms of levels of other contemporaneous macro variables, but it is still of some interest that exchange rate movements within bands are systematic and also not exclusively characterized by mean reversion. I missed in Flood and Taylor's survey a discussion of international macro problems or puzzles where a microstructure approach might help. A list of such problems/puzzles is discussed extensively in Andrew Rose's comment on Flood and Taylor. Let me just mention two such problems/puzzles that are on my own wish list. The effects on exchange rates of sterilized central bank intervention may be suitable to analyze with a microstructure approach. Most macro studies find little or no effect of sterilized intervention. If they are correct, why do central
301
Exchange Rate Economics: What's Wrong with the Macro Approach?
banks keep doing it? Some studies emphasize a signaling effect, namely, that sterilized interventions signal future nonsterilized interventions. If the signaling effect is the important one, why are central banks often secretive about their interventions? And why aren't there better ways to signal changes in future monetary policy? Some central bank officials have suggested to me that herd behavior in the foreign exchange market may be exploited by central banks. If the central bank secretly through intermediaries can convey the impression that someone other than the central bank is suddenly starting to buy the currency, the herd might follow. This way the central bank would then be able to push or pull the exchange rate in desired directions. These issues might be suitable for a microstructure approach. Another such issue is speculative attacks, in particular, central banks' defense against speculative attack. For instance, during the dramatic defense of the Swedish krona in September 1992, with a 500 percent overnight rate, the foreign exchange market for kronor and the Swedish Treasury bills market seemed paralyzed. Instead of a strong inflow of kronor, very little trade occurred. Huge spreads were quoted on Swedish T-bills, indicating that no one wanted to trade. One reason is that it became almost impossible to price a short-term bill since the uncertainty about the future overnight rate was enormous. Who· knew whether it would come down to reasonable levels within three, seven, or fourteen days? A related issue is to what extent the increased practice of trading rules, dynamic hedging, etc. affects central banks' defense against speculative attack. As the paper by Garber and Spencer (chap. 6 in this volume) demonstrates, this is another suitable area for a microstructure approach.
References Svensson, Lars E. O. 1993. Assessing target zone credibility: Mean reversion and devaluation expectations in the ERM, 1979-1992. European Economic Review 37:763-802. - - - . 1994. Why exchange rate bands: Monetary independence in spite of fixed exchange rates. Journal of Monetary Economics 33: 157-94.
9
Is There a Safe Passage to EMU? Evidence on Capital Controls and a Proposal Barry Eichengreen, Andrew K. Rose, and Charles Wyplosz
The 1992 and 1993 crises in the European Monetary System redirected attention toward proposals for regulating the foreign exchange market. Academics (including two authors of this paper) argued for a Tobin tax on foreign exchange transactions or the imposition of non-interest-bearing deposit requirements on banks with open positions in foreign exchange as a way of smoothing the transition to European monetary union (EMU) (see Eichengreen and Wyplosz 1993). European Commission President Jacques Delors mooted the idea of reimposing capital controls. The Monetary Subcommittee of the Committee on Economic and Monetary Affairs of the European Parliament called on the European Commission to submit detailed proposals for regulating foreign exchange transactions (European Parliament 1993). Others (Goldstein et al. 1993; Mussa and Goldstein 1994) voiced skepticism about the desirability and effectiveness of such measures. This controversy rekindled interest in the role played by capital controls in the operation of pegged exchange rate systems. Some authors (e.g., Wyplosz 1986; Giovannini 1989) have argued that controls played an important role in Barry Eichengreen is the John L. Simpson Professor of Economics and professor of political science at the University of California, Berkeley, a research fellow of the Centre for Economic Policy Research, and a research associate of the National Bureau of Economic Research. Andrew K. Rose is associate professor at the Haas School of Business, University of California, Berkeley, a research fellow of the Centre for Economic Policy Research, and a research associate of the National Bureau of Economic Research. Charles Wyplosz is professor of business at INSEAD (the European Institute of Business Administration) and a research fellow of the Centre for Economic Policy Research. For comments and suggestions the authors thank, without implicating, Javier Alonso, Vittorio Grilli, Assar Lindbeck, Luis Linde, Jorge de Macedo, Philippe Moutot, William Perraudin, Torsten Persson, Lars Svensson, Jose Vifials, and seminar participants at the Institute for International Economic Studies at the University of Stockholm, and the NBERICEPRlBank of Italy conference. Assistance was provided by Florence Beranger and Luis Freitas de Oliveira.
303
304
Barry Eichengreen, Andrew K. Rose, and Charles Wyplosz
virtually all systems of pegged exchange rates since World War II. In this view, controls reconciled a modicum of policy autonomy with the commitment to pegged exchange rates, provided the authorities breathing space to organize orderly realignments, and made it easier to rebuff speculative attacks not grounded in fundamentals. Others (e.g., Gros 1987; Gros and Thygesen 1992; Truman 1994) have argued that capital controls were always easy to evade and never played a major role in limiting exchange rate flexibility. In this paper, we seek to advance this debate. Using data for twenty-two countries over twenty-five years, we show that capital controls have been associated with significant differences in the behavior of such macroeconomic variables as budget deficits and money growth rates. This supports the view that, historically, controls have made a difference. This evidence provides the point of departure for the second half of the paper, where we argue the case for measures, specifically non-interest-bearing deposit requirements on lending to nonresidents, that mimic some of the effects of capital controls as a way of easing the transition to EMU. Our focus here is on temporary measures and on Europe rather than on the case for regulating foreign exchange transactions in general. It is motivated by the problem of how to complete the transition to EMU. We take this objective as given and ask whether non-interest-bearing deposit requirements are needed to achieve it. The argument for deposit requirements runs as follows. The Maastricht Treaty on European Union and the Single European Act to which it is a successor mandate the removal of capital controls by EU (European Union) countries and their maintenance of exchange rate stability for an extended period as prerequisites for participating in EMU. The removal of controls and the extended period of exchange rate stability may be incompatible, however, for the absence of controls increases the cost borne by monetary authorities seeking to defend themselves against speculative attacks of the sort that buffeted the EMS (European Monetary System) in 1992-93. 1 It is therefore necessary to alleviate this predicament. We provide evidence on the channels through which speculative pressures are transmitted and therefore on the appropriate nexus for intervention. We discuss political constraints associated with the treaty and suggest that they provide a justification for the selective use of deposit requirements. The rest of the paper is organized as follows. In section 9.1, we present evidence on the effects of controls. Section 9.2 discusses the transmission of speculative pressure and the feasibility of alleviating it through the imposition of non-interest-bearing deposit requirements on bank lending to nonresidents. Section 9.3 draws out the implications for the Maastricht Treaty and the 1996 Intergovernmental Conference (IGC). Section 9.4 is a brief conclusion. 1. Our discussion builds on recent theoretical contributions to the literature on speculative attacks such as Ozkan and Sutherland (1994) and Obstfeld (1994).
305
9.1
A Safe Passage to EMU?
Historical Evidence on the Operation of Capital Controls
In this section, we compare the behavior of macroeconomic variables during periods of tranquillity and speculative pressure. We ask whether there are differences in the behavior of such variables when capital controls are in place. A negative answer is consistent with the view that controls are an ineffectual policy instrument. Evidence that the behavior of macroeconomic variables differs significantly when controls are present does not establish that controls are responsible for those differences, of course; a government might prefer both controls and certain macroeconomic polices even if the two are causally unrelated. But a finding of differences in the stance of macroeconomic variables is at least consistent with the view that capital controls are a policy tool of economic significance. To analyze the behavior of economic variables around the time of speculative attacks, it is necessary to have a selection criterion for attacks that does not bias one toward finding certain patterns in the data. Large exchange rate changes are not the same thing as speculative attacks on pegged rates. For one thing, not all attacks succeed. In addition, large month-to-month changes in exchange rates are sometimes observed when rates are floating freely and it is impossible to launch an attack on official reserves because the authorities are not intervening. When exchange rates are pegged, attacks can be rebuffed by raising interest rates (relative to those prevailing abroad) and/or by committing international reserves. Examining only successful attacks might bias one toward a particular characterization of why attacks occur. In particular, considering only attacks that succeed is likely to lead one to conclude that controls are ineffective. An alternative is to construct an index of speculative pressure composed of a combination of exchange rate changes, reserve changes, and interest rate changes, as we did in Eichengreen, Rose, and Wyplosz (1994).2 Changes in exchange rates will be observed when the authorities are unwilling or unable to resist pressure to realign. (We consider only countries and periods when currencies were pegged under the provisions of explicit bands such as the Bretton Woods system, the Snake, and the EMS.) Increases in interest rates and declines in reserves will be observed when the authorities seek to defend the exchange rate against attack. We weight the three components of our index so that their conditional volatilities are equal. 3 We construct this measure using monthly data for the DECD 2. The present discussion of data and methodology is much abbreviated; interested readers are referred to this previous paper. 3. In our earlier paper, we conducted sensitivity analysis to gauge how much difference different weighting schemes made. Theory can be used to pin down the weights only if one adopts an empirical model of the connection between macroeconomic fundamentals and the exchange rate. The professional consensus is, however, that none of the existing models performs adequately for empirical work (see Meese and Rogoff 1983).
306
Barry Eichengreen, Andrew K. Rose, and Charles Wyplosz
countries and selected other economies drawn from the CD-ROM version of International Financial Statistics. We supplement this with information on capital controls from the IMF's Annual Report on Exchange Arrangements and Exchange Restrictions (volumes from 1967 through 1992).4 We take Germany as the reference country, computing changes in exchange rates, interest rates, international reserves, etc., relative to their German values. We specify a threshold for the index of speculative pressure (typically two standard deviations from the sample mean) and categorize as attacks all months in which its value exceeds that threshold. To avoid treating successive months when a currency came under attack as separate episodes, we define an exclusion window (typically plus or minus six months) and disregard crises other than the first that fall within the window. As a control group of noncrises against which our crises can be compared, we take all other nonoverlapping periods that are left once the episodes of speculative attack are removed. 5 These data have a number of limitations. First, published series on international reserves are imperfect. Central banks sometimes report only the gross foreign assets of the monetary authorities. Since a standard procedure is to arrange for standby credits in foreign currency, this is a potentially serious problem. When the authorities intervene, they draw on their credit lines without having to sell reported foreign assets. Even countries that provide data on foreign liabilities omit a number of operations that are typically undertaken during periods of speculative pressure, such as off-balance-sheet transactions like swaps and forward market intervention. 6 Moreover, intervention by foreign central banks may be hard to detect. Because we analyze changes in the reserves of each country relative to changes in German reserves, we will pick up intervention by the Bundesbank in support of foreign currencies. Intervention by third countries will not be detected, however. Insofar as Germany has been the strong-currency country on which the bulk of foreign intervention obligations have fallen (especially within the EMS), this will not be a serious problem. But, even in the EMS, intervention has been undertaken by third countries (e.g., by the Netherlands), which we will not capture. Moreover, monthly data may not be of a sufficiently fine periodicity to identify every attack (especially unsuccessful ones). Pressure against pegged currencies can mount quickly and be repelled within the month through interest rate increases or foreign exchange market intervention. If an attack is launched
4. Our countries were chosen on the basis of data availability and include the United States, the United Kingdom, Austria, Belgium, Denmark, France, Germany, Italy, the Netherlands, Norway, Sweden, Switzerland, Canada, Japan, Finland, Greece, Ireland, Portugal, Spain, Australia, South Africa, India, and South Korea. 5. In our earlier paper, we conducted sensitivity analysis varying the width of the exclusion window as well as the two-standard-deviation threshold for identifying crises. 6. Only comprehensive data on exchange market intervention, which is currently made available on a limited basis by only a few central banks, would solve this problem.
307
A Safe Passage to EMU?
and repelled in a matter of days, the average behavior of interest rates and international reserves over the month may not reveal its intensity. Finally, the available measures of controls provide only blunt indicators of their prevalence. Here we use the IMF's binary indicator of the presence of restrictions on capital transactions. 7 This crude measure provides minimal information about the intensity of controls. Given the scope for measurement error, we would not be surprised if the variable had little explanatory power; correspondingly, we take seriously any positive results. The list of speculative crises that results from the application of this methodology to data for post-1966 exchange rate pegs in the twenty-two countries is discussed in Eichengreen, Rose, and Wyplosz (1994). It includes prominent devaluations and realignments of OEeD currencies but also a number of episodes in which interest rates were increased significantly and/or international reserves were run down. We start by considering the distributions of macroeconomic variables. We first examine crises and ask whether the behavior of these variables when there were capital controls in place differs significantly from their behavior in the absence of controls. Given the limited sample size, we provide two nonparametric tests of the equality of distributions: the Kolmogorov-Smimov test, which considers the entire distribution, and the Kruskal-Wallis test, which focuses on the median. We provide a t-test for the equality of the sample means in the presence and absence of controls. We then compute identical test statistics for periods of tranquility (i.e., noncrises). Finally, we provide an analogous set of statistics for actual realignments and changes in exchange rate regime, which we dub events to distinguish them from crises. Table 9.1 reports the basic results. The left-most panel considers attacks ("crises"), while the right-most panel refers to tranquil periods ("noncrises"). Using the Kolmogorov-Smimov statistic, we cannot reject except at the 26 percent confidence level the null that the distribution of fiscal ratios (the ratio of the budget balance to GDP) is identical for crises that took place in the presence and absence of capital controls. The same is true for the level of the real exchange rate, for the level of the interest rate differential, and for the differential rate of growth of foreign exchange reserves. On the other hand, we can reject the null that the inflation differential, the smoothed trade balance (the ratio of exports over imports), domestic credit, and money growth are distributed equally for crises that took place in the presence and absence of capital controls. 8 Parametric tests reject the null of equal means for inflation, the trade balance, and money growth. The differential rate of money growth was 1.9 percent (annualized) for speculative crises with controls in place and -6.6 percent 7. The absence of Euromarket data for most of the sample means that offshore-onshore interest rate differentials, another potential indicator of controls, are unavailable to us. 8. We smooth the trade data using a centered seven-month moving average to eliminate noise.
308
Barry Eichengreen, Andrew K. Rose, and Charles Wyplosz
Table 9.1
Evidence on Capital Controls during Speculative Attacks (Crises) Crises
Fiscal ratio Real exchange rate Inflation XlM Credit growth Money growth Interest rate Reserve growth
K-S
K-W
.26 .63 .01 .00 .17 .00 .81 .65
.68 .38 .05 .00 .10 .00 .56 .68
Noncrises
-.59 .78 -3.13 6.65 -1.47 -4.98 -.78 .17
K-S
K-W
.05 .00 .00 .00 .00 .00 .00 .25
.04 .00 .00 .00 .00 .00 .00 .41
-1.71 2.14 -4.10 6.63 -3.19 -4.35 -3.31 .77
Note: "K-S" denotes probability of rejection of null hypothesis (of equality of distribution across controls and absence of controls), using the nonparametric Kolomogorov-Smirnov test; a low value is inconsistent with the null hypothesis. "K-W" denotes probability of rejection of null hypothesis (of equality of distribution across controls and absence of controls), using the nonparametric Kruskal-Wallis test. "t" denotes a t-test of the null hypothesis of equality of first-moments across controls and absence of controls; a positive number indicates that the sample mean in the absence of capital controls is higher than the sample mean in the presence of capital controls. Throughout, a six-month exclusion window and a two-standard-deviation event delimiter are used.
for crises where controls were absent. Similarly, the rate of growth of domestic credit (relative to Germany) was faster and trade deficits larger when controls were in place, while inflation rates were higher and more variable. The role of capital controls is more striking still when we consider the noncrisis observations in the right-hand panel of table 9.1. We are able to reject the null of equal distributions and equal means for each variable except reserves and possibly budget deficits. Rates of growth of money and credit are faster, real overvaluation is greater, and budget and trade deficits are larger for countries not experiencing speculative attacks but with capital controls in place. This evidence is consistent with the view that controls made a difference. Countries with controls in place followed more expansionary monetary policies, as manifest in faster growth of money and credit and higher rates of inflation. One might expect to see the strongest evidence of the effectiveness of controls in the behavior of interest rate differentials and the growth of foreign exchange reserves, with countries applying controls enjoying lower interest rates and smaller reserve losses. In fact, interest rates appear to have been higher rather than lower, which may be explicable in.terms of the existence of a political risk premium in countries with controls in place. We cannot reject the null that the level of foreign exchange reserves was unaffected by the presence of controls. This may provide the key to understanding how countries utilize the instrument. Controls do not allow countries that pursue policies inconsistent with a peg to keep their exchange rate unchanged forever. They do not prevent attacks, nor do they permit countries to
309 Table 9.2
A Safe Passage to EMU? Joint Probabilities of Crises and Capital Controls
Noncrises Crises Total
No Controls
Controls
Total
165 (24) 21 (3) 186 (27)
345 (49) 171 (24) 516 (73)
510 (73) 192 (27) 702 (100)
Note: Chi-square test (1) test of independence = 33 (P = .00). Percentages are given in paren-
theses.
avoid reserve losses or interest rate increases when attacks occur. 9 Controls merely render expansionary monetary policies viable for a longer period by attenuating the link between crises and exchange rate regime collapse. 10 This characterization is corroborated by table 9.2. It reports the percentage of periods (for crises and noncrises) when controls were in place. It shows that the incidence of crises was proportionally higher when controls were present. A chi-square test confirms that this difference is statistically significant. In table 9.3, we shift our focus from "crises" to "events." "Crises" are identified by our index of speculative pressure irrespective of whether there has been a change in the exchange rate. An "event;' in contrast, corresponds to a realignment or a change in exchange rate regime. 11 The analysis of events in table 9.3 confirms the findings of table 9.1 above, strengthening the case that controls have a clear effect. Table 9.4 is an analogue to tables 9.1 and 9.3 above. It too reports a series of tests of the null hypothesis of equality of distributions of macroeconomic variables in the presence and absence of controls. But, unlike tables 9.1 and 9.3, which examine crises and events, table 9.4 looks at successful and unsuccessful attacks. A successful attack is a crisis that coincides with an event (more precisely, with the absence of a nonevent); an unsuccessful attack is a crisis that is not an event. The impression conveyed by table 9.4 is similar to that of table 9.1; capital controls are associated with significant differences in macroeconomic behavior, especially looser monetary policy. Table 9.5 is analogous to tables 9.1, 9.3, and 9.4 but conditions on the presence or absence of capital controls rather than testing for differences in distributions. Whereas tables 9.1, 9.3, and 9.4 condition on crises, events, and suc9. It would be nice to be able to compare the rate of reserve loss in the presence and absence of controls. But, since our data are monthly, we cannot differentiate between short and violent attacks of the kind likely to be associated with free capital mobility and the slower erosion of reserves that may take place in the presence of controls. 10. This effect of controls is modeled formally by Wyplosz (1986), who emphasizes the distinction between supporting an unviable exchange rate and lengthening the period between crises. 11. The realignment can be in either direction, and the change in regime can be associated with an appreciation or a depreciation.
310
Barry Eichengreen, Andrew K. Rose, and Charles Wyplosz
Table 9.3
Evidence on Capital Controls: Events and Nonevents Nonevents
Events
Fiscal ratio Real exchange rate Inflation X/M
Credit growth Money growth Interest rate Reserve growth
K-S
K-W
.01 .04 .08 .00 .12 .01 .07 .38
.01 .06 .01 .00 .15 .00 .50
-2.11 1.65 -3.19 4.56 -1.53 -3.77
Al
.30
-040
K-S
K-W
.38 .00 .00 .00 .00 .00 .00 .38
.13 .03 .00 .00 .01 .00 .00 .70
-1.39 1.38 -4.33 8.16 -2.59 -4.03 -3.60 .68
Note: "K-S" denotes probability of rejection of null hypothesis (of equality of distribution across controls and absence of controls), using the nonparametric Kolomogorov-Smirnov test; a low value is inconsistent with the null hypothesis. "K-W" denotes probability of rejection of null hypothesis (of equality of distribution across controls and absence of controls), using the nonparametric Kruskal-Wallis test. "t" denotes a t-test of the null hypothesis of equality of first-moments across controls (events) and absence of controls (nonevents); a positive number indicates that the sample mean in the absence of capital controls is higher than the sample mean in the presence of capital controls. Throughout, a six-month exclusion window and a two-standard-deviation event delimiter are used.
Table 9.4
Capital Controls and Successful and Unsuccessful Attacks Unsuccessful Attacks
Successful Attacks
Fiscal ratio Real exchange rate Inflation X/M
Credit growth Money growth Interest rate Reserve growth
K-S
K-W
.02 .07 .17 .00 .13 .01 .13 .30
.04 .11 .04 .00 .16 .00 .60 .39
-1.75 1.26 -2.79
4046 -1048 -3.46 -.26 .33
K-S
K-W
043
.46 .81 .06 .00 .06 .01 .18 .65
.84 .01 .00 .14 .01 .29 .73
-.83 .04 -3.24 5.71 -2.12 -3.49 -1.98 -.32
Note: "K-S" denotes probability of rejection of null hypothesis (of equality of distribution across controls and absence of controls), using the nonparametric Kolomogorov-Smirnov test; a low value is inconsistent with the null hypothesis. "K-W" denotes probability of rejection of null hypothesis (of equality of distribution across controls and absence of controls), using the nonparametric Kruskal-Wallis test. "t" denotes a t-test of the null hypothesis of equality of first-moments across controls and absence of controls; a positive number indicates that the sample mean in the absence of capital controls is higher than the sample mean in the presence of capital controls.
cessful attacks, table 9.5 conditions on the presence or absence of controls. It tests null hypotheses such as, "Successful attacks are different from unsuccessful attacks in the presence of controls." Controls again appear to make a difference in the sense that their presence is associated with statistically significant differences in the distributions of a number of macroeconomic variables. By comparison, differences in macroeconomic behavior are more unusual in the absence of controls.
311 Table 9.5
A Safe Passage to EMU? More on Capital Controls No Capital Controls
Capital Controls K-S Successful vs. Unsuccessful Attack: Fiscal ratio .08 Real exchange rate .65 Inflation .00 .06 XlM Credit growth .00 Money growth .87 Interest rate .00 .10 Reserve growth Crises vs. Noncrises: .00 Fiscal ratio .14 Real exchange rate Inflation .17 XlM .00 .04 Credit growth Money growth .87 .39 Interest rate Reserve growth .11 Events vs. Nonevents: Fiscal ratio .99 Real exchange rate .07 Inflation .00 XlM .92 Credit growth .02 Money growth .04 Interest rate .00 Reserve growth .00
K-W
K-S
K-W
.04 .71 .02 .03 .00 .66 .18 .93
2.18 -.16 -2.74 -2.12 -3.60 -.51 -.79 -.77
.21 .57 .15 .01 .23 .80 .18 .99
.17 .73 .47 .04 .13 .09 .09 .81
.96 -.60 -1.03 1.92 -1.85 -1.92 -1.92 -.85
.01 .28 .87 .00 .03 .97 .41 .72
-2.55 -.90 .86 3.43 2.16 .01 .01 1.96
.03 .11 .03 .25 .26 .00 .80 .17
.06 .15 .24 .19 .21 .01 .88 .99
-1.38 -1.02 .14 -1.49 .97 2.79 -.39 .66
.95 .17 .00 .99 .02 .11 .02 .09
.56 -.74 -3.05 -.05 -2.74 -2.09 -1.63 .23
.10 .24 .13 .03 .66 .98 .00 .21
.05 .25 .16 .06 .64 .95 .01 .93
1.93 -1.39 -1.23 1.77 -.75 -.20 -2.03 .13
Note: "K-S" denotes probability of rejection of null hypothesis (of equality of distribution across,
e.g., successful and unsuccessful attacks), using the nonparametric Kolomogorov-Smimov test; a low value is inconsistent with the null hypothesis. "K-W" denotes probability of rejection of null hypothesis of equality of distribution using the nonparametric Kruskal-Wallis test. "t" denotes a t test of the null hypothesis of equality of first-moments across, e.g., successful and unsuccessful attacks; a positive number indicates that the sample mean in the case of an unsuccessful attack is higher than the sample mean in the presence of a successful attack.
This body of evidence, taken together, is difficult to reconcile with the view that capital controls were ineffectual-that they were too easily evaded to provide authorities with significant policy autonomy in periods when exchange rates were pegged.
9.2 The Mechanics of Speculative Attacks In the last section, we reported evidence that capital controls make a difference. Here, we examine the channels of speculative activity in more detail in order to identify forms of intervention that are most likely to be effective in influencing the development of speculative attacks.
312
Barry Eichengreen, Andrew K. Rose, and Charles Wyplosz
9.2.1
An Illustration
The mechanics of currency speculation are described in Goldstein et al. (1993). Most transactions take the form of forward contracts-swaps and options, for example-rather than spot purchases and sales. As soon as one moves beyond partial equilibrium, however, it becomes obvious that virtually all such transactions involve the spot sale of the currency under attack coupled with borrowing in that currency. Consider an attack against the French franc. A firm or fund manager contracts with bank A to sell the franc against the deutsche mark forward. This transaction is shown in the balance sheet in table 9.6 for a forward rate of three francs/deutsche mark. Bank A now has a long open position in francs that it typically does not wish to take. It therefore sells forward to another bank (bank B) the francs it purchased from its customer, at the same time buying forward the deutsche marks it is obliged to deliver. While bank A is now hedged, bank B is in the same position as bank A at the previous step.12 Bank B will now attempt to cover its position by undertaking a similar transaction with another bank. There may be a series of such transactions. But the bank at the end of the chain (bank B, to keep matters simple) will still have to sell francs spot against deutsche marks. Since bank B must find the francs that it will sell spot, it must borrow them. (Typically, bank B will simultaneously borrow the francs and lend deutsche marks for one month to cover the maturity mismatch, but this is not essential to the argument.) As shown in table 9.6, bank B is hedged; it now holds in its portfolio the deutsche marks that it has contracted to sell to bank A and owes the francs that it is committed to buy. In this example, it makes no difference whether traders deal in derivatives and whether they are residents or nonresidents, aside from the fact that derivatives can be off balance sheet. Imagine that Monsieur Dupont, a French fund manager, buys on 15 June from his bank a Fr 100,000 European put option on the franc, to mature on September 1st. He can now sell francs and receive dollars. The bank selling this option is in the same position as bank A in the preceding example; it is committed to buying francs (normally against dollars) on September 1st. Hedging will therefore take the same form as before, with the bank borrowing Fr 100,000. Similarly, a swap is a combination of a spot and reverse forward transaction; Monsieur Dupont sells francs spot and buys them forward. His bank in effect lends him francs during that period and earns the rate of interest implicit in the forward discount. For present purposes, then, currency speculation can be described as being composed of the following elements. An agent takes an open position, usually against a bank. That position will have associated with it a spot sale of the 12. Bank A still faces the risk that its customer or bank B will not fulfill its contractual obligation, but this is not an exchange risk and is therefore not treated here.
313
A Safe Passage to EMU?
Table 9.6
Speculation
A. First Step Customer Assets Now 1 month
Bank A Liabilities
DM 100
Fr 300
Assets
Liabilities
Fr 300
DM 100
B. Second Step Customer Assets Now 1 month
Bank A Liabilities
DM 100
Fr 300
Assets
Liabilities
Fr 300 DM 100
DM 100 Fr 300
BankB Assets Now 1 month
Liabilities
Fr 300
DM 100
C. Last Step BankB
Now 1 month
Assets
Liabilities
DM 100 Fr 300
Fr 300 DM 100
currency under attack, forcing the central bank defending its currency peg to draw down its reserves. While the chain of subsequent transactions may involve different agents and financial instruments, it necessarily entails a loan of domestic currency originating in the home country. There are two places where currency can be obtained: from the banking system of the country in question (including its central bank) and from domestic-currency-denominated assets held abroad. This becomes apparent when it is acknowledged that any speculative attack necessarily entails the following transactions. Speculators first obtain from banks the currency that is to be sold on the spot market. Banks then borrow that currency on the money market. The only agent buying the currency in such periods is the central bank, which, in so doing, drains liquidity from the
314
Barry Eichengreen, Andrew K. Rose, and Charles Wyplosz
Table 9.7
Bank Lending and Reserve Movements
France,~ay-1\ugust 1992 France, September-October 1992 United Kingdom, July-September 1992
+ Net Bank Position
Foreign Exchange Losses
+28.7 +28.6 +13.0
-37.9 -21.0 -3.6
Note: "Net Bank Position" refers to foreign lending of domestic currency by domestic banks during period of speculative attacks (IFS line 11-16c). "Foreign Exchange Losses" refers to changes in net changes in foreign asset position of exchange reserves for commercial banks (sources: for France, changes in net position refers to short-term assets/liabilities and is from Banque de France, Bulletin Trimestriel, various issues; for the United Kingdom, Bank of England, Quarterly Bulletin). 1\11 figures are in U.S. $billion.
market. 13 If, to prevent interest rates from rising to politically unsupportable levels, the central bank sterilizes its exchange market operations and lends the domestic currency, it fuels additional speculation. Consolidating these transactions (canceling, among other items, interbank loans) reveals that what is left is domestic currency lending by the banking system to the rest of the world. The central bank lends on domestic markets to resident commercial banks, which lend to nonresidents. 9.2.2
Evidence
The importance of these transactions during periods of speculative activity is documented by table 9.7, which presents data for France and the United Kingdom during the 1992 ERM crisis. It is apparent that the net asset position in francs of French commercial banks increased by amounts broadly comparable to the reported foreign exchange losses of the Bank of France. 14 The role of banks as key players in periods of speculative crisis can be further documented by tracing the evolution of their portfolios. As episodes of speculative pressure, we again use the "crises" identified above. Figure 9.1 presents histograms depicting gross and net bank lending to nonresidents, distinguishing banks from nonbanks and gross from net lending. 15 We compare the rate of change of assets and liabilities during "crisis" and "noncrisis" peri13. The central bank may refuse to buy its currency spot. In that case, the exchange rate will depreciate, and the attack will succeed. 1\lternatively, the central bank may limit its loans to the banking system, and the interest rate will rise. This, the standard defense against a speculative attack, proved to be problematic during the E~S crises of 1992 and 1993, for reasons explored in Eichengreen and Wyplosz (1993) and Bensaid and Jeanne (1994). 14. For reasons discussed earlier, we know that published data on foreign exchange reserves are unreliable. We therefore checked fluctuations in reported reserves against the intervention data reported by 1\logoskofous (1993). 1\t $46 billion from July to 1\ugust 1992 and $228 billion from September to October 1992, these tell a consistent story. 15. These data come from the I~F and are open to the same limitations as those concerning central bank reserves (see above). Banks' assets are line 7a.d, their liabilities line 7b.d. We calculate net assets as line 7a.d minus line 7b.d. 1\ssets vis-a.-vis nonresident nonbanks are line 7ad.d, liabilities line 7bd.d. We calculate assets vis-a.-vis nonresident banks as 7ad.d minus line 7bd.d, and similarly for liabilities.
1
1
Crises
d
0
0
ooJt
Non-Crises
-.25
.3
.3
0
0
.5
i
-.5
-.25
Crises
0
I~~
.25
.5
Fig. 9.1
0
Non-Crises
ooorn
0
0 Crises
d
~ i~D
0 0
II]
Non-Crises
-1 1
2
Crises
I
0
1
Net Assets, Non-Banks
-1
2
0
-.4
-.2
u 0
.2
m~
Crises
-1.5
Non-Crises
0
1.
-1.5
D
Crises
0
=t
Gross Liabilities
0
.25
.15
-2
.5
0
.2
Gross Liabilities
0
.3
Net Assets, Non-Banks
-2
0
jo~
Non-Crises
-.2
Gross Assets, Banks
0
.25
.25
0
-.4
Gross Assets, Banks
0
.15
o0
2.5
0
0 2.5
.5
Net Assets
J.
-2.5
1
Net Assets
-2.5
.25
.3
o
.25
0
.25
.4
.4
1.5
1.5
0
Histograms of international liquidity movements
International Liquidity
-2
-1
0
~lrn
Non-Crises
1
-2
-1
II 0
Crises
-.25
0
D~&rn
Non-Crises
o
[fj
.25
2
-.25
~
0
Crises
nI~ ~i ~
.25
Gross Liabilities, Banks
0
.15
Gross Liabilities, Banks
0
.15
Net Assets, Banks
0
.15
.3
Net Assets, Banks
0
.15
.3
First-Differences o7f Monthly Natural Logarithms
Gross Assets, Non-Banks
o
Gross Assets, Non-Banks .25 i
-.5
o i
.25 ~
.25
.1'5
0
alanDo
Gross Assets
i 0 0 0 000 -.'3 -.15
a
.11
.2
.15
Non-Crises
0
~~JL,oo
-.15
Gross Assets
-.'3
o,
.1 i
.2
en en
0.J
:::>
01 0"1 0"1
S!r
0)
c:r>
0.J
~
~
z
en
0.J 0"1
;;;
en
en
0.J
~
s
.~"
/
r\
()
en en
0.J
C? :::>
'
--x-- Monthly increase in
net foreign assets of commerciaf banks denominated in Francs
,,,.'.
.
' _ ~' t ~~~ ¥~/ ;~,V/
/'
- - - - Monthly increase in net foreign assets of commercial banks denominated in foreign currencies
-200000
-150000
-100000
-50000
o
50000
100000
150000
200000
FRANCE
cn
~
Source: Bank of France
~
c::l
en OJ
,.......,
- : : ; - Monthly decrease in net foreign assets of Bank of France
cn
~
x.
~
'(6
5
~
Q.
'0 .::It. C
-
~X x~
c:
Q) (,)
to
co 0
~ Q)
0 0
~x
to
C
0)
'Q)
~
///
0 0
>(~
0
;jdV £66l
~ "-
C,
,=
~""""""
Q)
;jV~
/
0
£66l
to
~
(.) Q)
'"0
>
.>(
::c C 0
~
t
z
c.
en
CI'J
.::It. C
CO .Q
(ij
'0 Q)
E·
E' -
0
(.)
d3S Z66l
0 0
G) 0 0
to
C; 0)
'Q)
~
=
C
.;
Q)
rJ'J
~
,= rJ)
to
Q)
0
,=
>
::c C 0
~
Q..
~
= = ~
QI
CJ
~
~
.5
01.)
:a=
Z66l
= =
~ ~
~
= M
0 0
LO
0 0 0
0 0
0
LO
0 0
U(
0\
0 0 0
..I
SOl8Sdd JO SUO!"~8
~
~
318
Barry Eichengreen, Andrew K. Rose, and Charles Wyplosz
Table 9.8
The Behavior of International Liquidity during Crises Gross
Total system Banks Nonbanks Total liabilities Bank liabilities
K~S
K-W
.01 .74 .23 .00 .49
.01 .94 .47 .08 .61
Net
-2.33 .76 .59 2.12 -.25
K-S
K-W
.01 .56 .98
.02 .38 .92
-1.65 1.20 .73
Note: "K-S" denotes probability of rejection of null hypothesis (of equality of distribution across crises and noncrises) using the nonparametric Kolomogorov-Smimov test; a low value is inconsistent with the null hypothesis. "K-W" denotes probability of rejection of null hypothesis using the nonparametric test. "t" denotes a t-test of the null hypothesis of equality of first-moments across crises and noncrises; a positive number indicates that the sample mean in the absence of crises is higher than the sample mean during crises.
ods. In the upper-left-hand comer, for example, we present the distribution of growth rates of gross bank assets during tranquil periods and directly below during speculative attacks. Variability appears to be higher during attacks. Analogous differences are evident in the behavior of net assets but not gross liabilities. This is consistent with the view that banks are engaged in domesticcurrency lending to nonresidents during periods of speculative attack since, when we distinguish the position vis-a-vis nonresident banks and nonbanks, we see that the higher variability is attributable entirely to the gross asset positions of domestic banks vis-a-vis nonresident banks. 16 Table 9.8 provides Kolmogorov-Smimov, Kruskal-Wallis, and t-tests of the null that the variables depicted in figure 9.1 are identically distributed during crises and noncrises. The results indicate that total and net assets and liabilities have significantly different distributions during crises and noncrises. This is not true, however, of either bank or nonbank assets (or bank liabilities) separately. Figure 9.2 provides additional evidence for Spain and France during the 1992-93 EMS crises. The thick line shows the foreign exchange losses of the Bank of Spain. The various speculative episodes are evident, as is the reflux of reserves following each realignment. The thin line depicts foreign lending by banks-the increase in their net asset position vis-a-vis the rest of the world. It shows that reserve losses have as a counterpart commercial bank transactions. The figure for France presents Bank of France data that separate out bank loans according to their currency of denomination (francs vs. others). The comovement of commercial bank net lending in francs and foreign exchange reserve losses is unmistakable. 16. The IMF data do not discriminate between loans in domestic and foreign currencies. The preceding analysis of bank activities during attacks suggests that the surge of activity documented by the histograms is most likely to correspond to domestic currency loans.
A Safe Passage to EMU?
319 Table 9.9
Pension Funds' International Investments Country Australia Belgium Canada France Germany Hong Kong Ireland Japan Netherlands Switzerland United Kingdom United States
Value
% of Total Portfolio
11.8
24
.7
29
7.4
5.1 3.7
10 5 3
4.4
63
2.2 108.1 1.5 5.1 71.3
35 14 17 6 24 4
54.4
Source: Pension Fund Indicators, UBS Asset Management, London, April 1994.
We conclude that bank lending is a major channel through which currency traders obtain the assets that they sell during speculative attacks. It might be objected that there exists another source of these holdings, namely, those of nonbank agents, including households and firms. But households and firms require much of the money they hold for transactions purposes and lack the specialized knowledge of professional currency traders. The available data do not indicate much change in the money holdings of households and firms around the time of speculative attacks. What can be sold quickly, in principle, are the assets of pension funds and other institutional investors. It is difficult to ascertain the amounts held in different currencies by these entities. Table 9.9 provides the total value of nonlocal assets held by these funds. This $220 billion total is probably held mostly in U.S. dollars and German marks. Assume that 10 percent is held in French francs. If pension funds were to liquidate all their franc-denominated assets, this would represent sales of $22 billion. While this is a large amount, the assets of pension funds, once liquidated, cannot playa further role in speculative dynamics. Lending in domestic currency by banks, in contrast, can continue indefinitely so long as the central bank sterilizes its foreign exchange intervention. This is the distinction between an unlimited source of speculative capital and a one-time sale of assets.
9.3
Alleviating Speculative Pressure during the Transition to EMU
In earlier work, we argued that macroeconomic convergence was not a sufficient condition to preclude speculative crises affecting EMS currencies because of the possibility of self-fulfilling speculative attacks. Here, we have provided evidence consistent with the notion that capital controls are important for the timing and incidence of balance-of-payments crises. We have identified
320
Barry Eichengreen, Andrew K. Rose, and Charles Wyplosz
the channels that must be affected in order to contain speculative pressure. This section brings these elements together and draws out their policy implications. It analyzes the feasibility of restrictions on domestic-currency lending to nonresidents as a device for containing speculative pressure in the final stages of the transition to EMU. 9.3.1
The Problem of Self-Fulfilling Attacks
The Maastricht Treaty specifies conditions under which a country will qualify for participation in Europe's monetary union. One is that its exchange rate must remain within the "normal" ERM bands without being devalued for at least two years prior to entry. This means that, during the last two years of the transition, a balance-of-payments crisis that forces a country to devalue or to suspend its membership in the ERM effectively precludes its participation in EMU. To these worries, officials respond that countries need only adopt policies of convergence sufficient to ensure that their exchange rates are held within the normal ERM bands for the requisite period. The problem is that a commitment to policies of convergence and policy harmonization may not suffice to hold the exchange rate stable. This will be the case when there exists scope for self-fulfilling speculative attacks of the sort analyzed by Flood and Garber (1984) and Obstfeld (1986). In their models, even countries that are fully committed to exchange rate stability and have pursued policies consistent with the maintenance of stable rates may fall prey to speculative crises. I7 In theory, a central bank can discourage banks from lending to domestic or foreign residents by using the traditional instruments of monetary policy. It can limit the supply of loans relative to demand if it is willing to allow interest rates to rise. But, given the large capital gains available in short order in the event of a realignment, it may be necessary to allow interest rates to rise to very high levels, as illustrated by the case of Sweden in October-November 1992 and by Greece in May 1994. This may prove politically unsupportable and render a speculative attack self-fulfilling. The interest rate defense will therefore fail because the markets know that it is too costly to maintain. Consider a country willing to endure high interest rates and other forms of austerity now in return for qualifying for EMU later. Its past and current policies may be entirely consistent with the maintenance of exchange rate stability. If a speculative attack occurs, however, it will be forced to raise interest rates. The costs of austerity now rise relative to the benefits of EMU membership later, which may lead the government to conclude that the cost of qualifying for EMU is suddenly too high. Once it forsakes the lure of EMU membership, 17. Eichengreen and Wyplosz (1993), Obstfeld (1994), and Eichengreen, Rose, and Wyplosz (1994) suggest that evidence from recent ERM crises is not inconsistent with the predictions of these models.
321
A Safe Passage to EMU?
it has no reason to resist shifting policy in a less austere direction; and the markets, aware of its incentives, have reason to attack. Note that the shift in policy in a more expansionary direction is contingent; there is no reason for it to occur in the absence of the attack. In this setting, in other words, speculative attacks can be rational and self-fulfilling. Eichengreen and Wyplosz (1993) show that there is some evidence of these dynamics in 1992-93. The implication is that the Treaty of Maastricht may fail even if countries intend to follow macroeconomic policies fully consistent with its letter and spirit. The question, to which we now tum, is whether it might be possible to reduce the odds of this happening by throwing sand in the wheels of international finance. 9.3.2
A Proposal
The analysis of section 9.2 can be summarized by the observation that speculative attacks start with the opening of a position and end with a loan denominated in the currency under attack. Discouraging position taking might appear to be a promising approach to dealing with the problems that result. But positions can be booked anywhere in the world so long as domestic currency transfers are possible at low cost. If France were to impose a tax on foreign exchange transactions in Paris, for example, it would be easy to shift francs to London and carry out the same transactions there. A solution is to make use of the fact that all speculative sales must be matched by fresh provision of the currency under attack. Except for francs made available by the liquidation of existing offshore asset positions, which are by definition of limited size, the rest comes from new lending by French financial institutions-hence the idea to impose an explicit or implicit tax on domestic-currency lending to nonresidents. The interest rate defense discourages speculation by making it expensive. This can equally be done by imposing a deposit requirement on domestic loans to nonresidents in domestic currency. The deposit could be proportional to the loan and would have to be maintained interest free at the central bank for the duration of the loan or for a fixed period. While the cost, in the first instance, is borne by the lending bank, part of it will be passed along to potential borrowers. A useful feature of this measure is that the opportunity cost of the noninterest-bearing deposit increases with the interest rate, which will rise during periods of speculative pressure. The interest rate defense will now be more powerful since it will not only increase the traditional interest parity threshold (at which the expected devaluation matches the interest differential) but also impose a cost on position taking. This proposal is open to obvious .criticisms. For one, any disruption to the free flow of capital has allocative and distributional costs. In the present case,
322
Barry Eichengreen, Andrew K. Rose, and Charles Wyplosz
however, these are likely to be small because long-term capital flows will be little affected. While lending to nonresidents will become more expensive, the additional cost, when spread over a long maturity, will be limited. 18 Non-interest-bearing-deposit requirements on bank lending to nonresidents are equivalent to an implicit widening of the exchange rate band. To illustrate, assume that the lower end of the French franc/deutsche mark band is at a rate of one (one hundred French francs per one hundred deutsche marks). 19 But, if the cost of the non-interest bearing deposit requirement passed along to the customer is ten francs per deutsche mark, this shifts the lower edge of the band to ninety. If the cost of the non-interest-bearing deposit is the equivalent of widening the band, why then not simply widen the band and avoid interfering with the operation of capital markets? The answer is that, by altering the incentives for the authorities to defend the currency peg, non-interest-bearing deposit requirements increase the exchange rate stabilizing effect featured in models of exchange rate target zones. Because deposit requirements introduce a wedge between on- and offshore rates, they reduce the cost to the authorities of using the interest rate to defend the peg. The knowledge that the authorities are more likely to defend the edge of the band reduces the incentive for speculators to test it. One might object that a policy that discriminates against loans to nonresidents runs counter to Article 73f of the Maastricht Treaty. Foreigners could protest an implicit tax not also levied on domestic borrowers. There is ambiguity about the proper interpretation of Article 73f, however, since the treaty allows temporary measures in case of emergency. 20 Nevertheless, the best .response would be explicitly to authorize such a measure during the remainder of Stage II. The treaty provides for an Intergovernmental Conference in 1996 to modify provisions that have proved undesirable. The IGC could provide the amendments required for the temporary establishment of deposit requirements when and where needed to protect the ERM and therefore ensure that the goals of the Maastricht process are achieved. Then there is the question of coverage. Could the measure be rendered ineffective by the diversion of domestic-currency loans to channels not covered by the deposit requirement? Recent Spanish experience illustrates the danger. 21 Between September and November 1992, the Bank of Spain imposed a measure similar to the one contemplated here. It applied a deposit requirement on new lending by banks to nonresidents through swaps. Swaps are the normal vehicle for short-term speculative lending; exempting lending for other pur-
18. For example, the cost of a ten-year loan will be increased by a tenth assuming that the interest rate is constant and the yield curve flat. 19. The example that follows is drawn from Garber and Taylor (1994). 20. It is unclear whether the treaty in fact rules out a scheme like that proposed here. Absent an amendment to the treaty that addressed this issue head on, the question of Maastricht compatibility would have to be adjudicated in the European Court of Law. 21. For a description, see Linde (1993) and Linde and Alonso (1993).
323
A Safe Passage to EMU?
18
17
15
14
13 07-00t
M-Oot
20-0ot
28-00t
3O-00t
OS-Nov
12-Nav
Ie-Hov
Fig.9.3 Spain: internal and external swap (one-day) interest rates, October-November 1993 Source: Linde (1993).
poses was meant to shield nonspeculative activity. The measure succeeded in discouraging speculation for a few days but then lost its effectiveness. Figure 9.3 shows the differential between domestic and off-shore interest rates on swaps in pesetas during this period. Within a week of the imposition of the deposit requirement, the differential fell to less than one hundred basis points, too small to deter speculation given the magnitude of the depreciation that was anticipated. Conversations with regulators and traders in Madrid and London have convinced us that there never was a scarcity of pesetas because Spanish banks sent pesetas to their London subsidiaries to circumvent the deposit requirement (see Freitas de Oliveira 1994). Thus, limiting the measure to lending to finance transactions in one instrument, even if the latter is the most widely used under normal circumstances, will prove futile since currency traders will shift to other instruments in response to the policy. Accordingly, deposit requirements must be applied to all domestic-currency loans to all nonresidents. Finally, there is the question of avoidance. Even if the measure applies to all bank lending to nonresidents, new nonbank mechanisms for channeling domestic currency offshore may be established in response to the imposition of a deposit requirement on lending to nonresidents. A French bank required to make non-interest-bearing deposits when lending francs to nonresidents could lend francs to French corporations, which in turn could lend them to nonresidents (including their own nonresident operations or nonresident branches of the initiating French bank). This raises the possibility that a scheme that started out as a deposit requirement on loans to nonresidents would be broadened into a deposit requirement on all loans extended through certain windows and, if lending was diverted to other windows, on all bank lending, which is surely undesirable.
324
Barry Eichengreen, Andrew K. Rose, and Charles Wyplosz
The extent of evasion is likely to depend on the length of the period for which the deposit requirement remains in effect. If that period is short, it may not pay to set up the back channels required for evasion. Firms may be unwilling to incur the costs of avoidance if the benefits are transitory; as Dixit (1991) has shown, even relatively small fixed costs can have potentially large effects on real and financial behavior. Hence, non-interest-bearing deposit requirements are most likely to be effective if their imposition is limited to the last two years of the transition to EMU. Clearly, no measure of the sort we describe here is ever 100 percent effective. It is important to note, however, that to slow down speculative activity and provide time for orderly realignments it is not necessary for the measure to be watertight. 22 The historical record indicates that capital controls have had measurable effects on macroeconomic activity even when they were less than totally effective.
9.4
Conclusion
Retrospective evidence on capital controls in section 9.1 verified that these measures affected the course of macroeconomic developments, contrary to the presumption that they were too easily evaded to have a discernible effect. Prospective analysis in section 9.3 suggested that it might be possible to simulate their effects for a transitional period by imposing non-interest-bearing deposit requirements on bank lending to nonresidents. We cannot emphasize too strongly that we conceive of this device as a temporary measure to be applied during the transition to monetary union in Europe. It is a third-best solution to which one is driven only if first- and secondbest responses are ruled out and the goal of EMU is taken as given. In Europe, where pegging exchange rates within normal bands for at least two years is a prerequisite for completing the transition to monetary union, such measures may be justified by the considerable efficiency advantages of the Single Market Program, whose political viability appears to hinge in turn on the establishment of a single currency. One of the "convergence criteria" of the Maastricht Treaty mandates that countries hold their exchange rates within their normal fluctuation bands for two years without experiencing "exceptional tensions." Even if this provision is interpreted as allowing countries to realign in response to speculative pressures not of their own making without being disqualified from participating in EMU, measures like those described here would be 22. Fieleke (1994) dismisses as ineffectual the capital controls applied by Ireland, Spain, and Portugal in 1993 on the grounds that "all three countries were obliged to devalue within months after imposing or intensifying controls." Leaving aside whether these countries' controls were well designed, this criticism misses the point that these three countries were all able to realign and stay in the ERM, whereas countries that did not apply controls, like Italy and the United Kingdom, were driven out of the system.
325
A Safe Passage to EMU?
needed to provide time for the multilateral consultations that must precede orderly realignments and to prevent self-fulfilling attacks from driving currencies out of the ERM.23 Non-interest-bearing deposit requirements on lending to nonresidents are not the first-best mechanism for completing the transition. The smoothest way of reaching that goal is to move there directly. Suppose that financial market participants awoke one Monday to the news that a subset of ED countries had formed a monetary union over the weekend, that the European Monetary Institute had been transformed into the European Central Bank, and that the latter would henceforth be the sole issuer of the participating countries' currencies, which it stood ready to exchange for one another at par. Transitional problems would be ruled out by ruling out the transition. In practice, however, this outcome is unlikely for political reasons. Germany insisted on the three-stage transition process of the Maastricht Treaty and the convergence criteria embedded in its protocol on monetary union precisely in order to rule out abrupt action. The second-best solution is to declare wide bands like those of the post-July 1993 EMS the "normal bands" referred to in the protocol and to move to monetary union after a subset of ED countries have held their currencies within bands of ± 15 percent for two years. This assumes, of course, that the difficulty of holding exchange rates within 15 percent bands is qualitatively different from holding them within 2 1/4 percent bands. The longer the ERM's new fluctuation bands have gone untested, the more confident European policymakers have become of this assumption. But there is reason to think that their confidence is unfounded-that an oil shock, a recession, or an electoral surprise could quickly cause wide bands to bind. Experience with floating exchange rates in the 1970s and 1980s showed that cumulative bilateral nominal exchange rate movements of 15 percent over a period of two years are not uncommon. The implication is that the Treaty of Maastricht can fail even if countries adopt macroeconomic policies consistent with its letter and spirit. And these dangers will certainly intensify in the runup to Stage III. Political brinkmanship will grow as the deadline nears, heightening doubts that exchange rates are really locked. 24 The markets will have good reason to anticipate last-minute realignments motivated by attempts to boost competitiveness before parities are locked in (Froot and Rogoff 1991). Any of these factors could defeat efforts to hold ERM currencies within 15 percent bands. Furthermore, German officials (who insisted on the convergence criteria to 23. It is useful to recall that the EMS has never lost a member as a result of a speculative attack so long as its weak currency countries were operating under capital controls. 24. For example, the German Constitutional Court has ruled that the final decision to go ahead with monetary unification belongs to the Bundestag. It is easy to guess how the markets will react if there is an off chance that the Bundestag is headed toward a negative vote.
326
Barry Eichengreen, Andrew K. Rose, and Charles Wyplosz
force their potential EMU partners to demonstrate their willingness to live with the consequences for macroeconomic policy of monetary union) are unlikely to regard 15 percent bands as a sufficiently stringent test of policymakers' resolve. 25 One might raise the same objection to the imposition of non-interestbearing deposit requirements on bank lending to nonresidents; these measures are tantamount to an implicit widening of the band in that they relax the external constraint on domestic policy. The difference is that non-interest-bearing deposit requirements bind only in periods of speculative pressure. The rest of the time, governments will have ample opportunity to demonstrate their commitment to the policies mandated by the Maastricht Treaty. A final objection to the proposal is that deposit requirements will weaken monetary discipline. Governments insulated from the discipline imposed by international financial markets may embark on policies which further destabilize exchange rates. That there exists the potential of moral hazard is clear from the analogy between our proposal and the standard argument for insurance: deposit requirements could ensure the EU against policy mistakes that would otherwise derail Stage II of the Maastricht process. If one thinks that the costs of failure are high, then an investment in insurance is justified. But, just as any sensible insurance company should monitor the behavior of its policyholders, the EU should monitor the behavior of governments receiving "deposit [requirement] insurance." Fortunately, it already has the appropriate mechanisms in place: the European Monetary Institute and the Monetary Committee, which are authorized to surveil the policies of EU countries, recommend corrective action, and levy penalties against governments that fail to comply. European policymakers will be inclined to shy away from any recommendation that entails amending the treaty. This "don't open the Pandora's box" mentality fails to come to grips with the lack of viability of the current Maastricht blueprint for completing the transition to monetary union. If, as we argue, an extended period of exchange rate stability within narrow bands is not feasible, then some provision of the treaty must be changed for the goal of monetary union to be achieved. One option is to add further safeguards sufficient for Germany and other reluctant participants to commence with monetary union immediately. Another is to accept the wide bands of the post-1993 EMS as the normal bands referred to in the protocol on monetary union, although gaining the agreement of these same reluctant countries will again require additional safeguards. Still another option is to authorize the temporary imposition of deposit requirements on lending to nonresidents. One way or another, the treaty will have to be revised. Of course, one can insist on a policy of "none of the above." But the implication is that the goal of European monetary unification will never be achieved.
25. The German Constitutional Court has also ruled that the Maastricht Treaty's so-called convergence criteria must be interpreted strictly, which throws into question the realism of this strategy.
327
A Safe Passage to EMU?
References Alogoskofous, George. 1993. The crisis in the European Monetary System and the future of EMU. Paper prepared for the Conference on the Impact of the European Monetary Unit, Barcelona. Bensaid, Bernard, and Olivier Jeanne. 1994. The instability of fixed exchange rate systems when raising the nominal interest rate is costly. Ecole Nationale des Ponts et Chaussees. Typescript. Dixit, Avinash. 1991. Irreversible investment with price ceilings. Journal of Political Economy 99:541-57. Eichengreen, Barry, Andrew Rose, and Charles Wyplosz. 1994. Speculative attacks on pegged exchange rates: An empirical exploration with special reference to the European Monetary System. University of California, Berkeley, and INSEAD. Typescript. Eichengreen, Barry, and Charles Wyplosz. 1993. The unstable EMS. Brookings Papers on Economic Activity, no. 1:51-143. European Parliament. Committee on Economic and Monetary Affairs. Monetary Subcommittee. 1993. Roumeliotis report. Strasbourg. Fieleke, Norman S. 1994. International capital transactions: Should they be restricted? New England Economic Review (March/April), 28-39. Flood, Robert, and Peter Garber. 1984. Collapsing exchange rate regimes: Some linear examples. Journal ofInternational Economics 17: 1-14. Freitas de Oliveira, Luis F. 1994. Deposit requirements as an alternative to curb speculative attacks in the EMS: A view of the Spanish experience and market reactions. INSEAD. Typescript. Froot, Kenneth, and Kenneth Rogoff. 1991. The EMS, the EMU, and the transition to a common currency. NBER Macroeconomics Annual 6:269-317. Garber, Peter, and Mark Taylor. 1994. "Sand in the wheels" policies in foreign exchange markets. Washington, D.C.: International Monetary Fund. Typescript. Giovannini, Alberto. 1989. How do fixed exchange rate regimes work? Evidence from the gold standard, Bretton Woods and the EMS. In Blueprints for exchange rate management, ed. Marcus Miller, Barry Eichengreen, and Richard Portes. New York: Academic. Goldstein, Morris, David Folkerts-Landau, Peter Garber, Liliana Rojas Suarez, and Michael Spencer. 1993. International capital markets, part I: Exchange rate management and international capital flows. Washington, D.C.: International Monetary Fund. Gros, Daniel. 1987. The effectiveness of capital controls: Implications for monetary autonomy in the presence of incomplete market separation. IMF Staff Papers 34:621-42. Gros, Daniel, and Niels Thygesen. 1992. European monetary integration from the European Monetary System to the European Monetary Union. London: Macmillan. Linde, Luis M. 1993. Las medias del Banco de Espana de septiembre y octubre de 1992 penaliando la especulacion cambiaria. Papeles de Economia Espanola no. 54. Madrid: Bank of Spain. Linde, Luis M., and Javier Alonso. 1993. Currency market and foreign exchange crises: A note in connection with the Group of Ten report of April 1993. Madrid: Bank of Spain. Typescript. Meese, Richard, and Kenneth Rogoff. 1983. Empirical exchange rate models of the seventies: Do they fit out of sample? Journal ofInternational Economics 14:3-24. Mussa, Michael, and Morris Goldstein. 1994. The integration of world capital markets. In Changing capital markets: Implications for policy, ed. Federal Reserve Bank of Kansas City. Kansas City, Federal Reserve Bank of Kansas City.
328
Barry Eichengreen, Andrew K. Rose, and Charles Wyplosz
Obstfeld, Maurice. 1986. Rational and self-fulfilling balance-of-payments crises. American Economic Review 76:72-81. - . - - . 1994. The logic of currency crises. Working Paper no. 4640. Cambridge, Mass.: National Bureau of Economic Research. Ozkan, ,F. Gulcin, and Alan Sutherland. 1994. A model of the ERM crisis. University of York. Typescript. Truman, Edward. 1994. Review of A retrospective on the Bretton Woods system. Journal of Economic Literature 32:721-23. Wyplosz, Charles. 1986. Capital controls and balance of payments crises. Journal of International Money and Finance 5: 167-79.
Comment
Jose Vifials
This is the latest in a series of very interesting papers written by the authors in the past few years that have enhanced our understanding of why foreign exchange market crises arise and of what can be done to preserve exchange rate stability. In the paper, the authors make two main points: first, that capital controls are effective in influencing key macroeconomic variables during prolonged periods of time; second, that a nonremunerated deposit on bank lending to nonresidents could mimic the main effects of capital controls and therefore be a useful tool for preserving EMS stability during the transition to EMU. In my comments, I will attempt to argue that the evidence presented by the authors is not as favorable as they claim to the alleged effectiveness or desirability of capital controls. I will also suggest that there are many reasons to suspect that their "Tobin tax"-type proposal for enhancing the future stability of the exchange rate mechanism (ERM) of the EMS is likely to be either ineffective or extremely distortionary from an economic viewpoint. How Effective Are Capital Controls? Section 9.1 of the paper presents empirical evidence on the effectiveness of capital controls during periods when currencies were pegged in an explicit manner. In particular, the observations are drawn from the Bretton Woods system, the Snake, and the EMS and correspond to DEeD countries. Instead of performing the traditional tests of covered interest rate differentials or onshore-offshore interest rate differentials to assess the effectiveness of capital controls, the authors use a more refined methodology, already applied in an earlier paper (Eichengreen, Rose, and Wyplosz 1994). In particular, they look at the behavior of certain key macroeconomic variables during both tranquil (what they call "noncrises") and turbulent (what they call "crisis") periods in foreign exchange markets and perform nonparametric tests to assess whether
Jose Vifials is head of economic studies at the Bank of Spain and a research fellow of the Centre for Economic Policy Research.
329
A Safe Passage to EMU?
the behavior of the chosen variables differs significantly depending on whether there are capital controls. After examining the empirical evidence, the authors conclude that the behavior of a number of macro variables (i.e., inflation, the money stock, the trade balance, and the real exchange rate) seems to differ with and without capital controls. They interpret this finding as supportive of the hypothesis that controls are effective in providing the authorities with some room for policy autonomy. However, my reading of their evidence is quite different. Indeed, in the light of the proposal that they make in section 9.1, the key question to ask is to what extent capital controls increase the capacity of the authorities to defend the central parity in bouts of turbulence so that they avoid having to raise interest rates to extremely high levels or deplete so fast the stock of foreign reserves. In other words, the key test on the effectiveness of controls would be to check whether the behavior of interest rates and foreign reserves differs during turbulent times in cases when controls are in place relative to cases when they are not. Consequently, when we go to table 9.1 and look at the behavior of interest rates and foreign reserves, what do we find? Surprisingly-in view of the optimism of the authors about the effectiveness of controls-what we observe from looking at the last two rows during turbulent times is that the behavior of interest rates and foreign reserves is not significantly different with and without capital controls! In particular, both variables are distributed similarly, and their mean values are not statistically different in both instances. In sum, I do not find any evidence suggesting that capital controls have provided for a smoother behavior of interest rates and foreign reserves in times of strong exchange market pressures. While it is true that the authors find that controls seem to matter for the evolution of other macroeconomic variables at such times, this is not what is relevant for answering the key question: Do controls facilitate the defense 'of the central parity by the authorities in turbulent times? In my view, their own evidence about the behavior of interest rates and foreign reserves suggests that the answer should be no. This does not mean, however, that controls do not matter. In fact, they do, although in a manner that I would describe as being far from economically desirable. In particular, if we look again at table 9.1, it is evident that, across tranquil and turbulent periods alike, capital controls tend to be associated with both higher inflation and higher trade deficits. In other words, capital controls and higher domestic and external imbalances go hand in hand as a result of the pursuit of more expansionary policies. And, while it is not possible to draw causal implications from observed correlations, my impression is that those countries that resort to capital controls do so in order to be able to run more expansionary policies. Finally, since such policies are associated with fundamental imbalances, capital controls eventually lead to exchange market turbulence and to unavoidable devaluations. All in all, the evidence presented by the authors is not inconsistent with the
330
Barry Eichengreen, Andrew K. Rose, and Charles Wyplosz
view that controls do not avoid the exacerbation of policy dilemmas incurred by the authorities at times of turbulence when defending the currency, especially when "self-fulfilling" attacks occur. In addition, over the medium term, capital controls seem to facilitate the pursuit of inflationary and divergent economic policies, which eventually provoke fully justified attacks on currencies, causing exchange market instability. In short, capital controls do not seem to facilitate the defense of exchange rate stability in the short term but contribute to undermine it in the medium term through the relaxation of policy discipline and coordination.
A Proposal to Save the ERM The authors believe that, since the ERM is exposed to "self-fulfilling" attacks, policy coordination and convergence are necessary but not sufficient conditions to preserve exchange rate stability. They also believe that defensive interest rate increases are not credible or effective and thus that the ERM is bound to fail unless something else is done. Their proposed solution is to impose a nonremunerated deposit on bank lending to nonresidents in order to close the main channel through which speculators obtain the weak currency in turbulent times. Besides it being highly debatable whether such a proposal is consistent with the legal provisions of the Treaty of Maastricht, I think that it poses several important economic and practical problems. The way in which I understand their "Tobin tax"-type proposal (see Tobin 1978) is by analogy with a tax on alcohol. Let us assume that the government believes that the consumption of alcohol should be reduced. It would be more efficient and effective to raise the tax on alcohol rather than to raise all consumption taxes. The logic for their proposal is similar: to defend the currency, it is preferable to design measures specifically aimed at "taxing" speculators rather than increasing the general level of interest rates, which penalizes speculators and nonspeculators alike and may be very costly in macroeconomic terms. In practice, however, the high degree of integration of national financial markets and the sophistication of market operators is very likely to render the Eichengreen-Rose-Wyplosz proposal ineffective very quickly, as the authors themselves seem to acknowledge. On the one hand, taxing only the loans granted by domestic banks in domestic currency would immediately provoke transfers of money to foreign branches, which, in turn, would rechannel the money to the ultimate borrowers. On the other hand, taxing only the loans extended by domestic banks to nonresidents would lead to the establish.ment of nominal. nonbank resident borrowers, which, in turn, would rechannel the money to nonresident borrowers. Consequently, if one wants the measure to be effective, one would have to be prepared to close all possible loopholes bytemporarily-taxing all bank credit on domestic currency. If this is the case, the proposed "tax" measure would be very disruptive since
331
A Safe Passage to EMU?
it would end up affecting all potential borrowers, thus negatively affecting economic activity, and eroding the single market in trade and financial services. In terms of the above-mentioned alcohol analogy, one would end up taxing consumption on all goods in order to discourage the consumption of alcohol. To conclude, while I agree with the authors that there are sources of potential instability in the ERM given its nature of a "fixed but adjustable" exchange rate arrangement and that something ought to be done about it, they have failed to convince me that (a) capital controls have been as effective in the past as they claim in helping defend exchange rate stability in the shorter-term, (b) that the medium-term effect of controls is economically desirable or favorable to exchange rate stability, and (c) that their "Tobin tax"-type proposal will be either effective or economically desirable as a way of stabilizing the ERM in future years. Obviously, there is no easy way out for preserving the stability of the ERM. Nevertheless, rather than imposing capital controls or other administrative measures to discourage capital flows, what ought to be done is to strengthen policy coordination in the monetary and nonmonetary fields so as to keep exchange rates broadly in line with fundamentals and to enhance the effectiveness of the mechanisms for preserving exchange rate cohesion in the ERM in those cases when it is threatened by speculative attacks. Restricting capital flows addresses just the symptoms but not the causes of underlying disturbances, and there is the serious risk that the presence of such restrictions may even exacerbate them over the medium term by relaxing policy discipline and by weakening international policy coordination. References Eichengreen, B., A. Rose, and C. Wyplosz. 1994. Speculative attacks on pegged exchange rates: An empirical exploration with special reference to the European Monetary System. Paper presented. at the European Summer Symposium in Macroeconomics, sponsored by the Centre for Economic Policy Research, Roda de Bani, May. Tobin, J. 1978. A proposal for international monetary reform. Eastern Economic Review 4, nos. 3-4 (July-October): 153-59.
Contributors
Fabio Bagliano Universita di Torino Dipartimento di Scienze Economiche e Finanziarie C.so Unione Sovietica 218bis 10134 Torino, Italy Andrea Beltratti Universita di Torino Dipartimento di Scienze Economiche e Finanziarie C.so Unione Sovietica 218bis 10134 Torino, Italy
Bernard Dumas Department of Finance HEC School of Management 78351 Jouy-en-Josas Cedex, France Barry Eichengreen Department of Economics 540 Evans Hall University of California Berkeley, CA 94720 Mark D. Flood Department of Finance Concordia University 1455 De Maissonneuve Blvd. West Montreal, QC H3G 1M8, Canada
Giuseppe Bertola Dipartimento di Scienze Economiche e Finanziarie Universita di Torino C.so Unione Sovietica 218bis 10134 Torino, Italy
Robert P. Flood Research Department International Monetary Fund 700 19th Street NW Washington, DC 20431
Lorenzo Bini-Smaghi European Monetary Institute Kaiserstrasse 29 Frankfurt a.M. D-60311, Germany
Silverio Foresi Department of Finance Stem School of Business New York University 44 West 4th Street New York, NY 10012
Zhaohui Chen London School of Economics University of London Houghton Street London WC2A 2AE, United Kingdom
Jeffrey Frankel University of California Department of Economics 549 Evans Hall, Room 3880 Berkeley, CA 94720
333
334
Contributors
Giampaolo Galli Direttore Centro Studi Confindustria Viale dell' Astronomia 30 00144 Roma, Italy
Richard K. Lyons Haas School of Business 350 Barrows Hall University of California Berkeley, CA 94720
Peter M. Garber Department of Economics, Box B Brown University 64 Waterman Street Providence, RI 02912
Antonio Mello Banco de Portugal Rua Febo Moniz, 4 1100 Lisboa, Portugal
Alberto Giovannini Lungotevere Arnaldo da Brescia, 11 00197 Roma, Italy Charles Goodhart London School of Economics University of London Houghton Street London WC2A 2AE, United Kingdom David A. Hsieh The Fuqua School of Business Duke University Durham, NC 27705 Takatoshi Ito Research Department International Monetary Fund 700 19th Street NW Washington, DC 20431 Philippe Jorion Graduate School of Management University of California, Irvine Irvine, CA 92717 Alan Kirman Department of Economics European University Institute 1-50016 San Domenico di Fiesole Florence, Italy Allan W. Kleidon Cornerstone Research 1000 El Camino Real Menlo Park, CA 94025
Richard Payne London School of Economics Financial Markets Group University of London Houghton Street London WC2A 2AE, United Kingdom William Perraudin Department of Economics Birkbeck College University of London 7-15 Gresse Street London W1P 1PA, United Kingdom Andrew K. Rose School of Business Administration 350 Barrows Hall University of California Berkeley, CA 94720 Michael G. Spencer Research Department IS 12-808 International Monetary Fund Washington, DC 20431 Antti Suvanto Bank of Finland Monetary Policy Department PO Box 160 FIN-00101 Helsinki, Finland Lars E. O. Svensson Institute for International Economic Studies Stockholm University 10691 Stockholm, Sweden Mark P. Taylor Department of Economics University of Liverpool Liverpool L69 3BX, United Kingdom
335
Contributors
Jose Vinals Banca de Espana Alcala, 50 28014 Madrid, Spain Paolo Vitale 308 Kings College Cambridge CB2 1ST, United Kingdom
Charles Wyplosz INSEAD Boulevard de Constance 77305 Fontainebleau Cedex, France
Author Index
Abramovitz, M., 81n4 Adams, P., 30n12 Adler, Michael, 236n5, 263 Admati, A. R., 41, 42n1, 43, 47, 54, 55-58, 65, 137, 139, 153, 184 Allen, Helen, 13, 286 Alogoskoufous, George, 314n14 Alonso, Javier, 322n21 Amihud, Y., 62
Bollerslev, Tim, 6, 24, 27, 42, 43, 44n4, 45, 48,54,56-57,124, 129n19, 134, 137, 155, 190 Boothe, Paul, 6 Bossaerts, P., 73 Brady Commission, 8, 209n1 Branson, William H., 268n5, 269 Brock, W. A., 55n13,56,61n18 Burnham, J. B., 44n4, 47
Backus, David K., 267 Bagehot, W. (pseud.), 54 Baillie, R. T., 124 Banerjee, A., 104 Bank for International Settlements (BIS), 108n2, 180, 211n3, 212, 213nn8,9,218, 220-21n17 Bank of England, 211n3, 314t Bank of Italy, 246 Banque de France, 314t Barclay, M. J., 42n1 Batten, J., 23 Baxter, Marianne, 273, 285 Beckers, S., 29 Bensaid, Bernard, 314n13 Bessembinder, H., 6, 24, 112, 129n19, 134, 153, 155 Bhar, R., 23 Biais, B., 75 Bikhchandani, S., 104 Bilson, John F. 0., 267, 276 Black, Fisher, 29 Blitz, J., 107, 108
Canina, L., 35 Chan, K. C., 62 Chesney, M., 30 Cheung, Yin-Wong,264 Chiang, Raymond, 213n8 Christie, W. G., 62 Clark, P., 22 Cohen, K., 130n20 Commodity Futures Trading Commission (CFTC),211n3 Cookson, Richard, 215 Copeland, T., 76 Cornell, B., 22, 25 Crnja, Z., 60n15 Cumby, Robert E., 281n18 Curcio, R., 129n19
337
Dacorogna, M. M., 140 Decupere, Danny, 13 De Grauwe, Paul, 13 De Long, J. Bradford,S Demos, A., 114 Demsetz, H., 203
338
Author Index
Deutsche Bundesbank, 211n3 Dickey, D., 29 Dimson, E., 135 Dini, Lamberto, 11 Dixit, Avinash, 324 Dominguez, Kathryn M., 269 Domowitz, Ian, 6, 42, 43, 44n4, 45, 48,54, 56-57,108,124,137,190,276 Donders, M. W. M., 40 Dornbusch, Rudiger, 236n5, 265, 266n3, 267, 281n18, 285 Duan, J.-C., 30 Dumas, Bernard, 213n8, 236 Easley, D., 25, 38nl, 79,86, 183, 198 Ederington, L. H., 40, 125 Eichengreen, Barry, 232n2, 246, 303, 305, 307,314n13,320n17,321,328 Engle, R., 27 Engle, Robert E, 286n21 European Parliament, 303 Fama, Eugene, 275,286 Federal Reserve System: Board of Governors, 211n3; Federal Reserve Bank, New York, 194 Fialkowsky, D., 114 Fieleke, Norman, 5, 324n22 Figlewski, S., 35 Figliuoli, L., 5, 44n4, 45n7, 89 Flood, M. D., 73, 110, 184nl Floo~RobertP,266n3,285,285-86n20,320
Folkerts-Landau, David, 211n3, 212n5 Foster, ED., 43n3, 56, 153 Frankel, Jeffrey, 4,5,11,13,24,25,59,60, 104, 231,236n5,261, 264, 267,268, 269,272,274nl1, 276,285,286 Freedman, R., 42nl Freitas de Oliveira, Luis E, 323 French, K., 5, 156 Frenkel,Jacob,39,263,264,267,281n18 Friedman, Milton, 274 Proot, Kenneth, 4,5, 13, 24, 59, 104,275, 276,285,286,325 Fuller, A., 29 Galai, D., 76 Garbade, Kenneth D., 99 Garber, Peter, 301, 320, 322n19 Garman, Marc, 9, 41, 59, 62, 65, 102,213 General Accounting Office (GAO), 210n2, 211n3 Gennotte, Gerard, 9, 43n2, 209n1
Ghysels, E., 39 Giovannini, Alberto, 11,276,303 Glass, G. R., 117 Glassman, Debra, 6, 24, 153 Glosten, L. R., 74, 79, 84, 86, 99, 187 Goldstein, Morris, 211n3, 212nn5,6, 217n14, 222n19,303,312 Goodhart, Charles A. E., 5, 13, 44n4, 45n7, 89,114,125, 129n19, 186, 190,286 Goodman, S., 13 Grabbe, 0., 213n8 Grammatikos, T., 22 Griffiths, M. D., 135 Gros, Daniel, 11, 304 Grossman, Sanford J., 9, 56n14, 73, 209n1 Group ofTen (G-10), 8, 222n19 Group of Thirty, 211 n3 Hakkio, Craig, 276 Halttunen, Hannu, 269 Hansch, 0., 62 Hansen, Lars P., 276 Hasbrouck, J., 112, 137, 138, 163-66, 185 Hausman, J., 185 Henderson, Dale W., 268n5 Heston, S., 31 Hillion, P., 73 Hirshleifer, David, 104 Ho, T. S. H., 62, 137, 138, 199 Hodrick, Robert J., 275n12, 276, 281 Hooper, Peter, 274n 11 Hsieh, D., 27nll, 102 Hull, J., 30 Hull, John C., 214n10 International Monetary Fund (IMF), 230, 246 Ito, Takatoshi, 5, 286n21 Jacklin, C., 43n2 Jasiak, J., 39 Jeanne, Olivier, 314n13 Jennergren, Peter, 213n8 Johnson, Harry G., 263 Jorion, Philippe, 25, 30n12, 276 Karpoff, J., 22n2 Kirman, A. P, 104 Kleidon, A. W., 43n2, 55n13, 56, 60nn15,16, 61n18, 63, 102 Kohlhagen, Steven, 9, 213 Kroner, Kenneth, 213n8 Krugman, Paul, 9, 13, 73,285n20
339
Author Index
Kurz, M., 104 Kyle, Albert, 55 Lai, Kon S., 264 Lamoureux, C., 32, 35 Lastrapes, W., 32, 35 Leach, C., 164 Lease, R., 138 Lee, C. M. C., 114 Lee, J. H., 40, 125 Lehmann, Bruce, 263 Leland, Hayne, 9, 25n6, 43n2, 209n 1 Levich, Richard, 39 Lewis, K. K., 60, 104,269,276 Lin, Wen-Ling, 286n21 Linde, Luis M., 322n21, 323f Litzenberger, R. H., 42n1 Lo, A., 185 Lothian, James R., 264 Lucas, Robert E., Jr., 269 Lyons, R., 5, 6, 24n4, 25n5, 44nn4,5,6, 47, 54n12, 62, 73,103,109,110,118, 155n28, 181, 184n1, 185, 186, 189, 195, 202,204,230,286 MacDonald, Ronald, 261, 262n 1, 268, 274, 277n14 Mclnish, T. H., 135, 140 MacKinlay, C., 185 McKinnon, Ronald, 5 McMahon, Patrick C., 264 McNown, Robert, 264 Madhavan, A., 155n28, 164, 185,203 Manaster, S., 185 Mann, S., 185 Mark, Nelson C., 264, 274 Marston, Richard C., 285 Masson, Paul R., 268 Masulis, R., 138 Meese, Richard A., 10,267,273,274, 305n3 Melino, A.; 30n14, 213n8 Melvin, Michael, 6, 24, 124, 129n19, 134, 155 Mendelson, H., 62 Milgrom, P. R., 74, 79,84,86,99,187 Miller, M. H., 56n14 Miller, Marcus, 13, 73 Morton, John, 274n11 Milller, U. A., 140 Mussa, Michael, 273, 281n18, 303 Naik, N., 62, 75, 213n8 Naslund, Bertil, 213n8 Neuberger, A., 75
Obstfeld, Maurice, 230, 281n18, 304n1, 320, 321n17 Officer, Lawrence H., 277n14 O'Hara, M., 25, 38n1, 62, 79,86, 183, 185, 198 Okunev, John, 213n8 Oldfield, G. S., 62 Ord, J. K., 135 Ozkan, F. GuIcin, 304n1 Page, J., 138 Perraudin, William R., 213n8 Petersen, M. A., 114 Pfleiderer, P., 42n1, 43, 47,54, 55-58, 65, 137, 139, 153, 184 Pictet, O. V., 115 Pitts, M., 6, 21, 22, 23, 27 Pomrenze, Jay L., 99 Ready, M. J., 114 Reuters, 191 Richardson, M., 23n3 Rogoff, Kenneth, 10, 11,267,268,273,274, 276,305n3,325 Roley, V. Vance, 5 Roll, R., 5, 137, 156 Romer, D., 43n2 Rose, Andrew, 11,59,60,231,246.261,285, 305,307,320n17,328 Saunders, A., 22 Schulmeister, Stephen, 13 Schultz, P. H., 62 Scott, E., 25, 26n8 Scott, L., 30 Securities and Exchange Commission (SEC), 209n1 Sharfstein, D. S., 105 Shastri, K., 30n12 Silber, William L., 99 Smidt, S., 155n28, 185,203 Smith, T., 23n3 Son, G., 62 Sorenson, Bent E., 213n8 Spencer, Michael, 301 Srinivas, P. S., 38 Stegun, I. A., 81n4 Stein,1. S., 104 Stigler, G. J., 203 Stock, J., 39, 163 Stockman, Alan C., 269, 270n8, 272, 273, 285
340
Author Index
Stoll, H., 62, 199 Stoughton, N., 30n12 Subrahmanyam, A., 42nl, 43, 47, 54, 55-57 Sultan, Jahangir, 213n8 Sutherland, Alan, 304n 1 Suvanto, A., 68, 70, 103 Svensson, Lars E. 0., 297, 299, 300f Tandon, K., 30n12 Tauchen, G., 6,21, 22, 23, 27 Taylor, Mark, 13,261, 262nl, 264,268,274, 281, 286, 322n19 Thaler, Richard, 275 Throop, Adrien W., 274 Thygesen, Niels, 11, 304 Tobin, James, 5, 262, 330 Truman, Edward, 304 Tucker, A., 25, 26n8 Turnbull, S., 30n14, 213n8
Viswanathan, S., 43n3, 56, 62, 75, 153 Vorst, T. C. E, 40 Wallace, Myles, 264 Warner, J. B., 42nl Wei, Shang-Jin, 6, 25 Welch, I., 104 Werner, I. W., 63 White, A., 30 White, H., 26 White, R. W., 135 Wiggins, J., 30 Wood, R. A., 135, 140 Wyatt, S., 30n12 Wyplosz, Charles, 11, 232n2, 246, 303, 305, 307,309nl0, 314n13, 320n17, 321, 328 Zhou, B., 124 Zhou, Z., 73
Subject Index
Admati-Pfleiderer asymmetric information model, 55-58 Arbitrage: avoidance, 203-5; conditions in foreign exchange trading for, 69-70; opportunities before ERM crisis, 230 Asset market models: characteristics of macro, 2; problems of macro, 2-3 Asset markets: approach to exchange rate research, 1-2; simplifying assumptions of, 2 Asymmetric information: costs of bid-ask spreads, 24; data for analysis of foreign exchange market, 45-47, 62; in foreign exchange market microstructure, 4-6, 41-42; in interdealer trading, 102-4 Asymmetric information models: analysis of volatility using, 42-43; differences of recent models from standard, 43; standard, 54-58,60-63 Balance-of-payments data: to analyze heterogeneous behavior, 254; shortcomings, 254-55 Black model of European options on futures, 29-30 Black-Scholes option pricing model, 30-31, 39-40 Bollerslev-Domowitz asymmetric information model, 56-57 Brokers: in currency market, 108; role in decentralized markets, 89; role in foreign exchange markets, 73-74
341
Capital controls: evidence during speculative attacks of, 307-11; expansionary macroeconomic policies with, 329-30; relation to balance-of-payments crises, 319 Central banks. See Intervention, central bank Currency markets: dynamic hedging in, 209; speculation, 311-19; swaps, 209 Currency risk: hedging, 209; with increased volatility, 21-22; management by fund managers, 216-17 D2000-2. See Reuters D2000-2 dealing system Data: characteristics of foreign exchange transactions data, 135-39; Reuters D2000-2 dealing system, 116, 124, 139-58; Reuters indicative quote system (FXFX), 116, 139-58, 181 Data sources: for analysis of asymmetric information, 44-47,59; analysis of foreign exchange market risk and turnover, 26; direct and brokered interdealer, and customer trading, 181; for index of speculative pressure, 305-7; for relation between exchange rate movements and macro fundamentals, 277, 280-81, 287; for Reuters D2000-2 analysis, 109-12; for transaction price model, 190-94 Dealers: in hot potato metaphor, 184; influence of information arrival on decisions, 68-70; in static and dynamic models of
342
Subject Index
Dealers (continued) interdealer trading, 75-88, 99-101. See also Marketmakers; Traders Dealing: automatic trading system, 113-14; electronic systems, 108, 181; in Reuters D2000-2, 114-24; study of relation to price revision, 163-65. See also Reuters D2000-2 dealing system; Reuters Dealing 2000-1 Delta, or hedge ratio: conditions for increased, 221; of a currency put option, 215, 221-22; role in dynamic hedging, 223 Derivations, transaction price model, 200-201 Derivatives: exchange-traded, 211-12; exchange-traded interest rate contracts, swaps, options, and forwards, 211; foreign exchange swaps, forwards, and options, 211; growth of market in, 210; option-pricing methods to construct, 209; stock index, 211
Efficient markets hypothesis: expectations component, 276; random walk transaction prices, 135; risk neutral, 275-76 Electronic Banking Service (EBS), 108 EMS. See European Monetary System (EMS) EMU. See European monetary union (EMU) Equilibrium model of exchange rates: allowances and intent, 262; empirical evidence, 272-73; real and nominal exchange rates in, 270-73; simple model, 270-72 ERM. See Exchange rate mechanism (ERM) European Monetary System (EMS): ERM crises (1992), 8, 232-33; exchange rate regime with narrow band (1992), 232. See also Exchange rate bands; Exchange rate mechanism (ERM) European monetary union (EMU), 303-4, 319-21,324-26 Event-uncertainty view: intertransaction time signals, 187; of trading intensity, 8, 18384,206 Exchange rate bands: with fixed rate regimes, 297; under Maastricht Treaty conditions, 320 Exchange rate mechanism (ERM): dynamic hedging sales in 1992 crisis, 223n21; exchange rate depreciation within bands, 297-98; interest rate differentials leading
up to crises, 230; Italy and Britain abandon (1992), 232-33; narrow band fixed exchange rate (1990), 232; proposal to preserve stability, 331; under terms of Maastricht Treaty, 320 Exchange rate models: asset market macro models, 2-3; based on macroeconomic theory fundamentals, 262-74; criticism of macroeconomic approach, 285-87; forecasting with macro-based models, 273-74; support for macroeconomic approach, 277-84 Exchange rate regimes: predictions with fixed rates, 297; predictions with floating rates, 296-97; related to equilibrium model of exchange rates, 273; speculative pressure against fixed rate, 219-25 Exchange rates: as flexible price in monetary model of exchange rates, 264-65; movements related to macroeconomic fundamentals, 277-84, 286-87; real and nominal rate in equilibrium model, 270-71; speculative attacks on, 305 Expectations: component in efficient market hypothesis, 276; in exchange rate market turnover, 21; formation in transaction price model with time, 187-88
Flexible-price monetary model. See Monetary model, flexible-price exchange rates Forecasts: bid-ask spreads as forecast of volatility, 33t, 34; GARCH model to forecast volatility, 3lt, 32-33, 39-40; information content to forecast volatility, 32-34; with monetary exchange rate models, 273-74 Foreign exchange: hedging exposure of, 212-13; in hot potato hypothesis of order-flow information, 187; optionpricing formula, 213-14. See also Prices, foreign exchange; Spreads, bid-ask Foreign exchange market: bid-ask spread in centralized and decentralized, 83; changing patterns of activity, 19-20; correlation of turnover and volatility, 19-21; crashes in decentralized, 84-86, 99; dealing banks in, 217-18; decentralized, 7, 74-75, 87-89, 185; distinction between brokered and direct trading, 180-81; efficiency in centralized and decentralized, 83-84; growth of, 19-20; intermarket connections with overlapping time zones,
343
Subject Index 70-71; marketmakers and retail market in, 44; microstructure analysis, 73; microstructure research, 286; noise trading and learning models of, 59-60, 66; price de-
termination in, 1; research in speculative efficiency, 286; speculative efficiency literature, 274; study of microstructure, 3-4. See also Currency markets; Currency risk; Interbank foreign exchange market; Prices, foreign exchange; Risk; Risk management; Risk premium; Spot foreign exchange market Futures market: contracts, 209; source of daily volume information, 19 GARCH model: to forecast market volatility, 31 t, 32-33, 39-40; to measure foreign exchange market volatility, 27-28 Garman/Kohlhagen option-pricing formula,
213-14,218 Globex trading system, 108 Glosten-Milgrom theory of microstructure, 7 Hedge ratio. See Delta, or hedge ratio Hedging: cross-hedges, 217; instruments, 209; money market hedges, 227; rolling hedges, 227; against short-term exchange rate movements, 216. See also Delta, or hedge ratio Hedging, dynamic: banks' use of techniques, 217-19; in crisis, 219-24; defined, 214; influence on trading volume and price movements, 219; during managed or fixed rate regime, 222-25; option pricing and delta hedging, 8-9, 215, 221-23; option-pricing theory in, 213 Heterogeneity: in foreign currency crises, 9-10; of information, 69-70; in market behavior leading to ERM crisis, 230; in market microstructure perspective, 4-5; model of various agents' behavior, 23435, 253-54; as research topic in microstructure of foreign exchange, 295 Hot potato view: clumped trading relation to, 197; intertransaction time signals, 187; of trading intensity, 8, 183-84, 206 Illiquidity, 219-24 Imperfect information aggregation models,
59-61 Implied standard deviation (ISD): as best estimate of future volatility, 21, 34-35;
Black-Scholes model as approximation to, 31, 39-40; to explain bid-ask spreads, 35; as predictor of future volatility, 25, 33 Information: in asset markets, 2-3; daily equilibria related to, 21; efficient processing of, 4-5; flow over time in transaction price model, 186-87; in forecasts of volatility, 32-34; heterogeneous, 69-70; in implied standard deviation of volatility, 25; of marketmakers in decentralized market, 74,87-89; marketmakers' sale of, 87-89; nonprice, 204-5; price in transaction price model with time, 186-87; in static and dynamic models of interdealer trading, 75-88, 99-101. See also Asymmetric information; Imperfect information aggregation models; Information flow model Information flow model: assumptions, 75-77; bid-ask spreads in centralized and decentralized markets, 83; dynamic model, 81; efficiency of centralized and decentralized markets, 83-84; filtering, 78-79; information rents, 79-80; martingale properties and volatility, 86-87; static form, 77-78,84; value of information propositions, 82-83 Instinet trading system, 108 Interbank foreign exchange market: characteristics of, 45-47; components of, 44-45; trading, 74 Interest rates: derivative instruments of, 211; differentials in pre-ERM crisis period, 230; effect of hedging strategies on defense of, 210, 219-20; in performance of uncovered interest rate parity, 281-83; raised in defense of fixed exchange rate, 220-25; smoothing within exchange rate bands, 297-300. See also Uncovered interest rate parity Intervention, central bank: effect of hedging strategies on, 210; during ERM crisis, 233; microstructure research of sterilized, 300-301; as research topic in microstructure of foreign exchange, 295; during speculative pressure, 306; with speculative pressure against currency, 219-25. See also Capital controls; Speculative attacks Inventory: inventory-based models, 61-62; inventory-carrying costs of bid-ask spreads, 24
344
Subject Index
Kolmogorov-Smirnov test, 307-11, 318 Kruskal-Wallis test, 307-11, 318 Lamoureux-Lastrapes model, 32, 33, 35 Learning models: asymmetric information in foreign exchange, 43-44, 59-60; to validate assumptions, 66 Liquidity: effect with central bank defense of interest rate, 219-24; in foreign exchange markets, 4-6; in standard asymmetric information models, 54-58; trading, 184, 186-87. See also Illiquidity Maastricht Treaty on European Union, 304; provision for Intergovernmental Conference (1996), 322; qualifications for participation, 320 Macroeconomic variables: behavior with capital controls, 305; tests for equality of distribution, 307-11, 318 Marketmakers: in decentralized foreign exchange market, 74; sale of information in interdealer market, 87-89; in transaction price model, 203-4. See also Brokers; Dealers; Traders Market microstructure, 41, 65; correlation of trading volume and volatility, 4-5; Glosten-Milgrom theory of, 7; proposed research topics in, 12, 286, 295-96; reasons to study, 2-4 Market microstructure models. See Information flow model; Trading model, interdealer Martingale properties of price: in decentralized market, 74; relation to volatility, 86-87 Microstructure, market. See Market microstructure Minex dealing system, 108 Mixing variable of information arrival: measurement of volatility with, 39; role in volume-volatility model, 38-39. See also Mixture of distribution hypothesis (MDH) Mixture of distribution hypothesis (MDH): assumptions, 22; correlation of unexpected risk and volume model, 21; volatilityvolume relation, 22-24, 31-34. See also Mixing variable of information arrival Monetary model, flexible-price exchange rates: assumptions, 262; empirical evidence, 267-68
Monetary model, sticky-price exchange rates: additions and allowances, 262; characteristics, 265; convergent saddle path, 266, 267f; empirical evidence, 267-68; monetary shocks, 266-67 Noise trading: models of, 59; as research topic in foreign exchange microstructure, 295; volatility and trading volume interaction, 5 Option pricing: methods in dynamic hedging strategies, 209; put pricing formula, 213-16 Option-pricing model, 29. See also Implied standard deviation (lSD) Options: exchange-traded, 209; over-thecounter (OTC), 209, 218; put delta value, 215; with stochastic volatility, 30-31 Order-processing costs, bid-ask spreads, 24 Over-the-counter (OTC) markets: exchangetraded derivatives, 211; options, 209, 212,218 Peso problem, 276, 299 Portfolio balance model of exchange rates: empirical evidence, 269; specifications, 262; substitutability of domestic and foreign assets in, 268-69 Predictions: of floating- and fixed-rate regimes, 296-97; of macro exchange rate models, 3 Prices, foreign exchange: D2000-2 and FXFX bid and ask transactions data, 135-58; determination, 1; formation in imperfect information aggregation models, 60; indicative quotes, 110, 113, 190; intertransaction time signals, 186-87; model of transaction prices with time, 185-90; of options with stochastic volatility, 30-31; in Reuters D2000-2 system, 114-24; volatility of bid-ask spreads, 24. See also Asymmetric information; Delta, or hedge ratio; Hedging; Hedging, dynamic; Information; Martingale properties of price; Spreads, bid-ask; Transaction price model Purchasing power parity (PPP): as exchange rate model, 263-64; regression of exchange rate movements and macro fundamentals, 288-90; relation between exchange rate movements and macro
345
Subject Index fundamentals, 277-80, 286-87, 294~95
Put pricing. See Option pricing Random walk: of asset prices, 137; efficient markets hypothesis, 135; of real exchange rates, 10-11 Rational bubbles, 276 Regime shifts, learning about, 276 Regularity of trading volume and volatility correlation, 4 Research: asset market approach for exchange rates, 1-2; in central bank intervention, 295; market microstructure, 12, 286, 295-96; noise trading suggestion for, 295; proposed future, 12-13; suggestion to study heterogeneity, 295 Reuters D2000-2 dealing system: brokered interdealer trading data, 180; characteristics of, 113-24; comparison with FXFX, 124-35; conditional heteroskedasticity in, 158-63; data from interdealer trading, 167-77f; development and competition of, 107-8 Reuters Dealing 2000-1: dealer direct quotes and trades, 191-92; for direct trading, 181 Reuters indicative quote system (FXFX): comparison with Reuters D2000-2 dealing system, 124-35; data, 44-45,66-67,71, 116,139-58,181;shortcomings,110,190 Risk: with high volatility, 21-22; model of relation to turnover, 22-24 Risk management: of bank foreign exchange books, 217-19; dynamic hedging techniques used by banks, 217-18; by fund managers, 216-17; GarmanlKohlhagen option pricing formula, 213-16; trading limits as, 194 Risk premium: assumption in portfolio balance model of exchange rates, 269; influence on uncovered interest rate parity, 275; risk aversion in models of, 276 Speculative attacks: behavior of macroeconomic variables with, 305; evidence on capital controls during, 307-8; mechanics of, 311-19, 321; proposal to deter, 32124,330 Speculative pressure: under fixed exchange rate regime, 219-25; during transition to EMU, 319-24
Speculative pressure index, 305-6 Spot foreign exchange market: ratio of dealer to customer trades, 180; trading in bid and ask dealing systems analysis, 180; trading volume in, 180, 194-95 Spreads, bid-ask: comparison of regressions on, 32t, 33-34; in decentralized market, 74-89; difference in cross-market midquote variance, New York-London, 53-54; difference in mid-quote variance, New York-London, 52-53; differences in New York-London (fifteen-minute intervals), 48, 52; distribution of New YorkLondon daily, 48, 51f; with expected and unexpected trading volume, 25; forecast of volatility, 33t, 34; with price volatility, 24; relation to volatility and volume, 24-25; standard deviations of New YorkLondon,48,51ttypesofco~~24
Sticky-price exchange rate model. See Monetary model, sticky-price exchange rates Stochastic volatility models, 30-31 Subrahmanyam asymmetric information model, 55-58 Tauchen-Pitts model of turnover and risk, 21, 33 Touch, the, 110 Traders: in models of imperfect information aggregation, 59~63; in standard asymmetric information models, 54-58; in static and dynamic models of interdealer trading, 75-81, 99-101. See also Dealers; Marketmakers Trading: brokered and direct, 180-81; day trading, 61, 68-69; event-uncertainty view of intensity, 183-84, 187; hot potato view of intensity, 183-84, 187; interdealer trade and information flow model, 75-81; on International Money Market futures market, 194-95; limits, 194; microstructure analysis, 41; relation of activity to bid-ask spread, 24-25; screen- and voice-based, 180-81. See also Eventuncertainty view; Hot potato view Trading, liquidity: effect of trader clumping of, 184; hot potato hypothesis, 186-87 Trading model, interdealer: assumptions, 75--77; bid-ask spreads in centralized and decentralized markets, 83; dynamic model, 81; efficiency of centralized and decentralized markets, 83-84; filtering,
346
Subject Index
Trading model, interdealer (continued) 78-79; information rents, 79-80; martingale properties and volatility, 86-87; static form, 77-78,84-85; value of information propositions, 82-83 Trading systems, electronic, 108-9 Trading volume: ARMA time-series model for, 28-29; correlation with volatility, 4-5; as measure of trading activity, 25; measurement of futures market, 38; regularity of correlation with volatility, 4-6 Transaction price model: estimates of order flow information content, 195-96; formation of expectations, 187-88; with role for intertransaction time, 185-86 Turnover: heterogeneous expectations in, 21; relation to risk, 22-24
Uncertainty: as determinant of bid-ask spread, 24; event-uncertainty view, 8, 183-84, 187 Uncovered interest rate parity: as condition for exchange rate market efficiency, 274-76; conditions for rejection of, 299; relation between exchange rate movements and macro fundamentals, 280-83, 286-87; in sticky-price model of exchange rates, 265
Volatility, 47-50; correlation with trading volume, 4-6; GARCH model of, 27-28; implied volatilities, 29-31; stochastic, 30-31; time-series model for, 27-28. See also Implied standard deviation (ISD); Noise trading
