Bootstrapping Stationary ARMA-GARCH Models

Kenichi Shimizu Bootstrapping Stationary ARMA-GARCH Models

VIEWEG+TEUBNER RESEARCH

Kenichi Shimizu

Bootstrapping Stationary ARMA-GARCH Models With a foreword by Prof. Dr. Jens-Peter Kreiß

VIEWEG+TEUBNER RESEARCH

Bibliographic information published by the Deutsche Nationalbibliothek The Deutsche Nationalbibliothek lists this publication in the Deutsche Nationalbibliografie; detailed bibliographic data are available in the Internet at http://dnb.d-nb.de.

Dissertation Technische Universität Braunschweig, 2009

1st Edition 2010 All rights reserved © Vieweg+Teubner | GWV Fachverlage GmbH, Wiesbaden 2010 Editorial Office: Dorothee Koch |Anita Wilke Vieweg+Teubner is part of the specialist publishing group Springer Science+Business Media. www.viewegteubner.de No part of this publication may be reproduced, stored in a retrieval system or transmitted, in any form or by any means, electronic, mechanical, photocopying, recording, or otherwise, without the prior written permission of the copyright holder. Registered and/or industrial names, trade names, trade descriptions etc. cited in this publication are part of the law for trade-mark protection and may not be used free in any form or by any means even if this is not specifically marked. Cover design: KünkelLopka Medienentwicklung, Heidelberg Printing company: STRAUSS GMBH, Mörlenbach Printed on acid-free paper Printed in Germany ISBN 978-3-8348-0992-6

This book is dedicated to my grandmother Sumi Yamazaki, who always teaches me effort, tolerance and optimism.

Geleitwort Im Jahre 1979 hat Bradley Efron mit seiner Arbeit Bootstrap Methods: Another Look at the Jackknife das Tor zu einem in den vergangenen 30 Jahren intensiv bearbeiteten Forschungsgebiet aufgestoßen. Die simulationsbasierte Methode des Bootstraps hat sich in den verschiedensten Bereichen als ein außerordentlich effizientes Werkzeug zur Approximation der stochastischen Fluktuation eines Schätzers um die zu schätzende Größe erwiesen. Präzise Kenntnis dieser stochastischen Fluktuation ist zum Beispiel notwendig, um Konfidenzbereiche für Schätzer anzugeben, die die unbekannte interessierende Größe mit einer vorgegebenen Wahrscheinlichkeit von, sagen wir, 95 oder 99% enthalten. In vielen Fällen und bei korrekter Anwendung ist das Bootstrapverfahren dabei der konkurrierenden und auf der Approximation durch eine Normalverteilung basierenden Methode überlegen. Die Anzahl der Publikationen im Bereich des Bootstraps ist seit 1979 in einem atemberaubenden Tempo angestiegen. Die wesentliche und im Grunde einfache Idee des Bootstraps ist die Erzeugung vieler (Pseudo-) Datensätze, die von ihrer wesentlichen stochastischen Struktur dem Ausgangsdatensatz möglichst ähnlich sind. Die aktuellen Forschungsinteressen im Umfeld des Bootstraps bewegen sich zu einem großen Teil im Bereich der stochastischen Prozesse. Hier stellt sich die zusätzliche Herausforderung, bei der Erzeugung die Abhängigkeitsstruktur der Ausgangsdaten adäquat zu imitieren. Dabei ist eine präzise Analyse der zugrunde liegenden Situation notwendig, um beurteilen zu können, welche Abhängigkeitsaspekte für das Verhalten der Schätzer wesentlich sind und welche nicht, um ausreichend komplexe, aber eben auch möglichst einfache Resamplingvorschläge für die Erzeugung der Bootstrapdaten entwickeln zu können. Herr Shimizu hat sich in seiner Dissertation die Aufgabe gestellt, für die Klasse der autoregressiven (AR) und autoregressiven moving average (ARMA) Zeitreihen mit bedingt heteroskedastischen Innovationen adäquate Bootstrapverfahren zu untersuchen. Bedingte Heteroskedastizität im Bereich von Zeitreihen ist ein überaus aktuelles Forschungsthema, da insbesondere Finanzdatensätze als ein wesentliches Merkmal diese Eigenschaft tragen. Zur Modellierung der bedingten Heteroskedastizität unterstellt der Autor die Modellklasse der ARCH- und GARCH-Modelle, die von Robert Engle initiiert und im Jahr 2003 mit dem renommierten Preis für Wirtschaftswissenschaften der schwedischen Reichsbank in Gedenken an Alfred Nobel ausgezeichnet wurde. Herr Shimizu stellt zu Beginn seiner Arbeit mit einem negativen Beispiel sehr schön dar, welchen Trugschlüssen man bei der Anwendung des Bootstraps unterliegen kann. Danach untersucht er konsequent zwei verschiedene Modellklassen

VIII

Geleitwort

(AR-Modelle mit ARCH-Innovationen und ARMA-Modelle mit GARCH-Innovationen) im Hinblick auf die Anwendbarkeit verschiedener Bootstrapansätze. Er beweist einerseits asymptotisch, d.h. für gegen unendlich wachsenden Stichprobenumfang, geltende Resultate. Andererseits dokumentiert er in kleinen Simulationsstudien die Brauchbarkeit der vorgeschlagenen Methoden auch für mittlere Stichprobenumfänge. Im letzten Kapitel seiner Arbeit geht der Autor dann noch einen deutlichen Schritt weiter. Hier wird die parametrische Welt verlassen und ein semiparametrisches Modell betrachtet. Es geht darum, einen unbekannten funktionalen und möglicherweise nichtlinearen Zusammenhang bezüglich der nichtbeobachtbaren Innovationen zu schätzen und dann in einem weiteren Schritt die Verteilung des Kurvenschätzers mit einem Bootstrap-Ansatz zu imitieren. Die vorliegende Arbeit, die am Institut für Mathematische Stochastik der Technischen Universität Braunschweig angefertigt wurde, erweitert die Möglichkeiten der Anwendung des Bootstrapverfahrens auf eine Klasse von autoregressiven Moving-Average-Zeitreihen mit heteroskedastischen Innovationen. Dabei werden die Verfahren nicht nur hinsichtlich ihrer asymptotischen Qualität, sondern auch in Simulationen für mittlere Stichprobenumfänge erprobt. Hierdurch werden nicht nur dem ambitionierten Anwender wertvolle Hinweise für einen adäquaten Ansatz des Bootstraps in sehr komplexen Situationen mit abhängigen Daten gegeben. Ich bin davon überzeugt, dass der Bereich des Bootstraps und allgemeiner des Resamplings für abhängige Daten noch für längere Zeit ein Feld intensiver akademischer Forschung und gleichzeitig fortgeschrittener praktischer Anwendung in zahlreichen Bereichen bleiben wird. Prof. Dr. Jens-Peter Kreiß Technische Universität Braunschweig

Acknowledgements I am indebted to my PhD supervisor Professor Dr. Jens-Peter Kreiß for proposing this interesting research topic and for many valuable comments. Without his constructive criticism this book would not exist. I also thank Dr. Frank Palkowski, Ingmar Windmüller and Galina Smorgounova for helpful discussions on the topic. I would like to acknowledge the generous support of the Ministry of Finance Japan, which enabled me to continue working in this research area. Last but not least, heartily thanks to Katrin Dohlus for sharing fun and pain in the last part of our dissertations and publications, and for her continuous enthusiastic encouragement. Very special thanks to my grandmother, parents and brother for their constant support and encouragement throughout this work. Kenichi Shimizu

Contents 1

. . . .

1 1 3 5 7

2

Bootstrap Does not Always Work 2.1 Estimation of Heteroscedasticity and Bootstrap . . . . . . . . . . 2.2 Result of a False Application: the VaR Model . . . . . . . . . . .

9 10 13

3

Parametric AR(p)-ARCH(q) Models 3.1 Estimation Theory . . . . . . . 3.2 Residual Bootstrap . . . . . . . 3.3 Wild Bootstrap . . . . . . . . . 3.4 Simulations . . . . . . . . . . .

4

5

Introduction 1.1 Financial Time Series and the GARCH Model . . . 1.2 The Limit of the Classical Statistical Analysis . . . 1.3 An Alternative Approach: the Bootstrap Techniques 1.4 Structure of the Book . . . . . . . . . . . . . . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

19 19 37 52 60

Parametric ARMA(p, q)- GARCH(r, s) Models 4.1 Estimation Theory . . . . . . . . . . . . . 4.2 Residual Bootstrap . . . . . . . . . . . . . 4.3 Wild Bootstrap . . . . . . . . . . . . . . . 4.4 Simulations . . . . . . . . . . . . . . . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

65 65 68 76 80

Semiparametric AR(p)-ARCH(1) Models 5.1 Estimation Theory . . . . . . . . . . 5.2 Residual Bootstrap . . . . . . . . . . 5.3 Wild Bootstrap . . . . . . . . . . . . 5.4 Simulations . . . . . . . . . . . . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

85 . 85 . 98 . 115 . 117

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

Appendix

121

A Central Limit Theorems

123

B Miscellanea

125

XII

Contents

Bibliography

127

Index

131

List o

igures

1.1 1.2 1.3 1.4

Euro to Yen exchangerate (1st January  2002 - 31st December 2007) . . √ Σ −1 . . . . . . . . . . . . . . . . . . −1 Ω T (b0 − b0 ) and N 0, Σ The basic idea of the √ bootstrap techniques . . . . . . . . . . . . . . . . √ T (b0 − b0 ) and T (b∗0 − b0 ) . . . . . . . . . . . . . . . . . . . . . .

. . . .

. . . .

1 4 5 6

2.1

 † ∗ ξ 0.01 (1, 10, 000) and ξ0.01 (1, 10, 000) . . . . . . . . . . . . . . . . . . . .

17

3.1 3.2 3.3

Residual Bootstrap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Wild Bootstrap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Two-step Wild Bootstrap . . . . . . . . . . . . . . . . . . . . . . . . . . .

62 63 64

4.1 4.2

Residual Bootstrap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Wild Bootstrap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

82 83

5.1 5.2

Residual Bootstrap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119 Wild Bootstrap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119

1 Introduction 1.1 Financial Time Series and the GARCH Model In financial time series analysis it is difficult to handle the observed discrete time asset price data Pt directly, because Pt are often nonstationary and highly correlated (Figure 1.1 (a)). Thus it is usual to analyse the rate of return on the financial instrument, which is generally stationary and uncorrelated.

110

130

150

170

(a) Original data

2002

2003

2004

2005

2006

2007

2008

2007

2008

−0.02

0.00

0.02

(b) Geometric returns

2002

2003

2004

2005

2006

Figure 1.1: Euro to Yen exchange rate (1st January 2002 - 31st December 2007)

K. Shimizu, Bootstrapping Stationary ARMA-GARCH Models, DOI 10.1007/978-3-8348-9778-7_1, © Vieweg+Teubner | GWV Fachverlage GmbH, Wiesbaden 2010

2

1 Introduction

Here it is common to use geometric returns 

Pt Xt = log Pt−1

 = log Pt − log Pt−1

(Figure 1.1 (b)) because of the property that compounded geometric returns are given as sums of geometric returns and that geometric returns are approximately equal to arithmetic returns rt =

Pt − Pt−1 Pt = − 1. Pt−1 Pt−1

Emprical studies on financial time series1 show that it is characterised by • serial dependence in the data, • conditional heteroscedasticity, i.e. the volatility changes over time, • heavy-tailed and asymmetric unconditional distribution. It was the autoregressive conditionally heteroscedastic (ARCH) model of Engle (1982) which first succeeded in capturing those “stylised facts of financial data” in a simple manner. The model has the advantage of easiness in obtaining the theoretical properties of interest, such as stationarity conditions and moments, and simplicity of numerical estimation. A few years later it was generalised by Bollerslev (1986) to the generalised ARCH (GARCH) model. A stochastic process {Xt } is called GARCH(p, q) if it satisfies the equation Xt = σt Zt , p

q

i=1

j=1

2 2 + ∑ β j σt− σt2 = b0 + ∑ bi Xt−i j,

where b0 > 0, bi ≥ 0, i = 1, ..., p, β j ≥ 0, j = 1, ..., q, and {Zt } is a sequence of independent and identically distributed (i.i.d.) random variables such that E (Zt ) = 0 and E (Zt2 ) = 1. The process is called ARCH(p) if q = 0. After this important discovery a variety of alternative conditionally heteroscedastic models have been proposed, however, the GARCH model is still the benchmark model of financial time series analysis because of its simplicity.2 1 2

See, e.g., Mandelbrot (1963), Fama (1965) and Straumann (2005). See Bollerslev, Chou and Kroner (1992) for the application of ARCH models to financial time series data.

1.2 The Limit of the Classical Statistical Analysis

3

1.2 The Limit of the Classical Statistical Analysis In the classical statistical analysis we obtain asymptotic properties of the GARCH estimators by the central limit theorem for martingale difference sequence.3 It is important to note that the theorem yields the asymptotic properties of the estimators. In practice, however, the number of observations is limited and thus the error in the normal approximation not negligible. In general, the smaller the number of data is, the larger is the approximation error. While in simpler models one obtains a good approximation with small number of observations, more complex models such as the ARMA-GARCH require large amounts of data for a satisfactory approximation. In short, by the classical statistical analysis it is possible to obtain an estimator for the GARCH model, but often difficult to analyse how accurate the estimator is. Example 1.1 (ARMA(1,1)-GARCH(1,1) Model) Let us consider the ARMA(1,1)-GARCH(1,1) model: Xt = a0 + a1 Xt−1 + α1 εt−1 + εt ,  εt = ht ηt , 2 ht = b0 + b1 εt−1 + β1 ht−1 ,

t = 1, ..., 2, 000.

We simulate now the error in the normal approximation. The simulation procedure is as follows. iid Simulate the bias ηt ∼ 35 t5 for t = 1, ..., 2, 000. a0 Let X0 = E (Xt ) = 1−a , h0 = ε02 = E (εt2 ) = 1−bb0−β , a0 = 0.141, 1 1 a1 = 0.433, α1 = −0.162, b0 = 0.007, b1 = 0.135 and β1 = 0.829 and calculate

1/2  2 Xt = a0 + a1 Xt−1 + α1 εt−1 + b0 + b1 εt−1 + β1 ht−1 ηt for t = 1, ..., 2, 000. From the data set {X1 , X2 , ..., X2,000 } obtain the quasi-maximum likelihood (QML) estimators

1 b0 b1 β1 ) . θ = (a0 a1 α Repeat (step 1)-(step 3) 1,000 times. 3

See chapter 4 for the further discussion.

4

1 Introduction Σ −1 Ω −1 ,4 where Compute the estimated variance Σ 

2 ∂ ht (θ ) and ∑ 2 ∂ b0 t=1 2ht (θ )   2 1 T 1 1 T εt4  1 ∂ ht (θ )  Ω= . ∑ h2 (θ ) − 1 T ∑ 2h2 (θ ) ∂ b0 2 T t=1 t t=1 t = 1 Σ T

T

1

√ Compare the simulated density of the QML estimator ( T (b0  −  −1 −1    b0 ), thin line) with its normal approximation (N 0, Σ Ω Σ , bold line).

0

1

2

3

4

As we see in Figure 1.2, the error is fairly large because the number of data is small.

−0.4

Figure 1.2:

4

−0.2

0.0

0.2

0.4

  √ Σ −1 −1 Ω T (b0 − b0 ) and N 0, Σ

See Theorem 4.1 and 4.2 for the asymptotic properties of the QML estimators for the ARMA(p, q)GARCH(r, s) model.

1.3 An Alternative Approach: the Bootstrap Techniques

5

1.3 An Alternative Approach: the Bootstrap Techniques '

$ Bootstrap World

Real World

- P(X  1 X2 ... Xn ) estimated probability model


P unknown probability model




random sampling

?

x = (X1 X2 ... Xn ) observed data







?

random sampling

?

x∗ = (X1∗ X2∗ ... Xn∗ ) bootstrap sample

?

θ = s(x) original estimator   √  d n(θ − θ ) −−−→ N μ (P), σ 2 (P)

θ ∗ = s(x∗ ) bootstrap estimator   √ ∗  d n(θ − θ ) −−−→ N μB (P), σB2 (P)

n→∞

n→∞

&

%

Figure 1.3: The basic idea of the bootstrap techniques

There is another approach to analyse the accuracy of the estimators. Efron (1979) proposed the bootstrap technique, in which one estimates the accuracy of an estimator by computer-based replications of the estimator. Figure 1.3 is a schematic diagram of the bootstrap technique.5 In the real world an unknown probability model P generates the observed data x by random sampling. We calculate the estimator θ of interest from x. The statistical properties of θ is obtained by the central limit theorem   √ d n(θ − θ ) −−−→ N μ (P), σ 2 (P) , n→∞

 (x) where we estimate unknown μ (P) and σ 2 (P) by their consistent estimators μ

2 and σ (x). On the contrary, in the bootstrap world we obtain the estimated probability model P from x. Then P generates the bootstrap samples x∗ by random sampling, from which we calculate the bootstrap estimator θ ∗ . Here the mapping from x∗ → θ ∗ , s(x∗ ), is the same as the mapping from x → θ, s(x). Note 5

The diagram is a modification of Figure 8.3 in Efron and Tibshirani (1993, p. 91).

6

1 Introduction

that in the real world we have only one data set x and thus can obtain one estimator θ, while in the bootstrap world it is possible to calculate x∗ and θ ∗ as many times as we can afford. That is, through a large number of bootstrap replica√ tions we can approximate the estimation error of the original estimator n(θ − θ ) √ by the bootstrap approximation n(θ ∗ − θ). A bootstrap technique is called to work consistent) if the distribution of the bootstrap approximation √(or to be weakly  L n(θ ∗ − θ) converges weakly in probability to the same distribution as the √  distribution of the estimation error of the original estimator L n(θ − θ ) .

0

1

2

3

4

Although it is difficult to prove it theoretically, empirical studies6 have shown √ that if the number of observations is small, the bootstrap approximation n(θ ∗ − √ θ) yields a better approximation to n(θ − θ ) than the normal approximation

2 (x)). Figure 1.4 also indicates that in Example 1.1 the residual boot (x), σ N (μ strap approximation (see section 4.2 and 4.4 for the simulation procedures) provides a slightly better approximation than the normal approximation (cf. Figure 1.2).

−0.4

−0.2

Figure 1.4: 6

0.0

0.2

0.4

√ √ T (b0 − b0 ) and T (b∗0 − b0 )

See, e.g., Horowitz (2003), Pascual, Romo and Ruiz (2005), Reeves (2005), Robio (1999), and Thombs and Schucany (1990).

1.4 Structure of the Book

7

Here it is important to note that a bootstrap technique does not always work. Whether a bootstrap works, depends on the type of bootstrap procedures, probability models and estimation methods. Especially in presence of conditional heteroscedasticity, which is one of the main characteristics of financial time series, naïve application of bootstrap could result in a false conclusion, as we will see in the following chapter. From this viewpoint this work aims to research the asymptotic properties of the two standard bootstrap techniques - the residual and the wild bootstrap - for the stationary ARMA-GARCH models.

1.4 Structure of the Book This work is motivated by the following three research results: First, Kreiß (1997) investigated the AR(p) model with heteroscedastic errors and proved weak consistency of the residual and the wild bootstrap techniques. Second, Maercker (1997) showed weak consistency of the wild bootstrap for the GARCH(1,1) model with the quasi-maximum likelihood estimation. Third, Franke, Kreiß and Mammen (2002) examined the NARCH(1) model and proved weak consistency of the residual and the wild bootstrap techniques. In this work we sketch the results and try to extend them to the general bootstrap theory for heteroscedastic models. This book is organised as follows. Following a brief introduction to the heteroscedastic time series models, chapter 2 explains potential problems of bootstrap methods for heteroscedastic models and points out the risk of a false application with the example of the VaR theory. The third chapter treats the parameter estimation for the parametric AR(p)-ARCH(q) models with the ordinary least squares estimation. For the AR(p)-ARCH(q) models the residual and the wild bootstrap techniques are adopted and the consistency of the estimators investigated. For both methods a suitable procedure and the consistency criterium are determined. The theoretical analysis is then confirmed by simulations. In the following two chapters the results of chapter 3 are extended to two further models, which are the parametric ARMA(p, q)-GARCH(r, s) models with the quasi-maximum likelihood estimation on the one hand and the semiparametric AR(p)-ARCH(1) models with the Nadaraya-Watson estimation, in which heteroscedasticity parameters are nonparametric, on the other. For both models a suitable procedure and the consistency criterium are determined. Then the analysis is confirmed by simulations.

2 Bootstrap Does not Always Work As we have seen in the previous chapter, bootstrap is without doubt a promising technique, which supplements some weak points of classical statistical methods. In empirical studies, however, the limit of bootstrap tends to be underestimated, and the technique is sometimes regarded as a utility tool applicable to all models. Let us see a typical misunderstanding of bootstrap in econometric literature. The bootstrap is a simple and useful method for assessing uncertainty in estimation procedures. Its distinctive feature is that it replaces mathematical or statistical analysis with simulation-based resampling from a given data set. It therefore provides a means of assessing the accuracy of parameter estimators without having to resort to strong parametric assumptions or closed-form confidence-interval formulas. (...) The bootstrap is also easy to use because it does not require the user to engage in any difficult mathematical or statistical analysis. In any case, such traditional methods only work in a limited number of cases, whereas the bootstrap can be applied more or less universally. So the bootstrap is easier to use, more powerful and (as a rule) more reliable than traditional means of estimating confidence intervals for parameters of interest.1 It is important to note, however, that bootstrap does not always work. As mentioned in the previous chapter, whether a bootstrap method works, depends not only on the type of bootstrap technique, but also on the chosen model and the selected estimation method. In fact, we always need to check whether the bootstrap technique combined with the statistical model and the estimation method works or not. From this viewpoint it is naïve to believe that bootstrap does not require mathematical or statistical analysis. When does bootstrap work? When does bootstrap fail to work? Classical examples are explained in Mammen (1992). In this paper we concentrate on the case where we estimate heteroscedasticity, which is indispensable for financial risk management. As we will see, when estimating heteroscedasticity, the third

1

Dowd (2005, p. 105).

K. Shimizu, Bootstrapping Stationary ARMA-GARCH Models, DOI 10.1007/978-3-8348-9778-7_2, © Vieweg+Teubner | GWV Fachverlage GmbH, Wiesbaden 2010

10

2 Bootstrap Does not Always Work

and fourth moment of the bootstrap residuals play a substantial role, which requires more careful attention when applying bootstrap. In the first section we see possible mistakes in the application of bootstrap when estimating heteroscedasticity. In the second section we observe how a false application of bootstrap could lead to a grave mistake in practice.

2.1 Estimation of Heteroscedasticity and Bootstrap Assume that {Xt , t = 2, ..., T } are generated by the following AR(1)-ARCH(1) model: Xt = aXt−1 + εt ,  εt = ht ηt ,

(2.1)

2 , ht = b0 + b1 εt−1

where b0 > 0, b1 ≥ 0 and {ηt } is a sequence of i.i.d. random variables such that E (ηt ) = 0, E (ηt2 ) = 1, following a symmetric distribution, i.e. E (ηt3 ) = 0, and E (ηt4 ) =: κ < ∞.2 For simplicity, we adopt the ordinary least squares (OLS) method for parameter estimation throughout this section. Note that the OLS estimation consists of two steps, in which the ARCH part is estimated based on the residuals εt = Xt − aXt−1 , t = 2, ..., T. Now we apply the residual and the wild bootstrap to the AR(1)-ARCH(1) model and see the cases where bootstrap fails to work.

2.1.1 Simple Residual Bootstrap Does not Work The simplest residual bootstrap method is carried out as follows:3 Obtain the OLS estimator a and calculate the residuals

εt = Xt − aXt−1 ,

t = 2, ..., T.

Compute the standardised residuals  ε t = εt − μ 2 3

For further assumptions see §3.1.1. See, e.g., Efron and Tibshirani (1993, p. 111), Shao and Tu (1996, p. 289) and Kreiß (1997).

2.1 Estimation of Heteroscedasticity and Bootstrap

11

for t = 2, ..., T , where = μ

1 T ∑ εt . T − 1 t=2

Obtain the empirical distribution function FT (x) based on ε t defined by FT (x) :=

1 T ∑ 1 (ε t ≤ x) . T − 1 t=2

Generate the bootstrap process Xt∗ by computing Xt∗ = aXt−1 + εt∗ ,

εt∗ ∼ FT (x), iid

t = 2, ..., T.

Calculate the bootstrap estimator a∗ = (x x)−1 x x∗ ,  −1 E   e∗ ,   E) b ∗ = (E where x = (X1 X2 ... XT −1 ) , x∗ = (X2∗ X3∗ ... XT∗ ) , b ∗ = (b ∗0 b ∗1 ) , e∗ = (ε2∗2 ε3∗2 ... εT∗2 ) and ⎛ ⎞ 2 1 ε1 ⎜ 2 ⎟ ε2 ⎟ ⎜1 =⎜. E .. ⎟ ⎜. ⎟. ⎝. . ⎠ 2 1 ε T −1 In this case we obtain T

) E∗ (εt∗ ) = ∑ (εt − μ t=2

=

1 T −1

1 T ∑ εt −μ T − 1 t=2    = μ

=0

12

2 Bootstrap Does not Always Work

and T

 )2 E∗ (εt∗2 ) = ∑ (εt − μ t=2

=

1 T −1

1 T 2 1 T  εt − 2μ ∑ ∑ εt +μ 2 T − 1 t=2 T − 1 t=2    = μ

1 T 2 ∑ εt − μ 2 T − 1 t=2 2  = E (εt2 ) − E (εt ) + o p (1),    =

=0

where E∗ denotes the conditional expectation given the observations X1 , ..., XT . These properties yield weak consistency of the bootstrap estimators for the AR part (cf. the proof of Theorem 3.3). For the ARCH part, however, we do not obtain the necessary condition for weak consistency (cf. the proof of Theorem 3.4), namely   2 ) = 0. E∗ εt∗2 − (b 0 + b 1 ε t−1 In short, this simplest technique works for the AR part of the model (2.1), but not for the ARCH part. The problem of this method is that the bootstrap residuals εt∗2 ignore the heteroscedasticity of εt2 . To overcome the problem it is necessary to take a more complicated approach so that εt∗2 immitate the heteroscedasticity of εt2 correctly, which we will propose in section 3.2.

2.1.2 Wild Bootstrap Usually Does not Work Analogously to the residual bootstrap, the simplest wild bootstrap method is carried out as follows:4 Obtain the OLS estimator a and calculate the residuals

εt = Xt − aXt−1 ,

t = 2, ..., T.

Generate the bootstrap process Xt† by computing Xt† = aXt−1 + εt† ,

εt† = εt wt† , 4

See, e.g., Liu (1988) and Kreiß (1997).

iid

wt† ∼ N (0, 1),

t = 2, ..., T.

2.2 Result of a False Application: the VaR Model

13

Calculate the bootstrap estimator a† = (x x)−1 x x† ,  −1 E   e† ,   E) b † = (E

where x† = (X2† X3† ... XT† ) , b † = (b†0 b†1 ) and e† = (ε2†2 ε3†2 ... εT†2 ) . In this case we obtain E† (εt† ) = εt E† (wt† ) = 0 and 2

E† (εt†2 ) = εt E† (wt†2 ) = E (εt2 ) + o p (1), where E† denotes the conditional expectation given the observations X1 , ..., XT . These properties yield weak consistency of the bootstrap estimators for the AR part (cf. the proof of Theorem 3.5). For the ARCH part, however, we do not obtain the necessary condition for weak consistency (cf. the proof of Theorem 3.6), namely   2 2

t−1 2 ) = 0 E† εt†2 − (b 0 + b 1 ε t−1 ) = εt − (b0 + b1 ε and E† (εt†4 ) = 3E (εt4 ) + o p (1). There are two problems with the simplest method. Firstly, the bootstrap residuals εt†2 do not immitate the heteroscedasticity of εt2 accurately, which is analogous to the simplest residual bootstrap. Secondly, εt†2 have larger variance than εt2 . As we will see in section 3.3, it is possible to overcome the first problem by taking a more complicated approach. It is important to note, however, that the second problem still remains even if we take the more complicated approach (cf. Theorem 3.6 and 4.4). In conclusion: Wild bootstrap usually does not work when estimating heteroscedasticity (cf. Remark 3.6-3.7).

2.2 Result of a False Application: the VaR Model In this section we demonstrate the danger of a false application of bootstrap. Firstly, based on Hartz, Mittnik and Paolella (2006) we sketch the theory of applying bootstrap to the Value-at-Risk (VaR) forecast model. Then we simulate the result of a false application of bootstrap to the VaR forecast model.

14

2 Bootstrap Does not Always Work

2.2.1 VaR Model The VaR is one of the most prominent measures of financial market risk, which “summarizes the worst loss over a target horizon that will not be exceeded with a given level of confidence”.5 As a measure of unexpected loss at some confidence level, the VaR has been used by financial institutions to compute the necessary buffer capital. Since the Basel Committee adopted the internal-models-approach (1996), which allows banks to use their own risk measurement models to determine their capital charge, a large number of new models have been established. Among them the (G)ARCH model of Engle (1982) and Bollerslev (1986) is one of the simplest but still successful methods to obtain the VaR. Assume that {Xt , t = 1, ..., T } are generated by the following ARMA(p, q)GARCH(r, s) model: p

q

i=1

j=1

s

r

j=1

i=1

Xt = a0 + ∑ ai Xt−i + ∑ α j εt− j + εt ,  εt = ht ηt ,

2 ht = b0 + ∑ b j εt− j + ∑ βi ht−i ,

where b0 > 0, b j ≥ 0, j = 1, ..., s, βi ≥ 0, i = 1, ..., r, and {ηt } is a sequence of i.i.d. random variables such that E (ηt ) = 0, E (ηt2 ) = 1, following a symmetric distribution, i.e. E (ηt3 ) = 0, and E (ηt4 ) =: κ < ∞.6 The k-step-ahead VaR forecast is −1   ξ

λ (k, T ) = Φ (λ , XT +k , hT +k ), where Φ−1 (λ , μ , σ 2 ) denotes the λ × 100% quantile of the the normal distribution with mean μ and variance σ 2 , p

q

i=1 s

j=1 r



j ε i XT +k−i + ∑ α X T +k = a0 + ∑ a T +k− j   T +k− j 2 + ∑ βi h h T +k = b0 + ∑ b j ε T +k−i j=1

with Xt = Xt , εt = 0 and εt = ht for t > T . 2

5 6

Jorion (2007, p. 17). For further assumptions see §4.1.1.

i=1

and

2.2 Result of a False Application: the VaR Model

15

The k-step-ahead VaR forecast ξ

λ (k, T ) is a point estimator. It is possible to

construct confidence intervals of ξλ (k, T ) through computing the bootstrapped kstep-ahead VaR forecast −1 ∗ ∗ ∗ ξ

λ (k, T ) = Φ (λ , XT +k , hT +k ),

where p

q

i=1 s

j=1 r

∗ ε XT∗+k = a∗0 + ∑ a∗i XT +k−i + ∑ α j T +k− j

and

2

∗  h∗T +k = b∗0 + ∑ b∗j ε T +k− j + ∑ βi hT +k−i , j=1

i=1

∗ , ..., α

q∗ , numerous times with different bootstrapped estimators a∗0 , a∗1 , ..., a∗p , α 1 ∗ ∗ ∗

∗ .

, ..., β b , b , ..., b∗ and β 0

1

s

1

r

2.2.2 Simulations Let us consider the following ARMA(1,1)-GARCH(1,1) model: Xt = a0 + a1 Xt−1 + α1 εt−1 + εt ,  εt = ht ηt , 2 + β1 ht−1 , ht = b0 + b1 εt−1

t = 1, ..., 10, 000.

The simulation procedure is as follows. Note that the model and (step 1)-(step 3) are identical with the simulations in section 4.4. iid Simulate the bias ηt ∼ 35 t5 for t = 1, ..., 10, 000. a0 Let X0 = E (Xt ) = 1−a , h0 = ε02 = E (εt2 ) = 1−bb0−β , a0 = 0.141, a1 = 1 1 0.433, α1 = −0.162, b0 = 0.007, b1 = 0.135 and β1 = 0.829 and calculate

1/2  2 + β1 ht−1 ηt Xt = a0 + a1 Xt−1 + α1 εt−1 + b0 + b1 εt−1 for t = 1, ..., 10, 000. From the data set {X1 , X2 , ..., X10,000 } obtain the QML estimators a0 , a1 ,

1 , b0 , b1 and β1 . α

16

2 Bootstrap Does not Always Work

Compute the one-step-ahead VaR forecast −1   ξ

λ (1, T ) = Φ (λ , XT +1 , hT +1 )

with λ = 0.01 and T = 10, 000. Adopt the residual bootstrap of section 4.2 and the wild bootstrap of section 4.3 based on the data set and the estimators of (step 3). Then compute the bootstrapped one-step-ahead VaR forecast −1 ∗ ∗ ∗ ξ

λ (1, T ) = Φ (λ , XT +1 , hT +1 )

and

ξλ† (1, T ) = Φ−1 (λ , XT†+1 , h†T +1 ), where

∗ εT , XT∗+1 = a∗0 + a∗1 XT + α 1

∗ h

h∗T +1 = b∗0 + b∗1 εT + β 1 T, 2

and  

XT†+1 = a†0 + a†1 XT + α1† εT ,   2

h†T +1 = b†0 + b†1 εT + β1† h

T, with λ = 0.01 and T = 10, 000. Repeat (step 5) 2,000 times and construct a 90% confidence interval for  † ∗  ξ0.01 (1, 10, 000) and ξ0.01 (1, 10, 000). Figure 2.1 displays the density functions of the bootstrapped one-step-ahead ∗ VaR forecast with the residual bootstrap (ξ 0.01 (1, 10, 000), thin line) and the wild  † ∗ bootstrap (ξ0.01 (1, 10, 000), bold line). The 90% confidence interval for ξ 0.01 (1,  † 10, 000) is [−0.5752, − 0.5600] while that for ξ0.01 (1, 10, 000) is [−0.5737, −0.5614]. As we will see in section 4.4, the wild bootstrap does not work for the ARMA(1, 1)-GARCH(1, 1) model, which results in the apparently narrower  ∗ (1, 10, 000) in this simconfidence interval for ξ † (1, 10, 000) than that for ξ 0.01

0.01

ulation. That is, the wild bootstrap technique underestimates the existing risk and could lead a financial institution to the false decision to take the too much risk that is not covered by her capital charge.

17

0

20

40

60

80

100

120

2.2 Result of a False Application: the VaR Model

−0.58

−0.57

−0.56

−0.55

 † ∗ Figure 2.1: ξ 0.01 (1, 10, 000) and ξ0.01 (1, 10, 000)

3 Parametric AR(p)-ARCH(q) Models

In this chapter we consider the parametric AR(p)-ARCH(q) model based on ARCH regression models of Engle (1982). For simplicity, we estimate the parameters by the ordinary least squares (OLS) method and adopt the two-step estimation for the ARCH part, in which the parameters of the ARCH part are estimated based on the residuals of the AR part. In the first section we sketch the estimation theory for the parametric AR(p)-ARCH(q) model with the OLS method and prove asymptotic normality of the OLS estimators. In the following two sections possible applications of the residual and the wild bootstrap are proposed and their weak consistency investigated. These theoretical results are confirmed by simulations in the last section.

3.1 Estimation Theory Since Engle (1982) introduced the ARCH model, asymptotic theories for ARCHtype models have been presented by many researchers. Although most of them adopted the maximum likelihood method for parameter estimation, there are some studies on the OLS method. Weiss (1984, 1986) investigated the asymptotic theory for ARMA(p, q)-ARCH(s) models with the least squares estimation. Pantula (1988) examined the asymptotic properties of the generalised least squares estimators for the AR(p)-ARCH(1) model. In this section we consider the AR(p)ARCH(q) model and sketch the asymptotic properties of the OLS estimators based on these past research results.

K. Shimizu, Bootstrapping Stationary ARMA-GARCH Models, DOI 10.1007/978-3-8348-9778-7_3, © Vieweg+Teubner | GWV Fachverlage GmbH, Wiesbaden 2010

20

3 Parametric AR(p)-ARCH(q) Models

3.1.1 Model and Assumptions Assume that {Xt , t = p + 1, ..., T } are generated by the p-th order autoregressive model with the ARCH(q) errors: p

Xt = a0 + ∑ ai Xt−i + εt , i=1

εt =

 ht ηt ,

(3.1)

q

2 ht = b0 + ∑ b j εt− j, j=1

where b0 > 0, b j ≥ 0, j = 1, ... q, and {ηt } is a sequence of i.i.d. random variables such that E (ηt ) = 0, E (ηt2 ) = 1, following a symmetric distribution, i.e. E (ηt3 ) = 0, and E (ηt4 ) =: κ < ∞. Let mt := ht (ηt2 − 1), then E (mt ) = 0, E (mt2 ) = (κ − 1)E (ht2 ) and the third equation of (3.1) can be rewritten as q

2 εt2 = ht + ht (ηt2 − 1) = b0 + ∑ b j εt− j + mt . j=1

For simplicity, we write the model in vector and matrix notation as follows: Xt

=

xt 

x

=

X

⎛

⎞ ⎛ X p+1 1 ⎜ X p+2 ⎟ ⎜ 1 ⎜ ⎟ ⎜ ⎜ .. ⎟ = ⎜ .. ⎝ . ⎠ ⎝. XT

Xp X p+1 .. .

X p−1 Xp .. .

... ... .. .

1 XT −1

XT −2

...

εt2

=

et 

e

=

E

⎛ 2 ⎞ 1 εq2 εq+1 2 ⎜ ε 2 ⎟ ⎜ 1 εq+1 ⎜ q+2 ⎟ ⎜ ⎜ . ⎟=⎜. .. ⎝ .. ⎠ ⎝ .. . 1 εT2 −1 εT2 ⎛

a

2 εq−1 εq2 .. . εT2 −2

... ... .. . ...

a

⎞

εt ,

+

u,

⎞ ⎛ a0 ε p+1 ⎜ a1 ⎟ ⎟ ⎜ ⎟ ⎜ ε p+2 ⎟ ⎟ ⎜ a2 ⎟ ⎜ ⎟ ⎟ ⎜ ⎟ + ⎜ .. ⎟ , ⎠ ⎜ .. ⎟ ⎝ . ⎠ ⎝ . ⎠ XT −p εT ap X1 X2 .. .

⎞

⎛

+

⎛

b

+

mt ,

b

+ ⎞

m,

⎛ ⎞ b0 mq+1 ⎜ b1 ⎟ ⎟ ⎜ ⎟ ⎜ mq+2 ⎟ ⎟ ⎜ b2 ⎟ ⎜ ⎟ ⎟ ⎜ ⎟ + ⎜ .. ⎟ , ⎠ ⎜ .. ⎟ ⎝ . ⎠ ⎝ . ⎠ mT εT2 −q bq

ε12 ε22 .. .

⎞

3.1 Estimation Theory

21

2 ε 2 ... ε 2 ) . where xt = (1 Xt−1 Xt−2 ... Xt−p ) , et = (1 εt−1 t−q t−2 To prove asymptotic normality of the estimators for the AR part, we make the following assumptions (cf. Nicholls and Pagan (1983), section 2):

{εt , t = 1, ..., T } are martingale differences with E (εs εt ) = 0,

s = t,

= E (ht ) < σ12

< ∞,

s = t.

The roots of the characteristic equation p

z p − ∑ ai z p−i = 0 i=1

are less than one in absolute value. For some positive δ1

 2+δ1   E c1  xt εt  < ∞,

where c1  ∈ R p+1 . The matrix A defined T   1 E xt xt  ∑ T →∞ T − p t=p+1

A := lim is invertible.

Analogously, we make the following assumptions for the ARCH part: {mt , t = 1, ..., T } are martingale differences with E (ms mt ) = 0, =

s = t,

(κ − 1)E (ht2 ) < σ22

< ∞,

The roots of the characteristic equation q

yq − ∑ b j yq− j = 0 j=1

are less than one in absolute value.

s = t.

22

3 Parametric AR(p)-ARCH(q) Models

For some positive δ2

 2+δ2   E c2  et mt  < ∞,

where c2  ∈ Rq+1 . The matrix B defined T 1 E (et et  ) ∑ T →∞ T − q t=q+1

B := lim is invertible.

Remark 3.1 (A2) is equivalent to Xt being causal1 , i.e. Xt can be written in the form ∞

Xt = μ + ∑ αi εt−i ,

∞

∑ |αi | < ∞,

i=0

i=0

μ = E (Xt ) =

a0 . p 1 − ∑i=1 ai

where

Remark 3.2 (A1) and (A3) together with Theorem B.1 yield T 1 a.s. εt → E (εt ) ∑ T − p t=p+1

and

T 1 a.s. εr εs → E (εr εs ). ∑ T − p r, s=p+1

3.1.2 OLS Estimation In the two-step estimation for the AR(p)-ARCH(q) model it is necessary to estimate the AR part firstly. The ARCH part is then estimated based on the residuals of the AR part. Nicholls and Pagan (1983) proved asymptotic normality of the OLS estimators for the AR(p) model with heteroscedastic errors. Based on their proof now we sketch the asymptotic theory for the AR(p)-ARCH(q) model with the two-step OLS estimation. 1 See,

e.g., Fuller (1996, p.108).

3.1 Estimation Theory

23

3.1.2.1 AR part The model to be estimated is p

Xt = a0 + ∑ ai Xt−i + εt ,

t = p + 1, ..., T.

(3.2)

i=1

The OLS estimator  a = (a0 a1 ... ap ) is obtained from the equation   a=



T

∑

xt xt

t=p+1

−1   X x. = X X



T

∑



xt Xt

t=p+1

Note that −1    X (Xa + u) a = X X   −1  Xu = a+ X X and thus 

 T − p( a − a) =

−1 

1 X X T−p

√

 1 X u . T−p

Theorem 3.1 Suppose that Xt is generated by the model (3.1) satisfying assumptions (A1)(A4). Then      d T−p  a−a − → N 0, A−1 VA−1 , where V = lim

T →∞

T   1 ∑ E ht xt xt  . T − p t=p+1

Proof. The proof is the mirror image of the Appendix in Nicholls and Pagan (1983). The difference between the model (3.1) and that of Nicholls and Pagan (1983) is that our model contains a constant term a0 and assumes that ht is not independent of xt . We observe  −1      1 1    √ XX Xu T − p a−a = T−p T−p

24

3 Parametric AR(p)-ARCH(q) Models

and show 1 p X X → A T−p

(3.3)

and √

  1 d X u → N 0, V T−p

(3.4)

separately. Firstly, observe ⎛

1

⎜X T ⎜ t−1 1 1 ⎜ Xt−2 X X = ∑ ⎜ ⎜ .. T−p T − p t=p+1 ⎝ . Xt−p

Xt−1 2 Xt−1 Xt−2 Xt−1 .. .

Xt−2 Xt−1 Xt−2 2 Xt−2 .. .

... ... ... .. .

Xt−p Xt−1

Xt−p Xt−2

...

From Remark 3.1 and 3.2 we obtain T 1 ∑ Xt−r T − p t=p+1

 T  ∞ 1 μ + ∑ αi εt−r−i ∑ T − p t=p+1 i=0  ∞ T 1 =μ + ∑ αi ∑ εt−r−i T − p t=p+1 i=0    a.s. −−→E (εt−r−i ) =

a.s.

−−→E (Xt−r )

⎞ Xt−p Xt−1 Xt−p ⎟ ⎟ Xt−2 Xt−p ⎟ ⎟. ⎟ .. ⎠ . 2 Xt−p

3.1 Estimation Theory

25

for r = 1, ..., p. Analogously, we obtain T 1 Xt−r Xt−s ∑ T − p t=p+1   ∞ ∞ T 1 = ∑ μ + ∑ αi εt−r−i μ + ∑ α j εt−s− j T − p t=p+1 i=0 j=0   T ∞ ∞ 1 2 = ∑ μ + μ ∑ αi εt−r−i + ∑ α j εt−s− j T − p t=p+1 i=0 j=0 ∞

∞

+ ∑ ∑ αi α j εt−r−i εt−s− j i=0 j=0



T ∞ T 1 1 =μ + μ ∑ αi ∑ εt−r−i + ∑ α j T − p ∑ εt−s− j T − p t=p+1 i=0 j=0 t=p+1       a.s. a.s. −−→E (εt−r−i ) −−→E (εt−s− j ) 2

∞

∞

∞

+ ∑ ∑ αi α j i=0 j=0

T 1 εt−r−i εt−s− j ∑ T − p t=p+1    a.s. −−→E (εt−r−i εt−s− j )

a.s.

−−→E (Xt−r Xt−s ) for r, s = 1, ..., p, therefore, 1 a.s. X X −−→ A. T−p Secondly, let c ∈ R p+1 ,

φt := √

1 c xt εt , T−p

t = p + 1, ..., T

and Ft−1 be the σ -field generated by φt−1 , ..., φ1 . Then           1    c E xt ht Ft−1 E ηt Ft−1 = 0 E φt Ft−1 = √ T−p

26

3 Parametric AR(p)-ARCH(q) Models

and T

∑

s2T :=

E (φt2 )

t=p+1

=

  T 1   c c E x x h t t t ∑ T − p t=p+1

→c Vc < ∞

as T → ∞,

that is, φt is a square integrable martingale difference sequence. Therefore, it suffices to verify the assumptions of a version of the central limit theorem for martingale difference sequences (Brown (1971), see Theorem A.1), namely    p  E φt2 Ft−1 → 1

(3.5)

 p   E φt2 1 |φt | ≥ δ sT → 0.

(3.6)

T

s−2 T

∑

t=p+1

and for every δ > 0 T

s−2 T

∑

t=p+1

On the one hand, (3.5) holds if  T

∑

s2T

E





 φt2 Ft−1



− s2T

p

→ 0.

(3.7)

t=p+1

Note that here we already know s2T → c Vc < ∞ as T → ∞. By direct computation we obtain T

∑

    E φt2 Ft−1 − s2T

t=p+1

   T 1    x c c x h − E x x h t t t t t t ∑ T − p t=p+1    T 1    =c ∑ xt xt ht − E xt xt ht c T − p t=p+1 =

a.s.

−−→0, which shows (3.7) and thus (3.5).

3.1 Estimation Theory

27

On the other hand, from (A3) we obtain the Lyapounov condition (see Corollary A.1)  2+δ   E φt  2+δ

T

lim

T →∞

1

∑

t=p+1 sT

1 = lim T →∞ (T − p)δ /2



2+δ  T 1    ε E x c t t ∑ T − p t=p+1



=0 for some positive δ , which shows (3.6). The Cramér-Wold theorem and the central limit theorem for martingale difference sequences applied to φ1 , ..., φT show √

  1 d X u → N 0, V T−p 

and the Slutsky theorem yields the desired result. 3.1.2.2 ARCH part 3.1.2.2.1 The imaginary case where εt are known

Firstly, we analyse the imaginary case where {εt , t = 1, ..., T } are known. The model to be estimated is then q

2 εt2 = b0 + ∑ b j εt− j + mt ,

t = q + 1, ..., T.

j=1

 = (b0 b1 ... bq ) is obtained from the equation The OLS estimator b   b=



T

∑

t=q+1

et et



−1   E e. = E E

T

∑

et εt2

t=q+1

Note that  −1   b = E E E (Eb + m)   −1  Em =b + E E

(3.8)

28

3 Parametric AR(p)-ARCH(q) Models

and thus 

 − b) = T − q(b



−1 

1 E E T −q

√

 1 E m . T −q

Using arguments analogous to Theorem 3.1, we obtain the following corollary. Corollary 3.1 Suppose that εt2 is generated by the model (3.8) satisfying assumptions (B1)(B4). Then      d T −q  b−b − → N 0, B−1 WB−1 , where

κ −1 T ∑ E (ht2 et et  ). T →∞ T − q t=q+1

W = lim

3.1.2.2.2 The standard case where εt are unknown Secondly, we analyse the standard case where {εt , t = 1, ..., T } are unknown. In this case we adopt the two-step estimation, where the residuals of the AR part p

εt = Xt − a0 − ∑ ai Xt−i ,

t = p + 1, ..., T

i=1

are used to estimate the ARCH part. The model to be estimated is now q

2 2 εt = b0 + ∑ b j ε t− j + nt ,

t = p + q + 1, ..., T,

(3.9)

j=1

or in vector and matrix notation 2 εt

⎛

 e

et   E

= ⎞ 2

=

⎛

2 1 ε ε p+q p+q+1 ⎜ ⎜ 2 2⎟ ⎜ ε ⎟ ⎜ 1 ε p+q+1 ⎜ p+q+2 ⎟ = ⎜ .. .. ⎜ ⎟ ⎜ .. ⎝ ⎠ ⎝. . . 2 2 εT 1 ε T −1

ε p+q−1 2 ε p+q .. . 2 ε T −2

2  2 2 where et = (1 ε t−q ) . t−1 ε t−2 ... ε

b

2

... ... .. . ...

+

nt ,

b + n, ⎛ ⎞ ⎞ ⎛ ⎞ b0 2 ε1 n p+q+1 ⎜ ⎟ 2 ⎟ b1 ⎟ ⎜ ⎟ ε2 ⎟ ⎜ ⎜ n p+q+2 ⎟ ⎟ ⎜ b2 ⎟ ⎟ + ⎜ .. ⎟ , .. ⎟ ⎜ . ⎟ ⎝ . ⎠ . ⎠⎜ ⎝ .. ⎠ 2 nT ε T −q bq

3.1 Estimation Theory

29

The OLS estimator b = (b 0 b 1 ...b q ) is obtained from the equation   T T 2 

b= ∑ et et ∑ et εt t=p+q+1

−1     E E e. = E Note that

−1       Eb  +n

E E b= E  −1 E  n E = b+ E

and thus 

t=p+q+1

− b) = T − p − q(b



−1 

1  E E T − p−q

√

 1   En . T − p−q

Remark 3.3 For simplicity, we denote

εt − εt = xt  ( a − a) =

p

∑ xt,k (ak − ak ),

k=0

where xt = (1 Xt−1 Xt−2 ... Xt−p ) = (xt,0 xt,1 xt,2 ... xt,p ) , a = (a0 a1 ... a p )

and

 a = (a0 a1 ... a p )

throughout this chapter without further notice.

is asymptotically as efficient as b  in §3.1.2.2.1, as we will In fact, the estimator b see in the following theorem. Theorem 3.2 Suppose that Xt is generated by the model (3.1) satisfying assumptions (A1)(A4) and (B1)-(B4). Then    d

− b) → T − p − q(b N 0, B−1 WB−1 .

Proof. From Corollary 3.1 we obtain T 1 p et et  → B, ∑ T − p − q t=p+q+1

  T 1 d √ et mt → N 0, W . ∑ T − p − q t=p+q+1

30

3 Parametric AR(p)-ARCH(q) Models

Together with the Slutsky theorem it suffices to show T T 1 1 et et  + o p (1), et et  = ∑ ∑ T − p − q t=p+q+1 T − p − q t=p+q+1

(3.10)

T T 1 1 et mt + o p (1). et nt = √ ∑ ∑ T − p − q t=p+q+1 T − p − q t=p+q+1

(3.11)

√

To prove (3.10), it is sufficient to show the following properties for all i, j ∈ Z: 1 √ T 1 √ T 1 √ T 1 √ T

T

∑



 2 2 ε − ε t−i t−i = o p (1),

  2 2 2  ε ε − ε ∑ t−i t− j t− j = o p (1),

(3.13)

  2 2 2   ε ε − ε t−i t− j ∑ t− j = o p (1),

(3.14)

t=1 T

t=1 T 

∑

 2 2 2 2 ε t−i ε t− j − εt−i εt− j = o p (1).

t=1

Firstly, the left hand side of (3.12) equals 1 √ T 1 =√ T 1 =√ T 1 ≤√ T

T

∑



2 2 ε t−i − εt−i

t=1 T 



  2 T + √ ∑ εt−i ε t−i − εt−i T t=1 t=1   T 2 2 T ∑ xt−i  (a − a) + √T ∑ εt−i xt−i  (a − a) t=1 t=1   p p T √ 2 T 2 2  − a) ∑ ∑ xt−i,k ∑ (ak − ak ) + T ∑ εt−i xt−i  T (a   t=1 k=0 t=1 k=0    =O p (1)

∑

ε t−i − εt−i

2

=o p (1)

√ = T 

p

1

T

p

2 +o p (1) ∑ (ak − ak )2 T ∑ ∑ xt−i,k

k=0

  =O p (T −1/2 )

=o p (1).

(3.12)

t=1 T

t=1 k=0

   =O p (1)

(3.15)

3.1 Estimation Theory

31

Here we obtain 2 T

zt−i  := 

T

∑ εt−i xt−i  = o p (1)

t=1



because E zt−i  = 0, 

E zt−i  zt−i



4 = T



1 T

T

∑E





2 εt−i xt−i  xt−i



= o(1),

t=1

and, therefore, from Chebyshev’s inequality we obtain for every δ > 0       1      P zt−i − E zt−i  > δ ≤ 2 E zt−i  zt−i = o(1), δ that is, according to the definition   p zt−i  → E zt−i  = 0. Secondly, the proof of (3.13) is the mirror image of (3.12). 1 √ T 1 =√ T 1 =√ T 1 ≤√ T

T

2 ∑ εt−i



2 2 ε t− j − εt− j



t=1 T

 2   2 T 2 ε t− j − εt− j + √ ∑ εt−i εt− j ε t− j − εt− j T t=1 t=1   T 2 2 T 2 2 xt−j  (a −  a) + √ ∑ εt−i εt− j xt−j  (a −  a) ∑ εt−i T t=1 t=1   p p T √ 2 T 2 2 2 2  − a) ∑ εt−i ∑ xt− j,k ∑ (ak − ak ) + T ∑ εt−i εt− j xt−j  T (a   t=1 t=1 k=0 k=0    =O p (1) 2 ∑ εt−i

=o p (1)

√ = T 

p

1

p

T

2 2 ∑ xt− ∑ (ak − ak )2 T ∑ εt−i j,k +o p (1)

k=0

  =O p (T −1/2 )

=o p (1).

t=1



k=0



=O p (1)



32

3 Parametric AR(p)-ARCH(q) Models

Thirdly, the left hand side of (3.14) equals 1 √ T 1 =√ T

T

t−i ∑ ε

t=1 T 

2

  2 2 ε t− j − εt− j  

∑

2 2 ε t− j − εt− j

1 =√ ∑ T t=1  

ε t− j − εt− j

1 =√ T 

t=1 T 

ε t−i − εt−i

T

∑

2

2

  2 2 2 ε t−i − εt−i + εt−i

  + 2εt− j ε t− j − εt− j

   2 + 2εt−i ε + εt−i t−i − εt−i

 2  2 ε ε t− j − εt− j t−i − εt−i

t=1

2 +√ T  1 +√ T 2 +√ T  4 +√ T  2 +√ T





=:s1 T

∑ εt−i



ε t− j − εt− j

t=1

2   ε t−i − εt−i





=:s2 T

2 ∑ εt−i

t=1 T

∑ εt− j



ε t− j − εt− j



2

ε t− j − εt− j

t=1



ε t−i − εt−i





=:s3 T

∑ εt− j εt−i



ε t− j − εt− j

t=1



ε t−i − εt−i



T



ε t− j − εt− j

 

=:s4 2 ∑ εt− j εt−i

2



t=1

(3.13)

= s1 + s2 + s3 + s4 + o p (1).

We prove now that s1 , s2 , s3 and s4 are o p (1) respectively.

3.1 Estimation Theory

33

If i =j 1 s1 = √ T 1 ≤√ T √ = T



T

∑

xt−j  (a −  a)

t=1 T



p

∑ ∑

t=1



2 

xt−i  (a −  a)

p

2 xt− j,k

k=0

p

∑ (ak − ak ) ∑

∑ (ak − ak )

k=0



p

2

k=0 2 2





=O p (T −3/2 )

1 T



T





p

2 xt−i,k

k=0

∑ (ak − ak )

2

k=0



p

p

∑ ∑

t=1

2

2 xt− j,k

k=0

∑

2 xt−i,k

k=0





=O p (1)

=o p (1).

T

2 s2 = √ T 2 ≤ T

∑ εt−i



xt−j  (a −  a)

2

xt−i  (a −  a)

t=1

T



∑

t=1



p

p

k=0

k=0

2  εt−i ∑ xt− j,k ∑ (ak − ak ) xt−i

p

1 =2 ∑ (ak − ak ) T k=0    2

T



∑

t=1

=O p (T −1 )

2

p

εt−i ∑

k=0



√

2  xt− j,k xt−i





=O p (1)

T (a −  a)

√ 

T (a −  a)   =O p (1)

=o p (1).

p

1 s3 = 2 ∑ (ak − ak ) T k=0    2

=O p (T −1 )

T



∑

t=1

εt− j 

p

∑

k=0

2  xt−i, k xt−j



=O p (1)



√ 

T (a −  a) = o p (1).   =O p (1)

34

3 Parametric AR(p)-ARCH(q) Models

4 s4 = √ T 4 ≤√ T 4 ≤√ T √

T

∑ εt− j εt−i xt−j  (a − a) xt−i  (a − a)

t=1 T t=1



T

∑ |εt− j εt−i |

t=1



p

∑

2 xt− j,k

1

T

k=0

p

=4 T 

 

 

 

∑ |εt− j εt−i |xt−j  (a − a) xt−i  (a − a) 



p

∑





p

∑

(ak − ak )2 k=0

k=0



p

k=0

p

2 2 ∑ xt−i,k ∑ (ak − ak )2 T ∑ |εt− j εt−i | ∑ xt− j,k

k=0



t=1



=O p (T −1/2 )

k=0



k=0



=O p (1)



=o p (1).

If i=j 1 s1 = √ T 1 ≤√ T √ = T 



T

∑

xt−i  (a −  a)

t=1 T



∑ ∑

t=1



2 xt−i,k

k=0

p

2

∑ (ak − ak )

2

k=0 2 2

∑ (ak − ak )

k=0

=o p (1).

4

p

p



=O p (T −3/2 )



1 T

T



∑ ∑

t=1



2

p

k=0

2 xt−i,k



=O p (1)

p

∑ (ak − ak )2

2 xt−i,k



3.1 Estimation Theory

35

2 s2 = s3 = √ T 2 ≤√ T 2 ≤√ T √

T

∑ εt−i



xt−i  (a −  a)

t=1 T 

3 3 



∑ εt−i xt−i  (a − a) 

t=1

  ∑ εt−i  T



t=1



∑

k=0

3/2

p

2

k=0

3/2

p

∑ (ak − ak )

2 xt−i,k

∑ (ak − ak )

=2 T 

3/2 

p





=O p (T −1 )

2

k=0

  ∑ εt−i  T

1 T

t=1





3/2

p

2 ∑ xt−i,k

k=0



=O p (1)



=o p (1). 4 s4 = √ T 4 ≤√ T

T

2 ∑ εt−i

t=1 T

∑

a) xt−i  (a − 



2 εt−i

t=1

√ =4 T 



p

∑



p

2 xt−i,k

k=0

∑ (ak − ak )

2

k=0

p

∑ (ak − ak )2

k=0

2

  =O p (T −1/2 )

1 T

p

T

2 2 ∑ εt−i ∑ xt−i,k

t=1



k=0



=O p (1)



=o p (1). Finally, from (3.13) and (3.14) we obtain (3.15). 1 √ T

T

2 2 (3.14) 1 = √ t−i ε t− j ∑ ε T t=1

T

(3.13) 1 2 2 t−i εt− j + o p (1) = √ ∑ ε T t=1

T

2 2 εt− ∑ εt−i j + o p (1).

t=1

To prove (3.11), note that q

2 mt = εt2 − b0 − ∑ b j εt− j, j=1 q

2 2 nt = εt − b0 − ∑ b j ε t− j j=1

q    2 2 2 = mt + εt − εt2 − ∑ b j ε t− j − εt− j ,



j=1

36

3 Parametric AR(p)-ARCH(q) Models

and thus we obtain for i = 1, ..., q √ =√

T 1 2 nt ε t−i ∑ T − p − q t=p+q+1

  T T 1 1 2 2 2 2   √ + − m ε ε ε t ∑ t t−i ∑ t εt−i T − p − q t=p+q+1 T − p − q t=p+q+1    (3.14)

= o p (1)

q



T



1 2 2 2  − ∑ bj √ t− j − εt− j ε t−i ∑ ε T − p − q j=1 t=p+q+1    (3.14)

= o p (1)

=√

T

1 2 t−i + o p (1). ∑ mt ε T − p − q t=p+q+1

Therefore, to prove (3.11), it suffices to show for i = 1, ..., q   T 1 2 2  √ m ε − ε t t−i ∑ t−i = o p (1). T − p − q t=p+q+1 Analogously to the proof of (3.13), we obtain 1 √ T 1 ≤√ T √ = T 

  2 2  m ε − ε t t−i ∑ t−i T

t=1 T



∑ mt ∑

t=1



p

p

2 xt−i,k

k=0

p

∑ (ak − ak )

2

k=0

1

 + 

2 T



T

∑ mt εt−i xt−i

t=1



=o p (1)





√ 

T (a −  a)   =O p (1)

p

T

2 +o p (1) ∑ (ak − ak )2 T ∑ mt ∑ xt−i,k

k=0

  =O p (T −1/2 )

t=1



k=0



=O p (1)



=o p (1). 

3.2 Residual Bootstrap

37

3.2 Residual Bootstrap Kreiß (1997) proved the asymptotic validity of the residual bootstrap technique applied to the Yule-Walker estimators for the AR(p) model with homoscedastic errors. It is also shown that the wild bootstrap technique applied to the OLS estimators for the AR(p) model with heteroscedastic errors is weakly consistent. In this and the following section we introduce possible ways to bootstrap the AR(p)ARCH(q) model based on Kreiß (1997). In this section a residual bootstrap technique is proposed and its consistency proved. A residual bootstrap method can be applied to the model (3.1) as follows: Obtain the OLS estimator  a and calculate the residuals p

εt = Xt − a0 − ∑ ai Xt−i ,

t = p + 1, ..., T.

i=1

and calculate the estimated heteroscedasCompute the OLS estimator b ticity q

2 h t = b 0 + ∑ b j ε t− j ,

t = p + q + 1, ..., T.

j=1

Compute the estimated bias

εt ηt = h t for t = p + q + 1, ..., T and the standardised estimated bias

η t =

 ηt − μ  σ

for t = p + q + 1, ..., T , where = μ

T 1 ηt ∑ T − p − q t=p+q+1

and

σ 2 =

T 1  )2 . (ηt − μ ∑ T − p − q t=p+q+1

Obtain the empirical distribution function FT (x) based on η t defined by FT (x) :=

T 1 1 (η t ≤ x) . ∑ T − p − q t=p+q+1

38

3 Parametric AR(p)-ARCH(q) Models

Generate the bootstrap process Xt∗ by computing p

Xt∗ = a0 + ∑ ai Xt−i + εt∗ i=1

∗ εt = h t ηt∗ ,

ηt∗ ∼ FT (x), iid

t = p + q + 1, ..., T,

or in vector and matrix notation x∗

=

X

⎞ ⎛ ∗ X p+q+1 1 X p+q ⎜X∗ ⎟ ⎜ 1 X p+q+1 p+q+2 ⎜ ⎟ ⎜ ⎜ . ⎟ = ⎜ .. .. ⎝ .. ⎠ ⎝ . . ∗ 1 X XT T −1

X p+q−1 X p+q .. .

... ... .. .

XT −2

...

⎛

⎛

a∗ = (X X)−1 X x∗ ,  −1 E   e∗ ,   E) b ∗ = (E   ∗2 ∗2 ε p+q+2 ... εT∗2 . where e∗ = ε p+q+1

E∗ (ηt∗ ) =

T

∑



t=1

=



 ηt − μ σ

1 T

 1 1 T  ηt −μ ∑  σ T t=1    = μ

= 0, E∗ (ηt∗2 ) =

T

∑

t=1

=



 ηt − μ σ

1 1 σ 2 T 

= 1,

2

1 T

T

∑ (ηt − μ )2

t=1



=σ 2

⎞

+

u∗ ,

⎛ ∗ ⎞ a ⎞ ε p+q+1 Xq+1 ⎜ 0 ⎟  a 1 ⎜ ⎟ ⎜ ∗ ⎟ Xq+2 ⎟ ⎟ ⎜ ⎟ ⎜ ε p+q+2 ⎟ .. ⎟ ⎜ a2 ⎟ + ⎜ .. ⎟ . . ⎟ ⎝ . ⎠ . ⎠⎜ ⎝ .. ⎠ XT −p εT∗ ap

Calculate the bootstrap estimator

Remark 3.4 ηt∗ has the following properties:

 a



3.2 Residual Bootstrap

39 T

∑

E∗ (ηt∗3 ) =



t=1

1 1 σ 3 T

=

 ηt − μ σ

3

1 T

T

∑ (ηt − μ )3

t=1

= o p (1), T

E∗ (ηt∗4 ) =

∑



t=1

=

1 1  σ4 T

 ηt − μ σ

4

1 T

T

∑ (ηt − μ )4

t=1

= κ + o p (1), where E∗ denotes the conditional expectation given the observations X1 , ..., XT .

Analogously to the previous section, we observe a + u∗ ) a∗ = (X X)−1 X (X = a + (X X)−1 X u∗ , that is,  T − p − q(a∗ −  a) =



−1 

1 X X T − p−q

 1 √ X  u∗ . T − p−q

Analogously, let mt∗ = εt∗2 − h t , then e∗

 E

=

⎞ ⎛ 2 ∗2 1 ε ε p+q+1 p+q ⎜ 2 ⎜ ε ∗2 ⎟ 1 ε p+q+1 ⎜ p+q+2 ⎟ ⎜ ⎜ ⎜ . ⎟=⎜. .. ⎝ .. ⎠ ⎝ .. . 2 εT∗2 1 ε T −1 ⎛

ε p+q−1 2 ε p+q .. . 2 ε T −2

+ b m∗ , ⎞ ⎞ b 0 ⎛ ∗ ⎞ 2 ε m p+q+1 p+1 ⎜ b ⎟ ⎟ ⎟ ⎜ 1 2 ⎜ ∗ ⎟ ε p+2 ⎟ ⎜ ⎟ ⎜ m p+q+2 ⎟ ⎟ ⎜ b2 ⎟ + ⎜ ⎟ . .. ⎟ ⎜ ⎟ ⎝ .. ⎠ . ⎠ ⎜ .. ⎟ ⎝ . ⎠ ∗ 2 m ε T T −q b q ⎛

2

... ... .. . ...

and  −1 E   e∗   E) b ∗ = (E  −1 E   (E   E) b

+ m∗ ) = (E  −1 E   m∗ ,   E) = b + (E

40

3 Parametric AR(p)-ARCH(q) Models

that is, 

= T − p − q(b ∗ − b)



−1 

1  E E T − p−q

 1   m∗ . √ E T − p−q

Remark 3.5 For simplicity, we denote

− et  b h t − ht =et  b    

=et  b − b + et  − et  b  

−b + =et  b =et 





b−b +

q

∑

 

e

t,k − et,k bk

k=0 q 

∑

 2 2

ε t−k − εt−k bk ,

k=1

where 2 2 2 et = (1 εt−1 εt−2 ... εt−q ) = (et,0 et,1 et,2 ... et,q ) , 2  2 2  et = (1 ε t−q ) = (e

t,q ) , t−1 ε t−2 ... ε t,0 e

t,1 e

t,2 ... e

b = (b0 b1 ... bq )

and

= (b 0 b 1 ... b q ) b

throughout this chapter without further notice. Lemma 3.1 Suppose that Xt is generated by the model (3.1) satisfying assumptions (A1)(A4) and (B1)-(B4). Then (i) (ii)

1 T 1 T

T

∑



 h t − ht xt xt  = o p (1),

t=1 T 

∑

 2 h t et et  − ht2 et et  = o p (1).

t=1

Proof. To prove (i), it suffices to show the following properties for every i, j ∈ Z: 1 T 1 T

T

∑



 h t − ht = o p (1),

(3.16)

t=1

T

∑ Xt−i

t=1



 h t − ht = o p (1),

(3.17)

3.2 Residual Bootstrap

41

1 T

T

∑ Xt−i Xt− j



 h t − ht = o p (1).

(3.18)

t=1

Firstly, the left hand side of (3.16) equals

1 T

T

∑



h t − ht



t=1 T

  1 T q   1 2  2  b − b + e ε − ε ∑ t ∑ ∑ t−k t−k b k T t=1 T t=1 k=1  q    1 T 1

b−b + ∑ bk + (b k − bk ) = et  ∑ T t=1     k=1     T =O p (1) =O(1) =O p (T −1/2 )  −1/2 =

=O p (T

T

∑



2 2 ε t−k − εt−k

t=1

)





(3.12)

= o p (T −1/2 )

=o p (1).

Secondly, the left hand side of (3.17) equals   1 T Xt−i h t − ht ∑ T t=1  q   1 1 T 

b − b = b k + X e t−i t ∑ ∑  T t=1       k=1 T =O p (1)  =O p (1) −1/2 =O p (T

   2 − ε2 X ε t−i ∑ t−k t−k T

t=1

)

=o p (1).

Here we obtain

1 T

T

∑ Xt−i

t=1

  2 2 ε t−k − εt−k = o p (1)



=o p (1)





42

3 Parametric AR(p)-ARCH(q) Models

because the left hand side equals  2 2 T   ε t−k − εt−k + t−k − εt−k ∑ Xt−i εt−k ε T t=1 t=1    T 2 1 2 T   (a −  a) = ∑ Xt−i xt−k (a −  a) + x ∑ X t−i εt−k  t−k    T t=1 T t=1 T

1 T

≤

∑ Xt−i



T

1 T

p

p

l=0

l=0

2 ∑ Xt−i ∑ xt−k,l ∑ (al − al )2

t=1



p

= ∑ (al − al ) l=0



2



=O p (T −1 )



1 T

T

p

t=1

l=0

∑ Xt−i ∑ 

=O p (1)



=O p (T −1/2 )

+ o p (1)

2 xt−k,l



=O p (1)

+o p (1)



=o p (1).

Thirdly, the left hand side of (3.18) equals   1 T Xt−i Xt− j h t − ht ∑ T t=1  q   1 1 T  = b − b + b k X X e t−i t− j t ∑ ∑  T T t=1       k=1 =O p (1)  =O p (1) −1/2 =O p (T

   2 − ε2 ε X X ∑ t−i t− j t−k t−k T

t=1

)

=o p (1).

Here we obtain 1 T

T

∑ Xt−i Xt− j

t=1



 2 2 ε t−k − εt−k = o p (1)



=o p (1)



3.2 Residual Bootstrap

43

because the left hand side equals T

1 T ≤

∑ Xt−i Xt− j



 +

t=1



T

1 T



T

2 T

 j εt−k xt−k ∑ Xt−i Xt−  

t=1 

t=1



= ∑ (al − al )

2

p

p

l=0

l=0

l=0



=O p (T −1 )



1 T

T

p

t=1

l=0

∑ Xt−i Xt− j ∑ 

=O p (1)



2 ∑ Xt−i Xt− j ∑ xt−k,l ∑ (al − al )2

p



xt−k  (a −  a)

2

(a −  a)   

=O p (T −1/2 )

+ o p (1)

2 xt−k,l



=O p (1)

+o p (1)



=o p (1). To prove (ii), it suffices to show the following properties for every i, j ∈ Z:  2 h t − ht2 = o p (1),

(3.19)

 2 2 2 2 h t ε t−i − ht εt−i = o p (1),

(3.20)

 2 2 2 2 2 2 h t ε t−i ε t− j − ht εt−i εt− j = o p (1).

(3.21)

1 T 1 T 1 T

T

∑

T

∑

T

∑



t=1



t=1



t=1

Firstly, the left hand side of (3.19) equals 1 T

T

∑

t=1



 1 2 h t − ht2 = T

T

∑



h t − ht

t=1

2

+

2 T

 

h h − h ∑ t t t = o p (1) T

t=1

because  2 T  ht ht − ht ∑ T t=1  q   2 T

− b + ∑ b k 1 b ht et  = ∑ T t=1     k=1  T =O p (1) =O p (1)  −1/2 =O p (T

=o p (1)

)

T

∑ ht

t=1



2 2 ε t−k − εt−k



=o p (1)

 

44

3 Parametric AR(p)-ARCH(q) Models

and 1 T 1 = T



T

h t − ht

∑

t=1 T

∑



 

q    2 2 2

− b ε

− b) + 2 ∑ et  b et  (b − ε t−k t−k bk

t=1

+

2

k=1



q

t−k ∑ (ε

k=1

1 ≤ T 



q

T

∑∑ 

=O p (1)

+ 

1 T

2



 1 2

( b − b ) +2 b − b ∑ k k T k=0     q

et,2 k

t=1 k=0



2 − εt−k )b k

2

T

=O p q

∑∑



(T −1 )

2

2 ε t−k − εt−k

t=1 k=1

2



q

∑ b k



−b b   

=O p (T −1/2 )

 2 2

 b ( ε − ε ) ∑ k t−k t−k q

k=1

2

=O p (1)

o p (1)



t=1



k=1

(3.13), (3.14)

=o p (1) + 2

∑ et 

   

 =

T

q

1 b k  T k=1 =O p (1) 

∑

T

∑ et 

  2 2 ε t−k − εt−k

t=1





(3.12), (3.13)

=

o p (T −1/2 )

=o p (1). Secondly, to prove (3.20) it suffices to show 1 T 1 T 1 T

T

2 ∑ εt−i



 2 h t − ht2 = o p (1),

t=1 T 

∑

2 h t − ht2

t=1 T

∑ ht2





 2 2 ε t−i − εt−i = o p (1),

 2 2 ε t−i − εt−i = o p (1).

(3.22) (3.23) (3.24)

t=1

The left hand side of (3.22) equals 1 T

 2  1 2 2

h − h ε t ∑ t−i t = T t=1 T

 2 2 T   2 2

h h = o p (1) − h + h − h ε ε t t t t t ∑ t−i ∑ t−i T t=1 t=1 T

3.2 Residual Bootstrap

45

because  2 T 2  εt−i ht ht − ht ∑ T t=1  q   1 2 T 2 

= b − b ε h e + b k t t ∑ ∑ t−i  T t=1       k=1 T =O p (1)  =O p (1) −1/2 =O p (T

  2 2 2  ε h ε − ε t t−k ∑ t−i t−k T

t=1



)



=o p (1)

=o p (1) and 1 T 1 = T

T

2 ∑ εt−i

t=1 T

∑

2 εt−i

h t − ht

2

 

q    2 2 2

− b) + 2 ∑ et  b − b ε et  (b t−k − εt−k bk

t=1

+ 



k=1



q

t−k ∑ (ε

2

2 − εt−k )b k

k=1

1 ≤ T 

∑



q

T

2 εt−i

t=1



∑

k=0

=O p (1)



q

∑

et,2 k

k=0

 

2



b k − bk

2





=O p (T −1 )

 q  1 T  2 2 2 +2 b−b et  ∑ b k (ε εt−i t−k − εt−k ) ∑ T t=1 k=1  q  2 q 2 T 1 2 2 2 + t−k − εt−k ∑ εt−i ∑ ε ∑ b k T t=1 k=1 k=1       =O p (1)

=o p (1)

 

−b =o p (1) + 2 b   

=O p (T −1/2 )

=o p (1).

q

1 b k  T k=1 =O p (1) 

∑

T

2 et  ∑ εt−i

t=1



2 2 ε t−k − εt−k



=o p (1)

 

46

3 Parametric AR(p)-ARCH(q) Models

The left hand side of (3.23) equals 1 T =

1 T

T

∑

  2  2 2 h t − ht2 ε t−i − εt−i

t=1 T 

∑

2 2 ε t−i − εt−i



2

h t − ht

+

t=1

2 T

T

∑

    2 2

ε t−i − εt−i ht ht − ht

t=1

=o p (1)

because 1 T 1 = T

  2 2 2

 h − ε ε − h t−i t t ∑ t−i T

t=1

 q       2 2 2   2 2

  e b − b ε − ε ( b − b) + 2 e ε − ε t t−k ∑ t−i t−i ∑ t t−k bk T

t=1

+ 



q

t−k ∑ (ε

2

2 − εt−k )b k

2

k=1

1 ≤ T 

T

∑



2

2 ε t−i − εt−i

t=1





q

∑

k=0



=O p (1)

k=1



et,2 k

q

∑ (b k − bk )2

k=0

 



=O p (T −1 )



  q   T   2 2  2 2

− b 1 ∑ ε b k (ε +2 b t−i − εt−i et t−k − εt−k ) ∑ T t=1 k=1  q   2 q 2 T  1 2 2 2 2 + t−i − εt−i ∑ ε t−k − εt−k ∑ ε ∑ b k T t=1 k=1 k=1       =O p (1)

=o p (1)

=o p (1) + 2



b−b    

=O p (T −1/2 )

=o p (1)

q

1 b k  T k=1 =O p (1) 

∑

T

∑

t=1



   2 2  2 2 ε t−i − εt−i et ε t−k − εt−k 

=o p (1)



3.2 Residual Bootstrap

47

and    2 T  2 2

ε t−i − εt−i ht ht − ht ∑ T t=1     2 T  2 2 

 h b − b ε − ε e = ∑ t−i t−i t t T t=1       −1/2 =O p (T

=O p (1)

q

T

1 b k  T k=1 =O p (1) 

+∑

∑



)

   2 2 2 2 ε t−i − εt−i ht ε t−k − εt−k

t=1





=o p (1)

=o p (1).

The left hand side of (3.24) equals

1 T 1 = T ≤

1 T

T

∑ ht2



ε t−i − εt−i

+

t=1 T

∑

ht2





xt−i (a −  a)

2

2 T

T

∑ ht2 εt−i



T

p

p

+

k=0

k=0

2 T

T

∑

= ∑ (ak − ak )



2

k=0



=O p (T −1 )



1 T

T

t=1 



ht2 



=O p (1)



(a −  a)   

=O p (T −1/2 )

+ o p (1)

p

∑ ∑

t=1



ht2 εt−i xt−i 

2 2 ∑ ht2 ∑ xt−i, k ∑ (ak − ak )

t=1

  ε t−i − εt−i

t=1



t=1

p



2

2 xt−i, k

k=0



=O p (1)

+o p (1)



=o p (1).

The equation (3.21) can be proved analogously.



48

3 Parametric AR(p)-ARCH(q) Models

Lemma 3.2 Suppose that Xt is generated by the model (3.1) satisfying assumptions (A1)(A4) and (B1)-(B4). Let c1  = (c1,0 , c1,1 , c1,2 , ..., c1,p ) ∈ R p+1 and c2  = (c2,0 , c2,1 , c2,2 , ..., c2,q ) ∈ Rq+1 , then

3  h t xt  = O p (1),

(i)

 T 1   c ∑ T − p − q t=p+q+1 1

(ii)

  T 1   3 c2 ht et  = O p (1). ∑ T − p − q t=p+q+1

Proof. Since |a + b| ≤ |a| + |b|, we have   |a + b|3 ≤ 8 max{|a|3 , |b|3 } ≤ 8 |a|3 + |b|3 , therefore,   T 1   3 h x c  t t 1 ∑ T − p − q t=p+q+1   3   3 T  8      ≤ ∑ c1 ( h t − ht )xt  + c1  ht xt  T − p − q t=p+q+1    (A3)

= O p (1)

=

T

8 ∑ T − p − q t=p+q+1

 p 3     ∑ c1,k ( h t − ht )xt,k  + O p (1) k=0

 3 p  8  p 3

h 8 |c | − h )x ( t t t,k  + O p (1) ∑ ∑ 1,k T − p − q t=p+q+1 k=0  3  p −3/2  T 1   p 3

8 ∑ |c1,k | ≤ min{b0 , b0 } ∑ (ht − ht )xt,k  +O p (1) T − p − q t=p+q+1 k=0    ≤

T

=O p (1)

=O p (1).

3.2 Residual Bootstrap

49

Analogously,   T 1   3 c2 ht et  ∑ T − p − q t=p+q+1  3   3 T 64       ≤ ( h − h )e + ( h − h )( e − e )  c c  t t t t t t t 2 2 ∑ T − p − q t=p+q+1 3   3       + c2 ht et  +c2 ht ( et − et )    (B3)

= O p (1)

=O p (1). Here we obtain the last equation because  3 q T 64  q 3 − h )e 8 |c | h (  t t 2,k t,k ∑ ∑ T − p − q t=p+q+1 k=0    =O p (1)

3  T 64  2 q−1 3 2  + 8 |c | ε − ε ) − h )( h ( ∑ ∑ 2,k t t t−k t−k  T − p − q t=p+q+1 k=1    q

=O p (1)

3  q T 64  2 q−1 3 2  + 8 |c | ( ε − ε ) h ∑ ∑ 2,k t t−k t−k  +O p (1) T − p − q t=p+q+1 k=1    =O p (1)

=O p (1).  Theorem 3.3 Suppose that Xt is generated by the model (3.1) satisfying assumptions (A1)(A4) and (B1)-(B4). Then the residual bootstrap of section 3.2 for the AR part is weakly consistent, i.e.    d T − p − q(a∗ −  a) − → N 0, A−1 VA−1 in probability.

Proof. From Theorem 3.1 we already have 1 p X X − → A. T − p−q

50

3 Parametric AR(p)-ARCH(q) Models

Together with the Slutsky theorem it suffices to show   T 1 1 d √ X  u∗ = √ εt∗ xt → N 0, V ∑ T − p−q T − p − q t=p+q+1 Now we observe 1 1 h t ηt∗ xt , εt∗ xt = √ y∗t := √ T − p−q T − p−q

in probability.

t = p + q + 1, ..., T.

For each T the sequence y∗1 , ..., y∗T is independent because η1∗ , ..., ηT∗ is independent. Here we obtain 1 h t xt E∗ (ηt∗ ) = 0 E∗ (y∗t ) = √ T − p−q and   E∗ y∗t y∗t 

T

∑

t=p+q+1

=

T 1 h t xt xt  E∗ (ηt∗2 ) ∑ T − p − q t=p+q+1

=

  T T 1 1 h t − ht xt xt  ht xt xt  + ∑ ∑ T − p − q t=p+q+1 T − p − q t=p+q+1    Lemma 3.1

=

o p (1)

p

− →V. Let c ∈ R p+q+1 and s2T :=

T

∑

 2 p  E∗ c y∗t − → c Vc,

t=p+q+1

then we have the Lyapounov condition for δ = 1 1   ∗ 3 E c yt 3 ∗ t=p+q+1 sT T

∑

=

 3 T    E∗ |ηt∗ |3 1

  h c x t t ∑  3 1/2 T − p − q t=p+q+1 sT (T − p − q)    Lemma 3.2

=

=o p (1).

O p (1)

3.2 Residual Bootstrap

51

The Cramér-Wold theorem and the central limit theorem for triangular arrays applied to the sequence c y∗1 , ..., c y∗T show T

∑

  d y∗t → − N 0, V

in probability.

t=p+q+1

 Theorem 3.4 Suppose that Xt is generated by the model (3.1) satisfying assumptions (A1)(A4) and (B1)-(B4). Then the residual bootstrap of section 3.2 for the ARCH part is weakly consistent, i.e.    d

− T − p − q(b∗ − b) → N 0, B−1 WB−1 in probability.

Proof. From Theorem 3.2 we already have 1 p − E E → B. T − p−q Together with the Slutsky theorem it suffices to show √

  T 1 d ∗   m∗ = √ 1  E m → N 0, W e t ∑ t T − p−q T − p − q t=p+q+1

Now we observe 1 1 h t (ηt∗2 − 1) et , z∗t := √ m∗ et = √ T − p−q t T − p−q

in probability.

t = p + q + 1, ..., T.

For each T the sequence z∗1 , ..., z∗T is independent because η1∗ , ..., ηT∗ is independent. Here we obtain   1 h t et E∗ (ηt∗2 − 1) = 0 E∗ z∗t = √ T − p−q and from Lemma 3.1 T

∑

  E∗ z∗t z∗t 

t=p+q+1

    2 T 1 E∗ ηt∗4 − 2E∗ (ηt∗2 ) + 1 h t et et  ∑ T − p − q t=p+q+1    T 1 2  = κ − 1 + oP (1) ∑ ht et et + o p (1) T − p − q t=p+q+1 =

p

− →W.

52

3 Parametric AR(p)-ARCH(q) Models

Let c ∈ R p+q+1 and s2T : =

 2 p  E∗ c z∗t − → c Wc,

T

∑

t=p+q+1

then we have the Lyapounov condition for δ = 1 1   ∗ 3 ∑ 3 E∗ c zt  t=p+q+1 sT  3 T  1   1 ∗2

√ ht (ηt − 1) = ∑ E∗ c et  3 T − p−q t=p+q+1 sT   T E∗ |ηt∗2 − 1|3 1   3  = 3 h e  c t t ∑ sT (T − p − q)1/2 T − p − q t=p+q+1    T

Lemma 3.2

=

O p (1)

=o p (1). The Cramér-Wold theorem and the central limit theorem for triangular arrays applied to the sequence c z∗1 , ..., c z∗T show T

∑

  d z∗t → − N 0, W

in probability.

t=p+q+1



3.3 Wild Bootstrap Analogously to the previous section, we propose a wild bootstrap technique and investigate its consistency. A wild bootstrap method can be applied to the model (3.1) as follows: Obtain the OLS estimator  a and calculate the residuals p

εt = Xt − a0 − ∑ ai Xt−i , i=1

t = p + 1, ..., T.

3.3 Wild Bootstrap

53

and calculate the estimated heteroscedasCompute the OLS estimator b ticity q

2 h t = b 0 + ∑ b j ε t− j ,

t = p + q + 1, ..., T.

j=1

Generate the bootstrap process Xt† by computing p

Xt† = a0 + ∑ ai Xt−i + εt†

εt†

=

i=1

h t wt† ,

iid

wt† ∼ N (0, 1),

t = p + q + 1, ..., T,

or in vector and matrix notation x†

=

X

⎞ ⎛ † X p+q+1 1 X p+q ⎟ ⎜ ⎜ † ⎜ X p+q+2 ⎟ ⎜ 1 X p+q+1 ⎟ ⎜ . .. ⎜ .. ⎟ = ⎜ . ⎝ . ⎠ ⎝ .. † 1 X T −1 XT ⎛

⎛

X p+q−1 X p+q .. .

... ... .. .

XT −2

...

 a

⎞

+

⎛ † ⎞ ⎞ a ε p+q+1 Xq+1 ⎜ 0 ⎟ ⎟ ⎜ a1 ⎟ ⎜ ε † Xq+2 ⎟ p+q+2 ⎟ ⎟ ⎜ a2 ⎟ ⎜ ⎜ . .. ⎟ ⎜ ⎟ + ⎜ . ⎟ . ⎟ . ⎟ . ⎠⎜ ⎝ .. ⎠ ⎝ . ⎠ XT −p εT† ap

Calculate the bootstrap estimator a† = (X X)−1 X x† ,  −1 E   e† ,   E) b † = (E   †2 †2 ε p+q+2 ... εT†2 . where e† = ε p+q+1 Analogously to the previous section, we observe a† = (X X)−1 X (X a + u† ) = a + (X X)−1 X u† , that is,  T − p − q(a† −  a) =



−1 

1 X X T − p−q

u† ,

 1  † √ Xu . T − p−q

54

3 Parametric AR(p)-ARCH(q) Models

Let mt† = εt†2 − h t , then e†

 E

=

⎞ ⎛ 2 †2 ε p+q+1 1 ε p+q ⎟ ⎜ ⎜ †2 2 ⎜ ε p+q+2 ⎟ ⎜ 1 ε p+q+1 ⎟=⎜ ⎜ .. ⎜ .. ⎟ ⎜ .. ⎝ . ⎠ ⎝. . 2 εT†2 1 ε T −1 ⎛

ε p+q−1 2 ε p+q .. . 2 ε T −2

2

... ... .. .

+ b m† , ⎛ ⎞ ⎞ ⎛ † ⎞ b 0 2 m p+q+1 ε p+1 ⎜ b ⎟ ⎟ 2 ⎟⎜ 1 ⎟ ⎜ † ε p+2 ⎟ ⎜ ⎟ ⎜ m p+q+2 ⎟ ⎟ ⎟ ⎜ b2 ⎟ + ⎜ .. ⎟ ⎜ ⎟ ⎜ .. ⎟ . ⎟ ⎜ ⎠ ⎝ ⎠ . . .. ⎠ ⎝ 2 † mT ε T −q b q

...

and  −1 E   e†   E) b † = (E  −1 E   (E   E) b

+ m† ) = (E  −1 E   m† ,   E) = b + (E that is, 

= T − p − q(b † − b)



−1 

1  E E T − p−q

 1   m† . √ E T − p−q

Theorem 3.5 Suppose that Xt is generated by the model (3.1) satisfying assumptions (A1)(A4) and (B1)-(B4). Then the wild bootstrap of section 3.3 for the AR part is weakly consistent, i.e.    d T − p − q(a† −  a) − → N 0, A−1 VA−1 in probability.

Proof. The proof is the mirror image of Theorem 3.3. From Theorem 3.1 we already have 1 p X X − → A. T − p−q Together with the Slutsky theorem it suffices to show   T 1 1 d † √ X  u† = √ ε x → N 0, V ∑ t t T − p−q T − p − q t=p+q+1 Now we observe y†t

1 1 := √ εt† xt = √ T − p−q T − p−q



h t wt† xt ,

in probability.

t = p + q + 1, ..., T.

3.3 Wild Bootstrap

55

For each T the sequence y†1 , ..., y†T is independent because w†1 , ..., w†T is independent. Here we obtain 1 † h t xt E† (wt† ) = 0 E† (yt ) = √ T − p−q and    E† y†t y†t

T

∑

t=p+q+1

=

T 1 h t xt xt  E† (wt†2 ) ∑ T − p − q t=p+q+1

=

  T T 1 1 h t − ht xt xt  ht xt xt  + ∑ ∑ T − p − q t=p+q+1 T − p − q t=p+q+1    Lemma 3.1

=

o p (1)

p

− →V. Let c ∈ R p+q+1 and s2T :=

T

∑

 2 p  E† c y†t − → c Vc,

t=p+q+1

then we have the Lyapounov condition for δ = 1 1   † 3 ∑ 3 E† c yt  t=p+q+1 sT T

=

 3 T    E† |wt† |3 1

  h c x t t ∑  3 1/2 T − p − q t=p+q+1 sT (T − p − q)    Lemma 3.2

=

O p (1)

=o p (1). The Cramér-Wold theorem and the central limit theorem for triangular arrays applied to the sequence c y†1 , ..., c y†T show T

∑

  d y†t → − N 0, V

in probability.

t=p+q+1



56

3 Parametric AR(p)-ARCH(q) Models

Theorem 3.6 Suppose that Xt is generated by the model (3.1) satisfying assumptions (A1)(A4) and (B1)-(B4). Then the wild bootstrap of section 3.3 for the ARCH part has the following asymptotic property:    d

− T − p − q(b† − b) → N 0, B−1 τ WB−1 where τ =

in probability,

2 κ −1 .

Proof. The proof is the mirror image of Theorem 3.4. From Theorem 3.2 we already have 1 p − E E → B. T − p−q Together with the Slutsky theorem it suffices to show √

  T 1 d   m† = √ 1 E mt† et → N 0, W ∑ T − p−q T − p − q t=p+q+1

in probability.

Now we observe z†t := √

1 1 mt† et = √ h t (wt†2 − 1) et , T − p−q T − p−q

t = p + q + 1, ..., T.

For each T the sequence z†1 , ..., z†T is independent because w†1 , ..., w†T is independent. Here we obtain   1 h t et E† (wt†2 − 1) = 0 E† z†t = √ T − p−q and from Lemma 3.1 T

∑

   E† z†t z†t

t=p+q+1

 2    T 1 E† wt†4 − 2E† (wt†2 ) + 1 h t et et  ∑ T − p − q t=p+q+1  T 1  2 =2 ∑ ht et et + o p (1) T − p − q t=p+q+1 =

p

− →τ W.

3.3 Wild Bootstrap

57

Let c ∈ R p+q+1 and  2 p  E† c z†t − → c τ Wc,

T

s2T : =

∑

t=p+q+1

then we have the Lyapounov condition for δ = 1 T 1   † 3 ∑ 3 E† c zt  t=p+q+1 sT  3 T  1   1 †2

 √ h E = ∑ (w − 1) e c t t † t 3 T − p−q t=p+q+1 sT   T E† |wt†2 − 1|3 1   3  h = 3 e c  t t ∑ sT (T − p − q)1/2 T − p − q t=p+q+1    Lemma 3.2

=

O p (1)

=o p (1). The Cramér-Wold theorem and the central limit theorem for triangular arrays applied to the sequence c z†1 , ..., c z†T show   d z†t → − N 0, τ W

T

∑

in probability.

t=p+q+1

 Remark 3.6 Theorem 3.6 implies that the wild bootstrap of section 3.3 is weakly consistent only if κ = 3. In other words, if κ = 3, it is necessary to construct a wild p bootstrap in a different way so that E† (wt† ) = 0, E† (wt†2 ) = 1, E† (wt†3 ) → 0 p and E† (wt†4 ) → κ ; however, it is not always easy to obtain such wt† . In the next section we will see the result of a false application, in which the wild bootstrap with standard normal distribution is applied though κ = 9. Remark 3.7 A possible way to avoid the problem of the ‘one-step’ wild bootstrap is that we apply another bootstrap procedure again to the ARCH part. We can continue the ‘two-step’ wild bootstrap method as follows: Generate the bootstrap process εt‡2 by computing q 2 √

‡ εt‡2 = b 0 + ∑ b j ε t− j + κ − 1ht wt , j=1

iid wt‡ ∼

N (0, 1),

t = p + q + 1, ..., T,

58

3 Parametric AR(p)-ARCH(q) Models

or in vector and matrix notation e‡

 E

=

⎞ ⎛ ‡2 ε p+q+1 1 ⎟ ⎜1 ⎜ ε ‡2 ⎜ p+q+2 ⎟ ⎜ ⎜ . ⎟ =⎜ . ⎜ . ⎟ ⎜ . ⎝ . ⎠ ⎜ ⎝. εT‡2 1 ⎛

2 ε p+q 2 ε p+q+1 . . . 2  εT −1

ε p+q−1 2 ε p+q . . . 2  εT −2

2

... ... .. . ...

+ b n‡ , ⎛ ⎞ ⎞ b ⎞ ⎛ 2 ‡ 0 h ε p+q+1 w p+q+1 p+1 ⎜ ⎟ ⎟ ⎜ ‡ 2 ⎟ ⎜ b1 ⎟ ⎜ h p+q+2 w p+q+2 ⎟ ε p+2 ⎟ ⎜ ⎟ √ ⎟ ⎜ b2 ⎟ ⎟ ⎜ + κ −1⎜ ⎟. . ⎟ ⎜ ⎟ . ⎟ ⎜ . ⎟⎜ . ⎟ . ⎠ ⎝ . ⎠⎝ . ⎠ . . 2 ‡  εT −q hT wT b q

Calculate the bootstrap estimator   E)  −1 E   e‡ . b‡ = (E

Here we observe  −1 E   (E   E) b

+ n‡ ) b ‡ = (E  −1 E   n‡ ,

+ (E   E) =b that is, 

= T − p − q(b ‡ − b)



−1 

1  E E T − p−q

 1   n‡ . √ E T − p−q

Theorem 3.7 Suppose that Xt is generated by the model (3.1) satisfying assumptions (A1)(A4) and (B1)-(B4). Then the wild bootstrap of Remark 3.7 is weakly consistent, i.e.    d

− T − p − q(b‡ − b) → N 0, B−1 WB−1 in probability.

Proof. The proof is the mirror image of Theorem 3.4. From Theorem 3.2 we already have 1 p − E E → B. T − p−q Together with the Slutsky theorem it suffices to show    T κ −1 1 d   n‡ = √ E h t wt‡ et → N 0, W ∑ T − p − q t=p+q+1 T − p−q Now we observe z‡t

 :=

κ −1 ‡ ht wt et , T − p−q

t = p + q + 1, ..., T.

in probability.

3.3 Wild Bootstrap

59

For each T the sequence z‡1 , ..., z‡T is independent because w‡1 , ..., w‡T is independent. Here we obtain E‡ (z‡t ) = 0 and from Lemma 3.1    E‡ z‡t z‡t

T

∑

t=p+q+1

=

  T 2 κ −1 h t E‡ wt‡2 et et  ∑ T − p − q t=p+q+1   

=

 κ −1  2 h e e + o (1) p ∑ t t t T − p − q t=p+q+1 T



=1

p

→W. Let c ∈ R p+q+1 and s2T : =

T

∑

 2 p  E‡ c z‡t − → c Wc,

t=p+q+1

then we have the Lyapounov condition for δ = 1    T 1   ‡ 3 ∑ 3 E‡ c zt  t=p+q+1 sT ⎛  3 ⎞ T 1 ⎝  κ − 1 ‡  ⎠ ht wt et  = ∑ c E 3 ‡   T − p−q t=p+q+1 sT =

  T (κ − 1)3/2 1   3 ‡ 3  h e E |w | c  t t ‡ ∑ t s3T (T − q)1/2    T − p − q t=p+q+1   =O p (1)  Lemma 3.2

=

O p (1)

=o p (1). The Cramér-Wold theorem and the central limit theorem for triangular arrays applied to the sequence c z‡1 , ..., c z‡T show T

∑

  d zt ‡ → N 0, W

in probability.

t=p+q+1



60

3 Parametric AR(p)-ARCH(q) Models

3.4 Simulations In this section we demonstrate the large sample properties of the residual bootstrap of section 3.2 and the wild bootstrap of section 3.3. Let us consider the following parametric AR(1)-ARCH(1) model: Xt = a0 + a1 Xt−1 + εt ,  εt = ht ηt , 2 , ht = b0 + b1 εt−1

t = 1, ..., 50, 000.

The simulation procedure is as follows. iid Simulate the bias ηt ∼ 35 t5 for t = 1, ..., 50, 000. Note that E (ηt ) = E (ηt3 ) = 0, E (ηt2 ) = 1 and E (ηt4 ) = 9. a0 b0 Let X0 = E (Xt ) = 1−a , h0 = ε02 = E (εt2 ) = 1−b , a0 = 0.141, a1 = 0.433, 1 1 b0 = 0.007, b1 = 0.135 and calculate

1/2  2 ηt Xt = a0 + a1 Xt−1 + b0 + b1 εt−1 for t = 1, ..., 50, 000. From the data set {X1 , X2 , ..., X50,000 } obtain the OLS estimators a0 , a1 ,

b0 and b 1 . Repeat (step 1)-(step 3) 2,000 times. Choose one data set {X1 , X2 , ..., X50,000 } and its estimators {a0 , a1 , b 0 , b 1 }. Adopt the residual bootstrap of section 3.2 and the wild bootstrap of section 3.3 based on the chosen data set and estimators. Both bootstrap methods are repeated 2,000 times. √ Compare the simulated density of the OLS estimators ( T (a0 − a0 ) etc., thin lines) with the residual and the wild bootstrap approximations √ √  ( T (a∗ − a0 ), T (a† − a0 ) etc., bold lines). 0

0

Figure 3.1 illustrates that the residual bootstrap is weakly consistent. Moreover, Figure 3.1 (c)-(d) indicates that in finite samples the residual bootstrap yields more accurate approximation than the conventional normal approximation. The theoretical explanation of the result is leaved for further research.

3.4 Simulations

61

On the other hand, Figure 3.2 (c)-(d) shows the problem of the wild bootstrap for the ARCH part, which corresponds to the theoretical result in Theorem 3.6 with τ = 41 . Figure 3.3 also shows the problem of the two-step wild bootstrap, which seems to contradict Theorem 3.7. In fact, with T = 50, 000 the approximation error 1 T

T

∑ ηt4 − E (ηt4 )

t=1

is not negligible, which results in the large error of the bootstrap approximation. That is, although the two-step wild bootstrap works theoretically, it is of limited use for the AR(1)-ARCH(1) model.

3 Parametric AR(p)-ARCH(q) Models

0.0

0.0

0.2

0.1

0.4

0.6

0.2

0.8

1.0

0.3

1.2

1.4

62

−1.0

0.0

0.5

1.0

−4

√ √ T (a0 − a0 ) and T (a∗0 − a0 )

(b)

−2

√

0

T (a1 − a1 ) and

2

√

4

T (a∗1 − a1 )

0

0.00

2

0.02

4

0.04

6

0.06

8

0.08

10

(a)

−0.5

−0.15

(c)

−0.10

−0.05

0.00

0.05

0.10

0.15

√ √ T (b 0 − b0 ) and T (b ∗0 − b 0 )

−20

(d)

−10

√

0

T (b 1 − b1 ) and

Figure 3.1: Residual Bootstrap

10

√

T (b ∗1 − b 1 )

20

63

0.0

0.0

0.2

0.1

0.4

0.6

0.2

0.8

1.0

0.3

1.2

1.4

3.4 Simulations

−1.0

0.0

0.5

1.0

−4

√ √  T (a0 − a0 ) and T (a†0 − a0 )

(b)

−2

√

0

T (a1 − a1 ) and

2

4

√ † T (a1 − a1 )

0

0.00

2

0.02

4

0.04

6

8

0.06

10

0.08

12

0.10

(a)

−0.5

−0.15

(c)

−0.10

−0.05

0.00

0.05

0.10

0.15

√ √ T (b 0 − b0 ) and T (b†0 − b 0 )

−20

(d)

−10

√

0

T (b 1 − b1 ) and

Figure 3.2: Wild Bootstrap

10

√ † T (b1 − b 1 )

20

3 Parametric AR(p)-ARCH(q) Models

0

0.00

2

0.02

4

0.04

6

0.06

8

0.08

10

64

−0.15

(a)

−0.10

−0.05

0.00

0.05

0.10

0.15

√ √ T (b 0 − b0 ) and T (b‡0 − b 0 )

−20

(b)

−10

√

0

T (b 1 − b1 ) and

Figure 3.3: Two-step Wild Bootstrap

10

√

T (b‡1 − b 1 )

20

4 Parametric ARMA(p, q)- GARCH(r, s) Models In this chapter we extend the results of the previous chapter to the parametric ARMA(p, q)-GARCH(r, s) model estimated by the QML method. In the first section we sketch the estimation theory based on Francq and Zakoïan (2004). Then, analogously to the previous chapter, possible applications of the residual and the wild bootstrap are proposed and their weak consistency investigated. These theoretical results are confirmed by simulations in the last section.

4.1 Estimation Theory Since Bollerslev (1986) introduced the GARCH model, asymptotic theories for GARCH-type models have been presented by many researchers. Lee and Hansen (1994) and Lumsdaine (1996) examined asymptotic properties of the QML estimators for the GARCH(1, 1) model. Properties of the GARCH(p, q) model have been investigated by Berkes, Horváth and Kokoszka (2003) and Comte and Lieberman (2003). For the more general ARMA(p, q)-GARCH(r, s) model Ling and Li (1997) proved consistency and asymptotic normality of the local QML estimators. Properties of the global QML were investigated by Francq and Zakoïan (2004). See also Ling and McAleer (2003a) for vector ARMA-GARCH models, Ling and McAleer (2003b) for nonstationary ARMA-GARCH models, and Ling (2007) for self-weighted QML estimators for ARMA-GARCH models. In this section we introduce the ARMA(p, q)-GARCH(r, s) model and observe the asymptotic properties of the QML estimators based on Francq and Zakoïan (2004).

K. Shimizu, Bootstrapping Stationary ARMA-GARCH Models, DOI 10.1007/978-3-8348-9778-7_4, © Vieweg+Teubner | GWV Fachverlage GmbH, Wiesbaden 2010

66

4 Parametric ARMA(p, q)- GARCH(r, s) Models

4.1.1 Model and Assumptions Assume that {Xt , t = 1, ..., T } are generated by the following ARMA(p, q)GARCH(r, s) model: p

q

i=1

j=1

s

r

j=1

i=1

Xt = a0 + ∑ ai Xt−i + ∑ α j εt− j + εt ,  εt = ht ηt ,

(4.1)

2 ht = b0 + ∑ b j εt− j + ∑ βi ht−i ,

where b0 > 0, b j ≥ 0, j = 1, ..., s, βi ≥ 0, i = 1, ..., r, and {ηt } is a sequence of i.i.d. random variables such that E (ηt ) = 0, E (ηt2 ) = 1, following a symmetric distribution, i.e. E (ηt3 ) = 0, and E (ηt4 ) =: κ < ∞. The parameter space Φ = Φa × Φb is compact, where Φa ⊂ R p+q+1 and Φb ⊂ Rs+r+1 with R = (−∞, ∞) and 0 R0 = [0, ∞). Theoretically, εt and ht depend on the infinite past history of {Xt } or {εt }. For simplicity, however, we assume here the presample values of Xt , εt and ht , which does not affect the asymptotic results (see Bollerslev (1986)). Let θ 0 = (a0  , b0  ) , a0 = (a00 a01 ... a0p α01 ... α0q ) and b0 = (b00 b01 ... b0s β01 ... β0r ) be the true parameter vector. Denote θ = (a , b ) , a = (a0 a1 ... a p α1 ... αq ) and b = (b0 b1 ... bs β1 ... βr ) , and p

Aa (z) = 1 − ∑ ai zi , i=1

Ab (z) =

s

∑ b jz j

q

Ba (z) = 1 + ∑ α j z j ,

and

j=1

j=1 r

Bb (z) = 1 − ∑ βi zi . i=1

We further assume that

ηt2 has a non-degenerate distribution. If r > 0, Ab0 (z) and Bb0 (z) have no common root, Ab0 (1) = 0, b0s + β0r = 0 and ∑ri=1 βi < 1 for all b ∈ Φb . For all θ ∈ Φ Aa (z)Ba (z) = 0 implies |z| > 1. Aa0 (z) and Ba0 (z) have no common root, a0p = 0 or α0q = 0. ˚ where Φ ˚ denotes the interior of Φ. θ 0 ∈ Φ,

4.1 Estimation Theory

67

4.1.2 QML Estimation The log likelihood function of the model (4.1) for a sample of T observations is LT (θ ) =

1 T

T

∑ lt (θ )

with

t=1

ε2 1 lt (θ ) = − log ht (θ ) − t (θ ). 2 2ht

A QML estimator θ is any measurable solution of

θ = arg max LT (θ ). θ ∈Φ

Theorem 4.1 Suppose that Xt is generated by the model (4.1) satisfying assumptions (A1)(A5). Then   √ d T (θ − θ 0 ) − → N 0, Σ −1 Ω Σ −1 , where

 Σ=E  Ω=E

∂ 2 lt (θ 0 ) ∂θ∂θ



 =

∂ lt (θ 0 ) ∂ lt (θ 0 ) ∂θ ∂θ

Σ1 0



 =

0 Σ2 Ω1 0

 , 0 Ω2



with    1 ∂ εt (θ 0 ) ∂ εt (θ 0 ) ∂ ht (θ 0 ) ∂ ht (θ 0 ) 1 + E , ∂ a ht (θ 0 ) ∂ a ∂ a 2ht2 (θ 0 ) ∂ a   ∂ ht (θ 0 ) ∂ ht (θ 0 ) 1 , Σ2 = E ∂ b 2ht2 (θ 0 ) ∂ b     1 ∂ εt (θ 0 ) ∂ εt (θ 0 ) 1 κ −1 ∂ ht (θ 0 ) ∂ ht (θ 0 ) + E , E Ω1 = 2 ∂ a ht (θ 0 ) ∂ a ∂ a 2ht2 (θ 0 ) ∂ a   1 κ −1 ∂ ht (θ 0 ) ∂ ht (θ 0 ) Ω2 = . E 2 ∂ b 2ht2 (θ 0 ) ∂ b 

Σ1 = E

Proof. Francq and Zakoïan (2004), Theorem 3.2. Theorem 4.2 Suppose that Xt is generated by the model (4.1) satisfying assumptions (A1)-



68

4 Parametric ARMA(p, q)- GARCH(r, s) Models (A5). Then the matrices Σ and Ω in Theorem 4.1 can be consistently estimated by   1 T 1 ∂ ht (θ ) ∂ ht (θ ) 1 T 1 ∂ εt (θ ) ∂ εt (θ )

Σ1 = ∑ + ∑ , T t=1 2ht2 (θ ) ∂ a ∂ a T t=1 ht (θ ) ∂ a ∂ a  T 1 ∂ ht (θ ) ∂ ht (θ )

2 = 1 ∑ Σ , T t=1 2ht2 (θ ) ∂ b ∂ b   T T 4  ) ∂ ht (θ ) 1 1 1 ε ∂ h ( θ 1 t t

1 = Ω ∑ h2 (θ ) − 1 T ∑ 2h2 (θ ) ∂ a ∂ a 2 T t=1 t t=1 t  1 T 1 ∂ εt (θ ) ∂ εt (θ ) + ∑ , T t=1 ht (θ ) ∂ a ∂ a   1 T 1 ∂ ht (θ ) ∂ ht (θ ) 1 1 T εt4 

Ω2 = ∑ h2 (θ ) − 1 T ∑ 2h2 (θ ) ∂ b ∂ b . 2 T t=1 t t=1 t

Proof. Ling (2007, p. 854). Although Ling (2007) discussed the self-weighted QML estimator for the ARMA(p, q)-GARCH(r, s) model, we can analogously prove Theorem 4.2 assuming that the weight wt = 1. See also Theorem 5.1 in Ling and McAleer (2003a). 

4.2 Residual Bootstrap There have been numbers of works introducing bootstrap methods for ARMA or GARCH models. Kreiß and Franke (1992) proved asymptotic validity of the autoregression bootstrap technique applied to M estimators for the ARMA (p, q) model. Maercker (1997) showed weak consistency of the wild bootstrap for the GARCH(1, 1) model based on the asymptotic properties of the QML estimators that Lee and Hansen (1994) examined. Moreover, there are numbers of empirical applications of bootstrap methods for GARCH models, such as Reeves (2005), Pascual, Romo and Ruiz (2005), and Robio (1999). Despite the large number of empirical studies, there is a rather sparse literature investigating the validity of bootstrap techniques for ARMA-GARCH models.1 In this and the following section we introduce possible ways to bootstrap the stationary ARMA(p, q)GARCH(r, s) model based on the results of Francq and Zakoïan (2004). 1

Reeves (2005) showed weak consistency of the bootstrap prediction for the AR(1)-GARCH(1, 1) model assuming that the autoregression bootstrap estimators for the model are weakly consistent. As far as we know, however, weak consistency of the autoregression bootstrap for the AR-GARCH model has not been proved yet.

4.2 Residual Bootstrap

69

In this section a residual bootstrap technique is proposed and its weak consistency proved. A residual bootstrap method can be applied to the model (4.1) as follows: Obtain the QML estimator θ and calculate the residuals p

q

i=1

j=1

j ε εt = Xt − a0 − ∑ ai Xt−i − ∑ α t− j ,

t = 1, ..., T.

Compute the estimated heteroscedasticity s

r

j=1

i=1

2  ht = b0 + ∑ bj ε t− j + ∑ βi ht−i ,

t = 1, ..., T.

Compute the estimated bias

εt ηt = ht for t = 1, ..., T and the standardised estimated bias

η t =

 ηt − μ σ

and

σ 2 =

for t = 1, ..., T , where = μ

1 T

T

∑ ηt

t=1

1 T

T

∑ (ηt − μ )2 .

t=1

Obtain the empirical distribution function FT (x) based on η t defined by FT (x) :=

1 T

T

∑ 1 (η t ≤ x) .

t=1

Generate the bootstrap process Xt∗ by computing p

q

i=1

j=1

∗

j ε Xt∗ = a0 + ∑ ai Xt−i + ∑ α t− j + εt ,

εt∗ = ht ηt∗ ,

ηt∗ ∼ FT (x), iid

t = 1, ..., T.

70

4 Parametric ARMA(p, q)- GARCH(r, s) Models

Obtain the bootstrap estimator a∗ by the Newton-Raphson method

∗ (θ )−1 g∗ (θ ), a−H a∗ =  1 1 where g∗1 (θ ) =

∂ lt∗ (θ ) t=1 ∂ a T

1 T

∑

with   ∗2 ∂ lt∗ (θ ) εt  ∂ ht (θ ) εt∗  ∂ εt∗ (θ ) 1 = − (θ ) (θ ) − 1 ∂a ∂a ht ∂a 2ht (θ ) ht η ∗2 − 1 ∂ ht (θ ) η ∗ ∂ εt∗ (θ ) −√ t = t 2ht (θ ) ∂ a ht (θ ) ∂ a and

∗ (θ ) = 1 H 1 T

T

∑

t=1



1 ∂ εt∗ (θ ) ∂ εt∗ (θ ) ∂ ht (θ ) ∂ ht (θ ) + ∂ a ∂ a 2ht2 (θ ) ∂ a ht (θ ) ∂ a 1

Analogously, calculate the bootstrap estimator b∗

∗ (θ )−1 g∗ (θ ), b∗ =  b−H 2 2 where

∂ lt∗ (θ ) t=1 ∂ b T

1 g∗2 (θ ) = T

∑

with 1 ∂ lt∗ (θ ) = ∂b 2ht (θ ) =

 εt∗2  ∂ ht (θ ) (θ ) − 1 ht ∂b

ηt∗2 − 1 ∂ ht (θ ) 2ht (θ ) ∂ b

and

∗ (θ ) = 1 H 2 T



T

∑

t=1



∂ ht (θ ) ∂ ht (θ ) ∂ b 2ht2 (θ ) ∂ b 1

.

.

4.2 Residual Bootstrap

71

Remark 4.1 Analogously to Remark 3.4, we obtain E∗ (ηt∗ ) = 0, E∗ (ηt∗2 ) = 1 and E∗ (ηt∗k ) = E (ηtk ) + o p (1) for k = 3, 4, .... Remark 4.2 Here it is possible to build the bootstrap model q

p

Xt∗ = a0 + ∑ ai Xt−i + ∑ α j εt− j + εt∗ ,

εt =



i=1

ht ηt ,

j=1

εt∗

=



ht ηt∗ ,

s

r

j=1

i=1

ηt∗ ∼ FT (x), iid

2 ht = b0 + ∑ b j εt− j + ∑ βi ht−i

and obtain the QML estimator θ ∗ by the Newton-Raphson method

θ ∗ = θ − H∗ (θ )−1 g∗ (θ ), where

∂ LT∗ (θ ) 1 g∗ (θ ) = = ∂θ T H∗ (θ ) =

∂ lt∗ (θ )  ∗  ∗    = g1 , g2 (θ ), t=1 ∂ θ T

∑

∂ 2 LT∗ (θ ) 1 = T ∂θ∂θ

∂ 2 lt∗ (θ )  . t=1 ∂ θ ∂ θ T

∑

∗ instead of H∗ (θ ) for simplicity of the proof.

∗ and H However, we use H 1 2 Maercker (1997) also made use of this technique to bootstrap the GARCH(1,1) model.

To prove weak consistency of the residual bootstrap, we make use of the following result of Francq and Zakoïan (2004). Lemma 4.1 Suppose that Xt is generated by the model (4.1) satisfying assumptions (A1)(A5). Let φi denote the i-th element of vector a and ψ j the j-th of b, then for all i = 1, ..., p + q + 1 and j = 1, ..., r + s + 1,  3 1 T  ∂ εt (θ )  (i) ∑  ∂ φi  = O p (1), T t=1  3 1 T  ∂ ht (θ )  (ii) ∑  ∂ φi  = O p (1), T t=1  3 1 T  ∂ ht (θ )  (iii) ∑  ∂ ψ j  = O p (1). T t=1

72

4 Parametric ARMA(p, q)- GARCH(r, s) Models

Proof. In Francq and Zakoïan’s (2004) proof of Theorem 3.2 we obtain ! !  !  ∂ εt (θ ) ! ! !  ! < ∞, ! sup  !θ ∈U (θ 0 ) ∂ a ! 4

! !  !  1 ∂ ht (θ ) ! ! !  ! sup  ! < ∞, !θ ∈U (θ 0 ) ht (θ ) ∂ a ! 4

! !  !  1 ∂ ht (θ ) ! ! ! ! <∞ ! sup  !θ ∈U (θ 0 ) ht (θ ) ∂ b ! 4

for any neighbourhood U (θ 0 ) ∈ Φ. Since

∂ εt ∂ ht ∂ φi , ∂ φi

and

∂ ht ∂ψj

are continuous at θ 0

for all i = 1, ..., p + q + 1, j = 1, ..., r + s + 1 and θ → θ 0 , we obtain the desired result.  p

For a proof of weak consistency of the residual bootstrap we need a stronger condition than just to prove asymptotic normality of the QML estimators. E |ηt |6 < ∞. Theorem 4.3 Suppose that Xt is generated by the model (4.1) satisfying assumptions (A1)(A5) and (B1). Then the residual bootstrap of section 4.2 is weakly consistent, i.e.   √ d T (θ ∗ − θ ) − → N 0, Σ −1 Ω Σ −1 in probability,     b∗ . where θ ∗ = a∗ ,

Proof. Firstly, we prove weak consistency of the GARCH part. Observe √

∗ (θ )−1 √1  = −H T (b∗ − b) 2 T

∂ lt∗ (θ ) . t=1 ∂ b T

∑

2 and from Theorem 4.2 we already have Σ

2 −

∗ (θ ) = Σ → Σ 2 . Together Note that H 2 with the Slutsky theorem it is sufficient to show p

1 √ T

  ∂ lt∗ (θ ) d → − N 0, Ω 2 t=1 ∂ b T

∑

in probability.

4.2 Residual Bootstrap

73

Now we observe 1 ∂ lt∗ (θ ) z∗t : = √ T ∂b 1 ηt∗2 − 1 ∂ ht (θ ) , =√ T 2ht (θ ) ∂ b

t = 1, ..., T.

For each T the sequence z∗1 , ..., z∗T is independent because η1∗ , ..., ηT∗ is independent. Here we obtain  1 ∂ ht (θ )  ∗2 1 E∗ ηt − 1 = 0 E∗ (z∗t ) = √ T 2ht (θ ) ∂ b and T

∑ E∗



z∗t z∗t 



t=1

1  = E∗ (ηt∗4 ) − 2E∗ (ηt∗2 ) + 1 T

T

∑

t=1



∂ ht (θ ) ∂ ht (θ ) ∂ b 4ht2 (θ ) ∂ b 1



p

− →Ω 2 . Let c ∈ Rs+r+1 ,

 2 p T s2T := ∑ E∗ c z∗t → c Ω 2 c t=1

and ck denote the k-th element of vector c and ψk the k-th of b, then we have the Lyapounov condition for δ = 1 T 1   ∗ 3 ∑ 3 E∗ c zt  t=1 sT  3 E∗ |ηt∗2 − 1|3 T  1  ∂ ht (θ )  c =  ∑ ∂b  8s3T T 3/2 t=1  ht (θ ) ⎛  3 ⎞ E∗ |ηt∗2 − 1|3 1 ⎝ 1 T s+r+1 ∂ ht (θ )  ⎠ √ ≤ ∑  ∑ ck ∂ ψk  3 T T t=1  k=1 8s3T b0 ⎛  3 ⎞  T   E∗ |ηt∗2 − 1|3 8s+r 1 s+r+1 1 ∂ h ( θ )  t  3⎝ √ ≤ |c |  k ∑ ∑  ∂ ψk  ⎠ 3 T 3 T  t=1 k=1 8s b T 0

=o p (1).

74

4 Parametric ARMA(p, q)- GARCH(r, s) Models

We obtain the last equation from (B1) and Lemma 4.1. The Cramér-Wold theorem and the central limit theorem for triangular arrays applied to the sequence c z∗1 , ..., c z∗T show T

− N (0, Ω 2 ) ∑ z∗t → d

in probability.

t=1

Secondly, we prove weak consistency of the ARMA part analogously to the proof of the GARCH part. Observe √

∗ (θ )−1 √1 T (a∗ −  a) = −H 1 T Note that

∂ lt∗ (θ ) . t=1 ∂ a T

∑

q p  ∂ε ∂ εt∗ ∂  ∗ t = Xt − a0 − ∑ ai Xt−i − ∑ α j εt− j = ∂a ∂a ∂ a i=1 j=1

p

1 −

∗ (θ ) = Σ and, therefore, H → Σ 1 . Now it suffices to show 1

1 √ T

  ∂ lt∗ (θ ) d → − N 0, Ω 1 t=1 ∂ a T

∑

in probability.

We observe 1 ∂ lt∗ (θ ) y∗t := √ T ∂a  1 ηt∗2 − 1 ∂ ht (θ ) ηt∗ ∂ εt∗ (θ ) −√ =√ T 2ht (θ ) ∂ a ht (θ ) ∂ a  1 ηt∗2 − 1 ∂ ht (θ ) ηt∗ ∂ εt (θ ) −√ , =√ T 2ht (θ ) ∂ a ht (θ ) ∂ a t=1, ..., T. For each T the sequence y∗1 , ..., y∗T is independent because η1∗ , ..., ηT∗ is independent. We obtain E (y∗t ) =

1 √ T



E (η ∗ ) ∂ εt (θ ) E (ηt∗2 ) − 1 ∂ ht (θ ) −√ t ∂a 2ht (θ ) ht (θ ) ∂ a

=0

4.2 Residual Bootstrap

75

and   ∗ ∗ y E y ∑ ∗ t t T

t=1

1  = E∗ (ηt∗4 ) − 2E∗ (ηt∗2 ) + 1 T

T

∑

t=1



∂ ht (θ ) ∂ ht (θ ) ∂ a 4ht2 (θ ) ∂ a 1



  1 T 1 ∗3 ∗ − E∗ (ηt ) − E∗ (ηt ) ∑ 3/2 T t=1 2(ht ) (θ )   ∂ ht (θ ) ∂ εt (θ ) ∂ εt (θ ) ∂ ht (θ ) + ∂a ∂ a ∂a ∂ a  T  ) ∂ εt (θ ) 1 ∂ ε ( θ 1 t + E∗ (ηt∗2 ) ∑ T t=1 ht (θ ) ∂ a ∂ a

p

− →Ω 1 . Let c ∈ R p+q+1 ,

 2 p T u2T := ∑ E∗ c y∗t − → c Ω 1 c t=1

and ck denote the k-th element of vector c and φk the k-th of a, then we have the Lyapounov condition for δ = 1 T

1

3 

 

∑ u3 E∗ c y∗t 

t=1 T

⎛

  3 3 ⎞    ∗2 ∗   8 η ∂ εt (θ )  ⎠  η − 1 ∂ ht (θ )   ≤ 3 3/2 ∑ ⎝E∗ c t  + E∗ −c √ t   2ht (θ ) ∂ a   uT T t=1 ht (θ ) ∂ a        1  1 T  3  3 8 ∗2 3  ∂ ht (θ )  ∗ 3  ∂ εt (θ )  ≤ 3 3/2 ∑ E∗ |ηt − 1|  c c  + E∗ |ηt |  √   2ht (θ )  ht (θ ) ∂a  ∂a  uT T t=1 ⎛  3 ⎞  p+q+1 T  ∗2 3 p+q  1 1 E∗ |ηt − 1| 8  ∂ ht (θ )  ⎠ √ |ck |3 ⎝ ∑  ≤  ∑ 3 T t=1  ∂ φk  T k=1 u3T b0 ⎛  3 ⎞  p+q+1 T  ∗ 3 p+q+1  1 E∗ |ηt | 8 1  ∂ εt (θ )  ⎠ √ + lim = o p (1). |ck |3 ⎝ ∑   ∑ 3/2 T →∞ T t=1  ∂ φk  T k=1 u3 b0 T

T

We obtain the last equation from (B1) and Lemma 4.1.



76

4 Parametric ARMA(p, q)- GARCH(r, s) Models

4.3 Wild Bootstrap Analogously to the previous section, we propose a wild bootstrap technique and investigate its weak consistency. A wild bootstrap method can be applied to the model (4.1) as follows: Obtain the QML estimator θ and calculate the residuals p

q

i=1

j=1

j ε εt = Xt − a0 − ∑ ai Xt−i − ∑ α t− j ,

t = 1, ..., T.

Compute the estimated heteroscedasticity s

r

j=1

i=1

2  ht = b0 + ∑ bj ε t− j + ∑ βi ht−i ,

t = 1, ..., T.

Generate the bootstrap process Xt† by computing p

q

i=1

j=1

†

j ε Xt† = a0 + ∑ ai Xt−i + ∑ α t− j + εt ,

εt† = ht wt† ,

iid

wt† ∼ N (0, 1),

t = 1, ..., T.

Obtain the bootstrap estimator a† by the Newton-Raphson method

a† =  a − H†1 (θ )−1 g†1 (θ ), where 1 g†1 (θ ) = T with 1 ∂ lt† (θ ) = ∂a 2ht (θ ) = and 1

H†1 (θ ) = T

T

∑

t=1



∂ lt† (θ ) t=1 ∂ a T

∑



εt†2  (θ ) − 1 ht

∂ ht (θ ) εt†  ∂ εt† (θ ) − (θ ) ∂a ht ∂a

ηt†2 − 1 ∂ ht (θ ) η † ∂ εt† (θ ) −√ t 2ht (θ ) ∂ a ht (θ ) ∂ a 

1 ∂ εt† (θ ) ∂ εt† (θ ) ∂ ht (θ ) ∂ ht (θ ) + ∂ a ∂ a 2ht2 (θ ) ∂ a ht (θ ) ∂ a 1

.

4.3 Wild Bootstrap

77

Analogously, calculate the bootstrap estimator b†

†  −1 †  −H b† = b 2 (θ ) g2 (θ ), where

∂ lt† (θ ) ∑ t=1 ∂ b T

1 g†2 (θ ) = T with 1 ∂ lt† (θ ) = ∂b 2ht (θ ) =

εt†2  (θ ) − 1 ht



∂ ht (θ ) ∂b

ηt†2 − 1 ∂ ht (θ ) 2ht (θ ) ∂ b

and 1

H†2 (θ ) = T



T

∑

t=1



∂ ht (θ ) ∂ ht (θ ) ∂ b 2ht2 (θ ) ∂ b 1

.

Theorem 4.4 Suppose that Xt is generated by the model (4.1) satisfying assumptions (A1)(A5). Then the wild bootstrap of section 4.3 has the following asymptotic property:     √  d −1 T (a† −  a) − → N 0, Σ −1 + Ω Ω Σ in probability, 1 3 1 1   √

d −1  − T (b† − b) → N 0, Σ −1 in probability, 2 τ Ω2 Σ2 where Ω3 =



3−κ E 2

1 ∂ ht (θ 0 ) ∂ ht (θ 0 ) ∂ a 2ht2 (θ 0 ) ∂ a

 and

τ=

2 . κ −1

Proof. The proof is the mirror image of Theorem 4.3. Firstly, we observe √

and prove 1 √ T

†  −1 √1  = −H T (b† − b) 2 (θ ) T

  ∂ lt† (θ ) d − N 0, τ Ω 2 ∑ ∂b → t=1

∂ lt† (θ ) t=1 ∂ b T

∑

T

in probability.

78

4 Parametric ARMA(p, q)- GARCH(r, s) Models

Let 1 ∂ lt† (θ ) z†t : = √ T ∂b 1 wt†2 − 1 ∂ ht (θ ) , =√ T 2ht (θ ) ∂ b

t = 1, ..., T.

For each T the sequence z†1 , ..., z†T is independent because w†1 , ..., w†T is independent. We obtain E† (z†t ) = 0 and T

∑ E†



z†t z†t





t=1

 1 T  1 ∂ ht (θ ) ∂ ht (θ ) †4 †2 = E† (wt ) − 2E† (wt ) + 1 ∑ 4h2 (θ ) ∂ b ∂ b T t=1 t  T 1 1 ∂ ht (θ ) ∂ ht (θ ) = ∑ 2 T t=1 2ht (θ ) ∂ b ∂ b p

− →τ Ω 2 . Let c ∈ Rs+r+1 and

 2 p T s2T := ∑ E† c z†t − → c τ Ω 2 c, t=1

then we have the Lyapounov condition for δ = 1 T

1

 

3 

∑ s3 E† c z†t 

t=1 T E† |wt†2 − 1|3 T  1  ∂ ht (θ ) 3 = c   ∂b 8s3T T 3/2 t=1 ht (θ )

∑

=o p (1). Secondly, we observe √ 1

T (a† −  a) = −H†1 (θ )−1 √ T

∂ lt† (θ ) ∑ t=1 ∂ a T

and prove 1 √ T

  ∂ lt† (θ ) d − N 0, Ω 1 + Ω 3 ∑ ∂a → t=1 T

in probability.

4.3 Wild Bootstrap

79

Let 1 ∂ lt† (θ ) y†t := √ T ∂a  wt† ∂ εt (θ ) 1 wt†2 − 1 ∂ ht (θ ) −√ , =√ T 2ht (θ ) ∂ a ht (θ ) ∂ a

t = 1, ..., T.

For each T the sequence y†1 , ..., y†T is independent because w†1 , ..., w†T is independent. We obtain E† (y†t ) = 0 and   † † y E y † ∑ t t T

t=1

1  = E† (wt†4 ) − 2E† (wt†2 ) + 1 T

T

∑

t=1



∂ ht (θ ) ∂ ht (θ ) ∂ a 4ht2 (θ ) ∂ a 1



  1 T 1 − E† (wt†3 ) − E† (wt† ) ∑ 2(h )3/2 (θ ) T t=1 t   ∂ ht (θ ) ∂ εt (θ ) ∂ εt (θ ) ∂ ht (θ ) + ∂a ∂ a ∂a ∂ a  T 1 ∂ εt (θ ) ∂ εt (θ ) †2 1 + E† (wt ) ∑ T t=1 ht (θ ) ∂ a ∂ a   1 T 1 T 1 ∂ ht (θ ) ∂ ht (θ ) 1 ∂ εt (θ ) ∂ εt (θ ) = ∑ + ∑ T t=1 2ht2 (θ ) ∂ a ∂ a T t=1 ht (θ ) ∂ a ∂ a p

− →Ω 1 + Ω 3 .

Let c ∈ R p+q+1 and  2 p   T u2T := ∑ E† c y†t − → c Ω 1 + Ω 3 c, t=1

80

4 Parametric ARMA(p, q)- GARCH(r, s) Models

then we have the Lyapounov condition for δ = 1 3 T 1   ∑ u3 E† c y†t  t=1 T ⎛   3 3 ⎞    † †2 T   wt ∂ εt (θ )  ⎠ 8   w − 1 ∂ ht (θ )  ≤ 3 3/2 ∑ ⎝E† c t  + E† −c √    2ht (θ ) ∂ a  uT T t=1 ht (θ ) ∂ a  ⎛   3 3 ⎞    T   ∂ ht (θ )  ∂ εt (θ )  ⎠ 8 1  1 †  ≤ 3 3/2 ∑ ⎝E† |wt†2 − 1|3  c c  + E† |wt |3  √   2ht (θ )  ht (θ ) ∂a  ∂a  uT T t=1 =o p (1).  Remark 4.3 Theorem 4.4 implies that the wild bootstrap of section 4.3 is weakly consistent only if κ = 3. In other words, if κ = 3, it is necessary to construct a wild p bootstrap in a different way so that E† (wt† ) = 0, E† (wt†2 ) = 1, E† (wt†3 ) → 0 p and E† (wt†4 ) → κ ; however, it is not always easy to obtain such wt† . In the next section we will see the result of a false application, in which the wild bootstrap with standard normal distribution is applied though κ = 9.

4.4 Simulations In this section we demonstrate the large sample properties of the residual bootstrap of section 4.2 and the wild bootstrap of section 4.3. Let us consider the following ARMA(1,1)-GARCH(1,1) model: Xt = a0 + a1 Xt−1 + α1 εt−1 + εt ,  εt = ht ηt , 2 + β1 ht−1 , ht = b0 + b1 εt−1

t = 1, ..., 10, 000.

The simulation procedure is as follows. iid Simulate the bias ηt ∼ 35 t5 for t = 1, ..., 10, 000. a0 Let X0 = E (Xt ) = 1−a , h0 = ε02 = E (εt2 ) = 1−bb0−β , a0 = 0.141, a1 = 1 1 0.433, α1 = −0.162, b0 = 0.007, b1 = 0.135 and β1 = 0.829 and calculate 1/2  2 + β1 ht−1 ηt Xt = a0 + a1 Xt−1 + α1 εt−1 + b0 + b1 εt−1

4.4 Simulations

81

for t = 1, ..., 10, 000. From the data set {X1 , X2 , ..., X10,000 } obtain the QML estimators a0 , a1 ,

1 , b0 , b1 and β1 . α Repeat (step 1)-(step 3) 2,000 times.

1 , Choose one data set {X1 , X2 , ..., X10,000 } and its estimators {a0 , a1 , α b0 , b1 , β1 }. Adopt the residual bootstrap of section 4.2 and the wild bootstrap of section 4.3 based on the chosen data set and estimators. Both bootstrap methods are repeated 2,000 times. √ Compare the simulated density of the QML estimators ( T (a0 − a0 ) etc., thin lines) with the residual and the wild bootstrap approximations √  √ ( T (a∗ − a0 ), T (a† − a0 ) etc., bold lines). 0

0

Figure 4.1 indicates that the residual bootstrap is weakly consistent. On the other hand, Figure 4.2 shows the problem of the wild bootstrap, especially for the GARCH part, which corresponds to the theoretical result in Theorem 4.4 with τ = 1 4 . Note that in this simulation Ω 3 is much smaller than Ω 1 , i.e. Ω 1 ≈ Ω 1 + Ω 3 . Thus, we do not see the problem of the ARMA part, which theoretically exists, in Figure 4.2.

4 Parametric ARMA(p, q)- GARCH(r, s) Models

0.0

0.00

0.02

0.1

0.04

0.2

0.06

0.3

0.08

0.10

0.4

0.12

82

−3

−1

0

1

2

3

−10

√ √ T (a0 − a0 ) and T (a∗0 − a0 )

(b)

√

−5

0

T (a1 − a1 ) and

5

10

√ T (a∗1 − a1 )

0

0.00

0.02

1

0.04

2

0.06

3

0.08

0.10

4

(a)

−2

−10

√

0

1 − α1 ) and T (α

5

√

10

−0.4

∗ − α

1 ) T (α 1

(d)

−0.2

√

0.0

T (b0 − b0 ) and

0.2

0.4

√ T (b∗0 − b0 )

0.00

0.00

0.05

0.05

0.10

0.10

0.15

0.15

0.20

0.25

0.20

0.30

0.25

0.35

(c)

−5

−4

(e)

−2

0

2

−4

4

√ √ T (b1 − b1 ) and T (b∗1 − b1 )

(f)

−2

0

2

4

√ √ ∗  T (β1 − β1 ) and T (β

1 − β1 )

Figure 4.1: Residual Bootstrap

83

0.0

0.00

0.02

0.1

0.04

0.2

0.06

0.3

0.08

0.4

0.10

0.12

0.5

4.4 Simulations

−3

−1

0

1

2

3

√

0

T (a1 − a1 ) and

5

10

√ † T (a1 − a1 )

6

(b)

−5

5

0.10

4

0.08

3

0.06

2

0.04

1

0.02

0

0.00

−10

−5

0

5

10

−0.4

√ † √

1 − α1 ) and T (α T (α 1 − α1 )

(d)

−0.2

√

0.0

T (b0 − b0 ) and

0.2

0.4

√ † T (b0 − b0 )

0.0

0.0

0.1

0.1

0.2

0.2

0.3

0.3

0.4

0.4

0.5

(c)

−10

√ √  T (a0 − a0 ) and T (a†0 − a0 )

0.12

(a)

−2

−4

(e)

−2

0

2

4

−4

√ √  T (b1 − b1 ) and T (b†1 − b1 )

(f)

−2

0

2

4

√ √

T (β1 − β1 ) and T (β1† − β1 )

Figure 4.2: Wild Bootstrap

5 Semiparametric AR(p)-ARCH(1) Models So far we have analysed parametric ARCH and GARCH models. In this chapter the results are extended to semiparametric models, in which the ARCH part is nonparametric. In the first section we introduce the semiparametric AR(p)-ARCH(1) model and show the asymptotic properties of the estimators. Then, as in preceding chapters, possible applications of the residual and the wild bootstrap are proposed and their weak consistency proved. The theoretical results are confirmed by simulations in the last section.

5.1 Estimation Theory Franke, Kreiß and Mammen (2002) investigated asymptotic properties of the Nadaraya-Watson (NW) estimator for the first order nonparametric autoregressive conditional heteroscedasticity (NARCH(1)) model of the form Xt = m(Xt−1 ) + σ (Xt−1 )ηt , where {ηt } is a sequence of i.i.d. random variables with mean 0 and variance 1, and m and σ are unknown smooth functions. They applied the residual, the wild and the autoregression bootstrap techniques to the model and proved their weak consistency. Palkowski (2005) extended the analysis to Fan and Yao’s (1998) estimator and applied the wild bootstrap technique to the model. It is important to note that the NARCH(1) model is fundamentally different from Engle’s ARCH(1) model  Xt = a1 Xt−1 + ht ηt , 2 ht = b0 + b1 εt−1 ,

where {ηt } is a sequence of i.i.d. random variables with mean 0 and variance 1, and a1 , b0 and b1 are unknown parameters. While the NARCH model assumes that variance of the  disturbanceat time t is a function of the observable random variable at time t-1 σ 2 (Xt−1 ) , Engle’s ARCH model assumes that this is a function   2 of the square of the disturbance at time t-1 b0 + b1 εt−1 . Financial time series

K. Shimizu, Bootstrapping Stationary ARMA-GARCH Models, DOI 10.1007/978-3-8348-9778-7_5, © Vieweg+Teubner | GWV Fachverlage GmbH, Wiesbaden 2010

86

5 Semiparametric AR(p)-ARCH(1) Models

is characterised by volatility clustering, that is, “large changes tend to be followed by large changes, of either sign, and small changes tend to be followed by small changes".1 Volatility clustering suggests a time series in which successive disturbances are serially dependent2 and, therefore, we propose the new semiparametric AR-ARCH model, in which the AR part is parametric, the ARCH part is nonparametric and  variance of the disturbance at time t is a function of the disturbance at time t-1 σ 2 (εt−1 ) . Note that the new semiparametric AR-ARCH model is more general than Engle’s ARCH model in that the former analyses the relation2 ship between εt and εt−1 nonparametrically while the latter between εt2 and εt−1 3 parametrically.

5.1.1 Model and Assumptions Assume that {Xt , t = 0, ..., T } are generated by the p-th order parametric autoregressive model with the NARCH(1) errors: p

Xt = a0 + ∑ ai Xt−i + εt , i=1

εt = σ (εt−1 )ηt ,

(5.1)

where {ηt } is a sequence of i.i.d. random variables such that E (ηt ) = 0, E (ηt2 ) = 1, E (ηt4 ) =: κ < ∞ and σ is an unknown smooth function. For simplicity we assume here the presample values {Xt , t = −p, ..., − 1}. The process {εt } is assumed to be stationary and geometrically ergodic. Let π denote the unique stationary distribution. Stationarity and geometric ergodicity of {εt } follow from the following two conditions (cf. Franke, Kreiß and Mammen (2002), section 2.1): The distribution of the i.i.d. innovations ηt possesses a Lebesgue density pη , which satisfies inf pη (e) > 0 e∈C

for all compact sets C. 1 2 3

Mandelbrot (1963). See, e.g., Chang (2006). Note that this model differs from the semiparametric ARCH models of Engle and González-Rivera (1991) in that the former model is estimated with the NW method based on the estimated residuals of the AR part while the latter model adopts the discrete maximum penalised likelihood estimation technique of Tapia and Thompson (1978).

5.1 Estimation Theory

87

σ and σ −1 are bounded on compact sets and E |σ (e)η1 | < 1, |e| |e|→∞

lim sup that is,

σ 2 (e) < 1. 2 |e|→∞ e

lim sup

(A1) and (A2) ensure that the stationary distribution π possesses a strictly positive Lebesgue density, which we denote p. From (5.1) we obtain p(e) =

"

1 pη R σ (u)



e σ (u)

 p(u)du.

We further assume that

σ is twice continuously differentiable and σ , σ  and σ  are bounded. There exists σ0 > 0 such that σ (e) ≥ σ0 for all e ∈ R. E (η16 ) < ∞. pη is twice continuously differentiable. pη , pη and pη are bounded and sup |epη (e)| < ∞. e∈R

g, h → 0, T h4 → ∞, T h5 → B2 ≥ 0 and g ∼ T −α with 0 < α ≤

1 9

for T → ∞.

The kernel function K has compact support [-1, 1]. K is symmetric, nonnegative and three times continuously differentiable with K(1) = K  (1) = 0 # and K(v)dv = 1.

5.1.2 NW Estimation In the semiparametric AR(p)-ARCH(1) model it is necessary to estimate the parametric AR part firstly. The nonparametric ARCH part is then estimated based on the residuals of the AR part. Because we have already analysed the parametric AR(p) model with heteroscedastic errors in chapter 3, we concentrate here on the nonparametric ARCH part. For simplicity of the discussion we adopt the OLS estimation for the AR(p) part and make use of the results of chapter 3.

88

5 Semiparametric AR(p)-ARCH(1) Models

5.1.2.1 The imaginary case where εt are known Firstly, we analyse the imaginary case where {εt , t = 1, ..., T } are known. Let mt :=

ηt2 −1 √ , κ −1

then E (mt ) = 0, E (mt2 ) = 1 and √ εt2 = σ 2 (εt−1 ) + κ − 1σ 2 (εt−1 )mt ,

t = 1, ..., T.

(5.2)

The NW estimator of σ 2 is −1 T 2

2 (e) = T ∑i=1 Kh (e − εi−1 )εi , σ h T T −1 ∑ j=1 Kh (e − ε j−1 )

where Kh (•) = 1h K(•/h), K(•) is a kernel function and h denotes the bandwidth. Now we sketch the result of Franke, Kreiß and Mammen (2002) observing the decomposition  √ 

T h σh2 (e) − σ 2 (e) h T 2 2 T ∑i=1 Kh (e − εi−1 )σ (εi−1 )(ηi − 1) = 1 T T ∑ j=1 Kh (e − ε j−1 )   h T 2 2 T ∑i=1 Kh (e − εi−1 ) σ (εi−1 ) − σ (e) + 1 T T ∑ j=1 Kh (e − ε j−1 ) and analysing asymptotic properties of each part separately. Lemma 5.1 Suppose that εt is generated by the model (5.2) satisfying assumptions (A1)(A2) and (B1)-(B4). Then for all e ∈ R $   h T d Kh (e − εt−1 )σ 2 (εt−1 )(ηt2 − 1) − → N 0, τ 2 (e) , ∑ T t=1 #

where τ 2 (e) = (κ − 1)σ 4 (e)p(e) K 2 (v)dv.

Proof. Franke, Kreiß and Mammen (1997), Lemma 6.2 (i). See also Lemma 2.1 in Palkowski (2005).  Lemma 5.2 Suppose that εt is generated by the model (5.2) satisfying assumptions (A1)(A2) and (B1)-(B4). Then for all e ∈ R 1 T

T

p

→ p(e). ∑ Kh (e − εt−1 ) −

t=1

5.1 Estimation Theory

89

Proof. Franke, Kreiß and Mammen (1997), Lemma 6.3 (i). See also Lemma 2.3 in Palkowski (2005).  Lemma 5.3 Suppose that εt is generated by the model (5.2) satisfying assumptions (A1)(A2) and (B1)-(B4). Then for all e ∈ R $   p h T Kh (e − εt−1 ) σ 2 (εt−1 ) − σ 2 (e) − → b(e), ∑ T t=1   # where b(e) = B v2 K(v)dv p (e)(σ 2 ) (e) + 12 p(e)(σ 2 ) (e) .

Proof. Franke, Kreiß and Mammen (1997), Lemma 6.4 (i). See also Lemma 2.2 in Palkowski (2005).  Theorem 5.1 Suppose that εt is generated by the model (5.2) satisfying assumptions (A1)(A2) and (B1)-(B4). Then for all e ∈ R    √ 

b(e) τ 2 (e) d , 2 T h σh2 (e) − σ 2 (e) − →N . p(e) p (e)

Proof. Lemma 5.1-5.3 together with the Slutsky theorem yield the desired result.  5.1.2.2 The standard case where εt are unknown Secondly, we analyse the standard case where {εt , t = 1, ..., T } are unknown. In this case we adopt the two-step estimation, where the residuals of the AR part p

εt = Xt − a0 − ∑ ai Xt−i ,

t = 0, ..., T

i=1

are used to estimate the ARCH part. The NW estimator of σ 2 is then

σh2 (e) =

2 T −1 ∑Ti=1 Kh (e − ε i−1 )εi . T −1 ∑Tj=1 Kh (e − ε j−1 )

We observe  √  T h σh2 (e) − σ 2 (e) =



h T

  2 2 (e)  ) ε − σ ∑Ti=1 Kh (e − ε i i−1 1 T

∑Tj=1 Kh (e − ε j−1 )

and analyse the asymptotic properties of each part separately. In fact, the estimator

2 (e) in §5.1.2.1, as we will see in Theorem σh2 (e) is asymptotically as efficient as σ h 5.2.

90

5 Semiparametric AR(p)-ARCH(1) Models

Remark 5.1 Analogously to chapter 3, we denote

εt − εt = xt  ( a − a) =

p

∑ xt,k (ak − ak ),

k=0

where xt = (1 Xt−1 Xt−2 ... Xt−p ) = (xt,0 xt,1 xt,2 ... xt,p ) , a = (a0 a1 ... a p )

 a = (a0 a1 ... a p )

and

throughout this chapter without further notice.

Lemma 5.4 Suppose that Xt is generated by the model (5.1) satisfying assumptions (A1)(A2) and (B1)-(B4). Then for all e ∈ R $

h T

$

T

i−1 )εi ∑ Kh (e − ε

i=1

2

=

h T

T

∑ Kh (e − εi−1 )εi2 + o p (1).

i=1

Proof. Here it suffices to show $

h T

T

∑

 2 Kh (e − ε i−1 ) − Kh (e − εi−1 ) εi = o p (1)



(5.3)

i=1

and $

h T

  2 2   = o p (1). − K (e − ε ) ε ε i i−1 h ∑ i T

i=1

(5.4)

5.1 Estimation Theory

91

Firstly, a Taylor expansion for the left hand side of (5.3) yields $

h T

T

∑



 2 Kh (e − ε i−1 ) − Kh (e − εi−1 ) εi

i=1

     e − ε 1 T e − εi−1 i−1 =√ ∑ K εi2 −K h h T h i=1    e − εi−1 εi−1 − ε 1 T i−1 = √ ∑ K h h T h i=1    2 3 1  1  e − εi−1 εi−1 − ε εi−1 − ε i−1 i−1 + K ( e) εi2 + K 2 h h 6 h   T 1 2  e − εi−1 =√ ε K a − a) xi−1  ( ∑ i h T h3 i=1   2 T 1 2  e − εi−1  + √ ε K ( a − a) x i−1 ∑ i h 2 T h5 i=1  3 T 1 + √ εi2 K  ( e) xi−1  ( a − a) ∑ 6 T h7 i=1    √ 1 T εi2  e − εi−1  √ K ≤ T ( a − a) xi−1 ∑    T i=1 h3 h 1 + 2



1 T

ε2 ∑ √Ti h5 K  i=1 T



e − εi−1 h

   cp    1 + √ max K (u) T 6 T 2 h7 u∈R   =O(1)

T

∑

i=1



=O p (1)



p

∑

2 xi−1,k

k=0

p

T 



∑ |xi−1,k |

3

k=0



=O p (1)

e−ε

k=0



p

εi2

∑ (ak − ak )2



p

T 3/2 ∑ |ak − ak |3 ,

 

e−ε



=O p (1)

k=0



=O p (1)



where e denotes a suitable value between hi−1 and hi−1 , and c p a constant such that 0 < c p < ∞. Now let   εi2  e − εi−1  zi−1 := √ K xi−1  , h h3

92

5 Semiparametric AR(p)-ARCH(1) Models

then on the one hand  1  E zi−1  zi−1 T %  & xi−1  xi−1 εi4   2 e − εi−1 1 K = E T h3 h     1 2 ≤ 3 max K  (u) E xi−1  xi−1 εi4 T h u∈R      =O(1)

=O(1)

=o(1),

thus Remark B.1 yields

1 T

T

∑ zi−1  =

i=1

1 T

    E z ∑ i−1 Fi−2 + o p (1). T

i=1

5.1 Estimation Theory

93

On the other hand, because xi−1 is measurable to Fi−2 , we obtain   1 T   E zi−1  Fi−2 ∑ T i=1    1 T xi−1  εi2  e − εi−1  √ = ∑E K Fi−2  T i=1 h h3     1 T xi−1  2  e − εi−1  = ∑ √ E σ (εi−1 )K Fi−2  T i=1 h3 h   " 1 T xi−1  e−u = ∑√ σ 2 (u)K  p(u)du T i=1 h3 h =

1 T

=−

xi−1 

T

∑ √h3

"

i=1

1 T

xi−1  √ h i=1 T

∑

σ 2 (e − hv)K  (v)p(e − hv)(−h)dv "

σ 2 (e − hv)K  (v)p(e − hv)dv

"    1   xi−1   2 √ ∑ h σ (e) − σ 2 (e)hv + 2 σ 2 ( e )h2 v2 i=1   1  K (v) p(e) − p (e)hv + p ( e )h2 v2 dv 2 " 1 2 1 T   = − √ σ (e)p(e) K (v)dv ∑ xi−1 h    T i=1

=−

1 T

T

=0

  " √  + h σ 2 (e)p (e) + σ 2 (e)p(e) vK  (v)dv   



1 T

T

∑ xi−1



i=1

=−1

   √ 1 1   σ 2 (e)p ( e ) + σ 2 (e)p (e) + σ 2 ( e )p(e) − h3 2 2  " T 1 v2 K  (v)dv ∑ xi−1  T i=1    √ +

=O(1)

h5

2



σ

2



 "   1 2  3 

(e)p (e ) + σ (e )p (e) v K (v)dv T    

 √   " h7 2   1 4  σ (e )p (e ) v K (v)dv − 4 T    =o p (1),

=O(1)

=O(1)

T

∑ xi−1

i=1





T

∑ xi−1

i=1



94

5 Semiparametric AR(p)-ARCH(1) Models

and, therefore, we obtain T

1 T

∑ zi−1  = o p (1).

i=1

Analogously, we can prove 1 T



ε2 ∑ √Ti h5 K  i=1 T

e − εi−1 h



p

2 = o p (1), ∑ xi−1,k

k=0

which shows (5.3). Secondly, the left hand side of (5.4) equals $

h T

  εi2 − εi2

T

i−1 ) ∑ Kh (e − ε

i=1 T

   e − ε 1 i−1 εi2 − εi2 =√ ∑ K h T h i=1    1 T      ≤ max K(u) √ ∑ εi2 − εi2  u∈R T h i=1     2 2 T  1 T     ≤ max K(u) √ ∑ εi − εi + √ ∑ εi (εi − εi )  u∈R T h i=1 T h i=1      T  2 2 T   1     = max K(u) √ ∑ xi (a −  a) + √ ∑ εi xi (a −  a)  u∈R T h i=1 T h i=1     1 p 1 T p 2   2 T ∑ (ak − ak ) ≤ max K(u) √ ∑ ∑ xi,k T i=1 u∈R Th k=0 k=0           + max K(u)  u∈R

=O p (1)





√

T

2

εi xi  ∑ 2 T h i=1



=o p (1)



=O p (1)

√ T (a −  a)    =O p (1)

=o p (1). Here we obtain T 2 εi xi  = o p (1) yt  := √ ∑ T 2 h i=1

5.1 Estimation Theory

95

because     T 2 E yt  = √ E (ηi ) E σ (εi−1 )xi  = 0, ∑ T 2 h i=1    =0

    4 T E yt  yt = 2 ∑ E εi ε j xi  xj T h i, j=1  4 1 T  2   = ∑ E εi xi xi T h T i=1 =o(1), and, therefore, from Chebyshev’s inequality we obtain for every δ > 0      1    P yt  − E yt   > δ ≤ 2 E yt  yt = o(1), δ   p that is, according to the definition yt  → E yt  = 0.



Lemma 5.5 Suppose that Xt is generated by the model (5.1) satisfying assumptions (A1)(A2) and (B1)-(B4). Then for all e ∈ R

1 T

T

j−1 ) = ∑ Kh (e − ε

j=1

1 T

T

∑ Kh (e − ε j−1 ) + o p

j=1



 1 √ . Th

Proof. Analogously to Lemma 5.4, we prove $

h T

T

∑

i=1



 Kh (e − ε i−1 ) − Kh (e − εi−1 ) = o p (1).

96

5 Semiparametric AR(p)-ARCH(1) Models

A Taylor expansion for the left hand side yields $  h T   K (e − ε ) − K (e − ε ) i−1 i−1 h ∑ h T i=1      e − ε e − εi−1 1 T i−1 −K =√ ∑ K h h T h i=1    e − εi−1 εi−1 − ε 1 T i−1 = √ ∑ K h h T h i=1    2 3 1  1  e − εi−1 εi−1 − ε εi−1 − ε i−1 i−1 + K ( e) + K 2 h h 6 h   T 1 e − εi−1 =√ K a − a) xi−1  ( ∑ 3 h T h i=1   2 T 1  e − εi−1 + √ K a − a) xi−1  ( ∑ h 2 T h5 i=1  3 T 1 + √ K  ( e) xi−1  ( a − a) ∑ 6 T h7 i=1    √ 1 T 1  e − εi−1  √ K ≤ T ( a − a) xi−1 ∑    T i=1 h3 h   =O p (1)  1 + 2

 

=o p (1)

1 T

T

1

∑ √T h5 K

i=1





e − εi−1 h 





p

∑

k=0

T

∑

∑ (ak − ak )2

k=0



p



=O p (1)



p

∑ |xi−1,k |

T 3/2 ∑ |ak − ak |3

=O p (1)

=O p (1)

3

i=1 k=0

=O(1)

T 

=o p (1)

   cp    1 + √ max K (u) T 6 T 2 h7 u∈R  

p

2 xi−1,k





 

k=0





=o p (1), where e denotes a suitable value between that 0 < c p < ∞.

e−ε i−1 h

and

e−εi−1 h ,

and c p a constant such 

Theorem 5.2 Suppose that Xt is generated by the model (5.1) satisfying assumptions (A1)-

5.1 Estimation Theory

97

(A2) and (B1)-(B4). Then for all e ∈ R  √  d T h σh2 (e) − σ 2 (e) − →N



 b(e) τ 2 (e) , 2 . p(e) p (e)

Proof. The left hand side equals

√

  T h σh2 (e) − σ 2 (e) =



h T

  2 2 ∑Ti=1 Kh (e − ε i−1 ) εi − σ (e) 1 T

∑Tj=1 Kh (e − ε j−1 )

From Lemma 5.1-5.3 we already know $

h T

    d 2 2 2 K (e − ε ) ε − σ (e) → N b(e), τ (e) i−1 ∑ h i T

i=1

1 T

T

∑ Kh (e − ε j−1 ) → p(e). p

j=1

From Lemma 5.5 we already have

1 T

T

j−1 ) = ∑ Kh (e − ε

j=1

1 T

T

∑ Kh (e − ε j−1 ) + o p (1).

j=1

Together with the Slutsky theorem it suffices to show $ $ =

h T h T

  2 2   ε ε − σ K (e − ) (e) i i−1 h ∑ T

i=1

  2 2 − (e) + o p (1). K (e − ε ) ε σ i−1 h ∑ i T

i=1

and

.

98

5 Semiparametric AR(p)-ARCH(1) Models

From Lemma 5.4 and 5.5 we obtain $   h T 2 2   − K (e − ε ) ε σ (e) i i−1 h ∑ T i=1 $ √ 1 T h T 2 2   − K (e − ε ) ε T h σ (e) = i i−1 i−1 ) h ∑ ∑ Kh (e − ε T i=1 T i=1 $ h T = ∑ Kh (e − εi−1 )εi2 + o p (1) T i=1    √ 1 1 T 2 − T hσ (e) ∑ Kh (e − εi−1 ) + o p √T h T i=1 $   h T Kh (e − εi−1 ) εi2 − σ 2 (e) + o p (1) − o p (1). = ∑ T i=1 

5.2 Residual Bootstrap Franke, Kreiß and Mammen (2002) proved asymptotic validity of the residual, the wild and the autoregression bootstrap technique applied to the NW estimators for the NARCH(1) model. In this and the following section we introduce possible ways to bootstrap the semiparametric AR(p)-ARCH(1) model based on the results of Franke, Kreiß and Mammen (2002). In this section a residual bootstrap technique is proposed and its weak consistency proved. A residual bootstrap method can be applied to the model (5.1) as follows: Obtain the OLS estimator  a and calculate the residuals p

εt = Xt − a0 − ∑ ai Xt−i ,

t = 0, ..., T.

i=1

Obtain the NW estimator with a bandwidth g

σg2 (e) =

2 T −1 ∑Ti=1 Kg (e − ε i−1 )εi . T −1 ∑Tj=1 Kg (e − ε j−1 )

5.2 Residual Bootstrap

99

Compute the estimated bias  2 −1/2  εt

t = −1 κ −1 m 2 σg (ε t−1 ) for t = 1, ..., T and the standardised estimated bias mt =



t − μ m σ

for t = 1, ..., T , where = μ

1 T

T

∑ m t

σ 2 =

and

t=1

1 T

T

 )2 . mt − μ ∑ (

t=1

Obtain the empirical distribution function FT (x) based on mt defined by FT (x) :=

1 T

T

∑ 1 (mt ≤ x) .

t=1

Generate the bootstrap process εt∗2 by computing √ ∗ 2 εt∗2 = σg2 (ε t−1 ) + κ − 1σg (ε t−1 )mt , mt∗ ∼ FT (x), iid

t = 1, ..., T.

Build the bootstrap model

√ ∗2 ∗ εt∗2 = σ ∗2 (ε t−1 ) + κ − 1σ (ε t−1 )mt

' ∗2 and calculate the NW estimator σ h (e) with another bandwidth h −1 T K (e − ε )ε ∗2 ∑i=1 h i−1 i ' ∗2 (e) = T . σ h −1 T ∑Tj=1 Kh (e − ε j−1 )

Analogously to the previous section, we observe the decomposition  √ ' T h σh∗2 (e) − σg2 (e) √ h T ∗ 2   T ∑i=1 Kh (e − εi−1 ) κ − 1σg (εi−1 )mi = 1 T  T ∑ j=1 Kh (e − ε j−1 )   h T 2  2  T ∑i=1 Kh (e − εi−1 ) σg (εi−1 ) − σg (e) + 1 T  T ∑ j=1 Kh (e − ε j−1 )

100

5 Semiparametric AR(p)-ARCH(1) Models

and analyse the asymptotic properties of each part separately. For the proof of weak consistency of the residual bootstrap we need a stronger condition than just to prove asymptotic normality of the NW estimators. E |ηt |6 < ∞. Lemma 5.6 Suppose that Xt is generated by the model (5.1) satisfying assumptions (A1)(A2) and (B1)-(B4). Then for all e ∈ R     p σg2 (e) − → σ 2 (e)

 and

σg2



  p (e) − → σ 2 (e).

Proof. We observe the decomposition 

σg2



(e) =

1 T g2

∑Ti=1 K 

1 Tg

−



∑Tj=1 K 1 Tg

e−ε i−1 g



∑Ti=1 K



2 εi  

e−ε j−1 g



e−ε i−1 g



1 Tg



2 εi



∑Tj=1 K

1 T g2



∑Ti=1 K  2 



e−ε i−1 g



e−ε j−1 g

and   2 1 T  e−ε i−1   K εi ∑ g T g3 i=1 2   σg (e) = e−ε j−1 1 T T g ∑ j=1 K g   2 1 T  e−ε i−1 K εi ∑ 2 i=1 g Tg −2   2 e−ε j−1 1 T K ∑ j=1 Tg g      2 e−ε 1 1 T T  e−ε i−1 i−1  K ε K ∑ ∑ i T g i=1 g g T g3 i=1 −   2  e−ε j−1 1 T T g ∑ j=1 K g       2 e−ε 1 1 T i−1 εi T g2 ∑Ti=1 K  e−εgi−1 T g ∑i=1 K g +2 .    3 e−ε j−1 1 T K ∑ T g j=1 g

5.2 Residual Bootstrap

101

Together with the Slutsky theorem it suffices to show      1 T i−1  e − εi−1  e − ε (i) −K = o p (1), ∑ K T g2 i=1 g g       1 T i−1 2  e − εi−1 2  e − ε (ii) εi − K εi = o p (1), ∑ K T g2 i=1 g g      1 T i−1  e − εi−1  e − ε (iii) −K = o p (1), ∑ K T g3 i=1 g g       1 T i−1 2  e − εi−1 2  e − ε (iv) εi − K εi = o p (1). ∑ K T g3 i=1 g g A Taylor expansion for the left hand side of (i) yields      1 T i−1  e − ε  e − εi−1 −K ∑ K T g2 i=1 g g     2 1  1 T εi−1 − ε εi−1 − ε i−1 i−1  e − εi−1 + K ( e) = 2∑ K T g i=1 g g 2 g    2 T T 1 1 e − εi−1   = 3 ∑ K  a − a) + K ( e ) x ( a − a) xi−1  ( i−1 ∑ T g i=1 g 2T g4 i=1    √ 1 1 T  e − εi−1  ≤ K T ( a − a) xi−1  ∑    T i=1 T g6 g   =O p (1)  1 + 2

 

=o p (1)

1 T

p p 1  2 K ( e ) x T ∑ T g4 ∑ i−1,k ∑ (ak − ak )2 i=1 k=0 k=0     T

=O p (1)

=o p (1)

=o p (1), e−ε

e−ε

where e denotes a suitable value between gi−1 and gi−1 . To prove (ii) we need to show      1 T i−1  e − ε  e − εi−1 −K εi2 = o p (1) ∑ K T g2 i=1 g g

(5.5)

102

5 Semiparametric AR(p)-ARCH(1) Models

and 1 T  ∑K T g2 i=1



e − ε i−1 g



 2 εi − εi2 = o p (1).

(5.6)

Firstly, a Taylor expansion for the left hand side of (5.5) yields      1 T i−1  e − ε  e − εi−1 εi2 −K ∑ K T g2 i=1 g g     2   1 1 T ε ε − ε ε − ε e − i−1 i−1 i−1 i−1 i−1 + K  ( εi2 = 2 ∑ K  e) T g i=1 g g 2 g   1 T 2  e − εi−1 = 3 ∑ εi K a − a) xi−1  ( T g i=1 g 2 1 T 2    + ε K ( e ) x ( a − a) i−1 ∑ i 2T g4 i=1    √ 1 T εi2  e − εi−1   T ( a − a) ≤ K xi−1 ∑ 6    T i=1 T g g   =O p (1)  1 + 2

 

=o p (1)

1 T

p p εi2  2 ∑ T g4 K ( e) ∑ xi−1,k T ∑ (ak − ak )2 i=1 k=0 k=0     T

=O p (1)

=o p (1)

=o p (1),

where e denotes a suitable value between

e−ε i−1 g

and

e−εi−1 g .

Secondly, the left

5.2 Residual Bootstrap

103

hand side of (5.6) equals    1 T  e − ε i−1 2 − ε 2 ε K ∑ i i T g2 i=1 g    1 T      ≤ max K  (u) 2 ∑ εi2 − εi2  T g i=1 u∈R     2 1 T  2 T      ≤ max K (u) ∑ εi − εi + T g2 ∑ εi (εi − εi )  T g2 i=1 u∈R i=1      T  T  2 1 2      = max K (u) ∑ xi  (a − a) + T g2 ∑ εi xi  (a − a)  T g2 i=1 u∈R i=1    1  p T p 1    2 2 ≤ max K (u) 2 T ∑ (ak − ak ) ∑ ∑ xi,k Tg T i=1 u∈R k=0 k=0           + max K  (u)  u∈R

=O p (1)







T

2 T 3 g4





=O p (1)

 √   T a −  a     =O p (1)

∑ εi xi  

i=1



=o p (1)

=o p (1). A Taylor expansion for the left hand side of (iii) yields      1 T i−1  e − ε  e − εi−1 −K ∑ K T g3 i=1 g g =

1 T  εi−1 − ε i−1 ∑ K ( e) g T g3 i=1

1 T  ∑ K ( e)xi−1  (a − a) T g4 i=1    T   √  1     1   ≤ max K (u) T  a − a xi−1 ∑ T i=1 T g8 u∈R       =O p (1) =

=o p (1)

=o p (1), where e denotes a suitable value between

e−ε i−1 g

and

e−εi−1 g .

104

5 Semiparametric AR(p)-ARCH(1) Models

To prove (iv) it suffices to show

     1 T i−1  e − ε  e − εi−1 εi2 = o p (1) −K ∑ K T g3 i=1 g g

(5.7)

and 1 T  ∑K T g3 i=1



e − ε i−1 g

  2 εi − εi2 = o p (1).

(5.8)

Firstly, a Taylor expansion for the left hand side of (5.7) yields       e − e − ε ε 1 T i−1 i−1 εi2 − K  ∑ K  T g3 i=1 g g 1 T  εi−1 − ε i−1 2 εi K ( e) ∑ 3 T g i=1 g  1 T        ≤ max K  (u) 4 ∑ εi2 xi−1  ( a − a)  T g i=1 u∈R    T   √  εi2     1   ≤ max K (u) T  a − a xi−1 ∑ 8 T i=1 T g u∈R      =O p (1)  =

=o p (1)

=o p (1),

where e denotes a suitable value between

e−ε i−1 g

and

e−εi−1 g .

Secondly, the left

5.2 Residual Bootstrap

105

hand side of (5.8) equals    1 T  e − ε i−1 K εi2 − εi2 ∑ 3 T g i=1 g    1 T      ≤ max K  (u) 3 ∑ εi2 − εi2  T g i=1 u∈R     2 1 T  2 T      ≤ max K (u) ∑ εi − εi + T g3 ∑ εi (εi − εi )  T g3 i=1 u∈R i=1      T  2 1 2 T        = max K (u) ∑ xi (a − a) + T g3 ∑ εi xi (a − a)  T g3 i=1 u∈R i=1    1  p T p 1    2 2 ≤ max K (u) 3 T ∑ (ak − ak ) ∑ ∑ xi,k Tg T i=1 u∈R k=0 k=0           + max K  (u)  u∈R

=O p (1)

 



2

T





=O p (1)

  √  T a −  a     =O p (1)

εi xi   ∑ 3 6 T g i=1



=o p (1)

=o p (1).  Lemma 5.7 Suppose that Xt is generated by the model (5.1) satisfying assumptions (A1)(A2) and (B1)-(B4). Then for all e ∈ R h T

 2 h σg2 (ε = t−1 ) T

T

t−1 ) ∑ Kh2 (e − ε

t=1

T

∑ Kh2 (e − εt−1 )σ 4 (εt−1 ) + o p (1).

t=1

Proof. Here it suffices to show h T

T

∑



 2 4 Kh2 (e − ε t−1 ) − Kh (e − εt−1 ) σ (εt−1 ) = o p (1)

(5.9)

t=1

and h T

T

t−1 ) ∑ Kh2 (e − ε

t=1



 2 σg2 (ε − σ 4 (εt−1 ) = o p (1). t−1 )

(5.10)

106

5 Semiparametric AR(p)-ARCH(1) Models

Firstly, a Taylor expansion for the left hand side of (5.9) yields h T

T

∑



 2 4 Kh2 (e − ε ) − K (e − ε ) t−1 t−1 σ (εt−1 ) h

t=1

     1 T t−1 2 e − ε 2 e − εt−1 = σ 4 (εt−1 ) −K ∑ K T h t=1 h h    1 T  2  e − εt−1 εt−1 − ε t−1 = ∑ K T h t=1 h h  2 εt−1 − ε 1  2  t−1 ( e) σ 4 (εt−1 ) + K 2 h    1 T  2  e − εt−1  4 = 2∑ K εt−1 − ε t−1 σ (εt−1 ) T h t=1 h 2 1 T  2   4 + K ( e) εt−1 − ε t−1 σ (εt−1 ) ∑ 3 2T h t=1        1 T   2  4     K max (u) ε − ε max σ (u) ≤       t−1 t−1 ∑ T h2 t=1 u∈R u∈R       +

=O(1)

2 1 T  εt−1 − ε t−1 ∑ 3 2T h t=1



=O(1)

         max  K 2 (u) max σ 4 (u) u∈R u∈R      =O(1)

=O(1)

=o p (1), where e denotes a suitable value between equation from

e−ε t−1 h

and

e−εt−1 h .

 1 T    ε − ε  ∑ t−1 t−1  T h2 t=1  √ 1 1 T  |xt−1 | T | a − a| ≤√ ∑    T h4 T t=1   =O p (1)   =O p

=O p (1)

1 √ T h4



= o p (1)

We obtain the last

5.2 Residual Bootstrap

107

and 2 1 T   ε − ε ∑ t−1 t−1 T h3 t=1  p 1 1 T p 2 ≤ 3 xt−1,k T ∑ (ak − ak )2 ∑ ∑ Th T t=1 k=0 k=0       = Op

=O p (1)

1 T h3



=O p (1)

= o p (1).

Secondly, a Taylor expansion for the left hand side of (5.10) yields

h T

T

∑

Kh2 (e − ε t−1 )

  2 4 2 σg (ε − σ (εt−1 ) t−1 )

t=1 T

   % 2 1 t−1 2 e − ε 4 2 = σg (ε − σ (ε t−1 ) t−1 ) ∑K T h t=1 h &  4 ) − σ ( ε ) + σ 4 (ε t−1 t−1   % 2 1 T 2 e − ε t−1 2 = K σg2 (ε t−1 ) − σ (ε t−1 ) ∑ T h t=1 h   & 2 (ε ) − σ 2 (ε ) + 2 σ 2 (ε ) σ t−1 t−1 g t−1        1 T 2 e − ε t−1 4  K σ ( ε ) ε − ε + t−1 t−1 t−1 ∑ T h t=1 h 2 1  4   , + σ ( e) ε t−1 − εt−1 2 where e denotes a suitable value between ε t−1 and εt−1 . The first and the second

108

5 Semiparametric AR(p)-ARCH(1) Models

summand are smaller than    2 1  2  2 (u) − σ 2 (u) max T h σ max (u) K  g T h2 u∈R u∈R      =O p (1)

=O(1)

      2       +√ max T hσg2 (u) − σ 2 (u) max K 2 (u) max σ 2 (u) 3 u∈R u∈R u∈R Th        √

=O p (1)

=O(1)

=O(1)

=o p (1) and



      1 T    2   4  (u) ε t−1 − εt−1  max K (u) max  σ ∑ T h t=1 u∈R u∈R          =O(1)

=O p √ 1

=O(1)

T h2

     2 1 T      max K 2 (u) max  σ 4 (u) + ε t−1 − εt−1 ∑ 2T h t=1 u∈R u∈R          

=O p √1

=O(1)

=O(1)

Th

=o p (1), 

respectively, which shows (5.10). Lemma 5.8 Suppose that Xt is generated by the model (5.1) satisfying assumptions (A1)(A2) and (B1)-(B5). Then the residual bootstrap of section 5.2 has the following asymptotic property for all e ∈ R: $   √ h T ∗ d 2 Kh (e − ε → N 0, τ 2 (e) in probability. t−1 ) κ − 1σg (ε t−1 )mt − ∑ T t=1

Proof. We observe

φt∗ (e) :=

$

√ h  ∗ Kh e − ε κ − 1σg2 (ε t−1 t−1 )mt , T

t = 1, ..., T.

For each T the sequence φ1∗ , ..., φT∗ is independent because m∗1 , ..., m∗T is independent. Here we obtain   $h √ ∗ ∗ 2 Kh (e − ε E∗ φt (e) = t−1 ) κ − 1σg (ε t−1 )E∗ (mt ) = 0 T

5.2 Residual Bootstrap

109

and from Lemma 5.7   T s2T : = ∑ E∗ φt∗2 (e) t=1

=

h T

2   2 2 (ε )  K (e − ε ) κ − 1 σ E∗ (mt∗2 ) t−1 ∑ h g t−1    t=1 T

  = κ −1



 h = κ −1 T

=1

h T

T

∑



Kh2 (e − εt−1 )σ 4 (εt−1 ) + o p (1)

t=1

T

∑ Kh2 (e − εt−1 )σ 4 (εt−1 ) + o p (1)

t=1

p

→ τ 2 (e). Then we have the Lyapounov condition for δ = 1 1  ∗ 3 ∑ 3 E∗ φt (e) t=1 sT 3/2  ∗ 3 $  T   3 κ −1 E∗ mt  h3 1   2 = t−1 )σg (ε t−1 ) ∑ Kh (e − ε T T t=1 s3T   3/2  ∗ 3   3 κ −1 E∗ mt  1 1 T 3 e − ε  t−1  2 √ = σg (ε t−1 ) ∑K h s3T T h3 T t=1  3/2  ∗ 3  3 κ −1 E∗ mt  1 1 T  2  3 √ max K (u) = t−1 ) ∑ σg (ε s3T T h3 u∈R   T t=1    T



log3 T √ T h3 =o p (1).

=O p



=O p (1)

=O p (log3 T )

110

5 Semiparametric AR(p)-ARCH(1) Models

We obtain the last equation from (B5) and

σg2 (ε t−1 ) 2 ∑Ti=1 Kg (ε t−1 − εi−1 )εi ∑Tj=1 Kg (ε t−1 − ε j−1 )  ≤ max ε 2 =

i

1≤i≤T

  ≤2 max εi2 + (εi − εi )2 1≤i≤T

≤2 max εi2 +2 max 1≤i≤T 1≤i≤T   



2

p

∑ xi,k (ak − ak )

k=0

=O p (log T )

p

p

2 ∑ xi,k ∑ (ak − ak )2 1≤i≤T

≤O p (log T ) + 2 max 

≤O p (log T ) + 2 

k=0

1 T

T

k=0



p

∑∑

2 xi,k

i=1 k=0



=O p (1)

T 



p

∑ (ak − ak )

2

k=0



=O p (1)



=O p (log T ). The central limit theorem for triangular arrays applied to the sequence φ1∗ , ..., φT∗ shows T

∑ φt∗ (e) → N d



 0, τ 2 (e)

in probability.

t=1

 Lemma 5.9 Suppose that Xt is generated by the model (5.1) satisfying assumptions (A1)(A2) and (B1)-(B4). Then for all e ∈ R $

h T

T

  p 2 σg2 (ε → b(e). t−1 ) − σg (e) −

t−1 ) ∑ Kh (e − ε

t=1

5.2 Residual Bootstrap

111

Proof. The left hand side can be decomposed into the following three summands: $ $ =

T



 2 σg2 (ε t−1 ) − σg (e)



 σg2 (εt−1 ) − σg2 (e)

h T

t=1

h T

∑ Kh (e − εt−1 )

t−1 ) ∑ Kh (e − ε T

t=1

$

+ $ +

  2 σg2 (ε t−1 ) − σg (εt−1 )

T

h T

t=1

h T

∑

∑ Kh (e − εt−1 ) T



  2 Kh (e − ε σg2 (ε t−1 ) − Kh (e − εt−1 ) t−1 ) − σg (e) .

t=1

From Lemma 5.3 and 5.6 we obtain $

h T

T



∑ Kh (e − εt−1 )

t=1

 p σg2 (εt−1 ) − σg2 (e) → b(e).

112

5 Semiparametric AR(p)-ARCH(1) Models

The second summand equals $   h T 2 (ε ) − σ 2 (ε K (e − ε σ ) ) t−1 t−1 t−1 h ∑ g g T t=1      1 T e − εt−1 2 =√ ∑ K σg2 (ε t−1 ) − σ (ε t−1 ) h T h t=1     2 ) − σ ( ε ) − σg2 (εt−1 ) − σ 2 (εt−1 ) + σ 2 (ε t−1 t−1 

  e − εt−1   2 2 T g σg (ε t−1 ) − σ (ε t−1 ) h    T e − εt−1   2 1 1 T g σg (εt−1 ) − σ 2 (εt−1 ) − ∑√ K T t=1 gh h    e − εt−1   2 1 T 1 2 T g σ (ε + ∑√ K t−1 ) − σ (εt−1 ) T t=1 gh h       e − εt−1 1 T 1   √ K max T gσg2 (u) − σ 2 (u) ≤ ∑ T t=1 gh h u∈R   

1 = T

T

1 ∑ √gh K t=1

 +

1 T

T

1

∑ √gh K



t=1

e − εt−1 h

=O p (1)











max u∈R 



    T gσg2 (u) − σ 2 (u)   =O p (1)

e − εt−1   2  T g σ ( e)(ε t−1 − εt−1 ) h t=1       e − εt−1 1 T 1   √ K ≤ max T gσg2 (u) − σ 2 (u) ∑ T t=1 gh h u∈R    +

1 T

 +

T

1

∑ √gh K

1 T

T

1 ∑ √h K t=1

e − εt−1 h



=O p (1)

     √     xt−1   T a −  a max  σ 2 (u)    u∈R   =O p (1)

=o p (1). Now let

1 Zt−1 := √ K gh



e − εt−1 h

 ,

=O p (1)

5.2 Residual Bootstrap

113

then on the one hand 1  2  E Zt−1 T    1 2 e − εt−1 E K = T gh h 1 max K 2 (u) ≤ T gh u∈R =o(1), thus Remark B.1 yields 1 T

T

T

1

∑ Zt−1 = T ∑ E

t=1

   Zt−1 Ft−2 + o p (1).



t=1

On the other hand, we obtain 1 T

    E Z ∑ t−1 Ft−2 T

t=1

  e − εt−1  1 Ft−1 ∑ E √gh K  h t=1   " 1 T 1 e−u = ∑√ K p(u)du T t=1 gh h  " h 1 T K(v)p(e − hv)dv =− ∑ T t=1 g  "   1 T 1 h =− ∑ K(v) p(e) − p (e)hv + p ( e)h2 v2 dv T t=1 g 2    " " " 1 h5  h h3  p (e) vK(v)dv − p ( p(e) K(v)dv + e) v2 K(v)dv =− g g 2 g          

1 = T

T

=o(1)



=1

=0

=o p (1), and, therefore, we obtain 1 T

T

∑ Zt−1 = o p (1).

t=1

=O(1)

114

5 Semiparametric AR(p)-ARCH(1) Models

Analogously, we can prove 1 T

T

1 ∑ √h K t=1



e − εt−1 h

    xt−1   = o p (1),

which shows that the second summand is o p (1). The last summand equals $

  h T  2 σg2 (ε Kh (e − ε t−1 ) − Kh (e − εt−1 ) t−1 ) − σg (e) ∑ T t=1 $    h T  2 Kh (e − ε σg2 (ε = t−1 ) − Kh (e − εt−1 ) t−1 ) − σ (ε t−1 ) ∑ T t=1     2 − σg2 (e) − σ 2 (e) + σ 2 (ε σ ) − (e) t−1 $  h T   ≤ t−1 ) − Kh (e − εt−1 ) ∑ Kh (e − ε T t=1    Lemma 5.5

=

o p (1)

       T gσg2 (u) − σ 2 (u) + max σ 2 (u) max u∈R u∈R       =O p (1)

=O p (1)

=o p (1).  Theorem 5.3 Suppose that Xt is generated by the model (5.1) satisfying assumptions (A1)(A2) and (B1)-(B5). Then the residual bootstrap of section 5.2 is weakly consistent, i.e.  √ ' d T h σh∗2 (e) − σg2 (e) − →N



b(e) τ 2 (e) , p(e) p2 (e)

 in probability.

Proof. Lemma 5.5, 5.8 and 5.9 together with the Slutsky theorem yield the desired result. 

5.3 Wild Bootstrap

115

5.3 Wild Bootstrap Analogously to the previous section, we propose a wild bootstrap technique and prove its weak consistency. A wild bootstrap method can be applied to the model (5.1) as follows: Obtain OLS estimator  a and calculate the residuals p

εt = Xt − a0 − ∑ ai Xt−i ,

t = 0, ..., T.

i=1

Obtain the NW estimator with a bandwidth g

σg2 (e) =

2 T −1 ∑Ti=1 Kg (e − ε i−1 )εi . T −1 ∑Tj=1 Kg (e − ε j−1 )

Generate the bootstrap process εt†2 by computing √ † 2 εt†2 = σg2 (ε t−1 ) + κ − 1σg (ε t−1 )wt , iid

wt† ∼ N (0, 1),

t = 1, ..., T.

Build the bootstrap model

√ † †2 εt†2 = σ †2 (ε t−1 ) + κ − 1σ (ε t−1 )wt

' and calculate the NW estimator σh†2 (e) with another bandwidth h †2 T −1 ∑Ti=1 Kh (e − ε i−1 )εi ' σh†2 (e) = . T −1 ∑Tj=1 Kh (e − ε j−1 )

Analogously to the previous section, we observe the decomposition  √ ' T h σh†2 (e) − σg2 (e) √ † h T 2   T ∑i=1 Kh (e − εi−1 ) κ − 1σg (εi−1 )wi = 1 T  T ∑ j=1 Kh (e − ε j−1 )   h T 2  2  T ∑i=1 Kh (e − εi−1 ) σg (εi−1 ) − σg (e) + 1 T  T ∑ j=1 Kh (e − ε j−1 ) and analyse asymptotic properties of each part separately.

116

5 Semiparametric AR(p)-ARCH(1) Models

Lemma 5.10 Suppose that Xt is generated by the model (5.1) satisfying assumptions (A1)(A2) and (B1)-(B4). Then the wild bootstrap of section 5.3 has the following asymptotic property for all e ∈ R: $   √ h T † d 2 Kh (e − ε → N 0, τ 2 (e) in probability. t−1 ) κ − 1σg (ε t−1 )wt − ∑ T t=1

Proof. The proof is the mirror image of Lemma 5.8. We observe $ √ h  † † Kh e − ε φt (e) := κ − 1σg2 (ε t = 1, ..., T. t−1 t−1 )wt , T For each T the sequence φ1† , ..., φT† is independent because w†1 , ..., w†T is independent. Here we obtain   E† φt† (e) = 0 and   T s2T : = ∑ E† φt†2 (e) t=1

=

h T

p

2   κ − 1 σg2 (ε E† (wt†2 ) t−1 )   

T

t−1 ) ∑ Kh2 (e − ε

t=1

=1

→ τ (e). 2

Then we have the Lyapounov condition for δ = 1 T

 

1

3 

∑ s3 E† φt† (e)

t=1 T



=

 3 $  E† wt†  h3 1 3 T T sT

3/2

κ −1

3    2 t−1 )σg (ε t−1 ) ∑ Kh (e − ε T



t=1

=o p (1). The central limit theorem for triangular arrays applied to the sequence φ1† , ..., φT† shows   T d ∑ φt† (e) → N 0, τ 2 (e) in probability. t=1



5.4 Simulations

117

Theorem 5.4 Suppose that Xt is generated by the model (5.1) satisfying assumptions (A1)(A2) and (B1)-(B4). Then the wild bootstrap of section 5.3 is weakly consistent, i.e.    √ ' b(e) τ 2 (e) d , 2 T h σh†2 (e) − σg2 (e) − →N in probability. p(e) p (e)

Proof. Lemma 5.5, 5.9 and 5.10 together with the Slutsky theorem yield the desired result. 

5.4 Simulations In this section we demonstrate the large sample properties of the residual bootstrap of section 5.2 and the wild bootstrap of section 5.3. Let us consider the following semiparametric AR(1)-ARCH(1) model: Xt = a0 + a1 Xt−1 + εt ,

εt = σ (εt−1 )ηt ,

t = 1, ..., 10, 000,

a0 where X0 = E (Xt ) = 1−a . 1 The simulation procedure is as follows. Note that in (step 5) it would be better to choose some data sets and adopt the bootstrap techniques based on each data set 2,000 times respectively. It is, however, not easy to show the result in a figure. Therefore, we present here one characteristic result for each bootstrap method. iid

Simulate the bias ηt ∼ N (0, 1) for t = 1, ..., 10, 000. 2 , ε 2 = E (ε 2 ) = b0 , a = 0.141, a = Let σ (εt−1 ) = b0 + b1 εt−1 0 1 t 0 1−b1 0.433, b0 = 0.607, b1 = 0.135 and calculate 1/2  2 Xt = a0 + a1 Xt−1 + b0 + b1 εt−1 ηt for t = 1, ..., 10, 000. From the data set {X1 , X2 , ..., X10,000 } obtain the OLS estimators a0 , a1 and then the NW estimators σg2 (e), σh2 (e), where e = 0 with g = T −1/15 and h = T −1/5 with T = 10, 000. Repeat (step 1)-(step 3) 2,000 times.

118

5 Semiparametric AR(p)-ARCH(1) Models 



Choose one data set {X1 , X2 , ..., X10,000 } and its estimators σg2 (e), σh2 (e) . Adopt the residual bootstrap of section 5.2 and the wild bootstrap of section 5.3 based on the chosen data set and estimators. Both bootstrap methods are repeated 2,000 times. √  Compare the simulated density of the NW estimators ( T h σh2 (e) −  σ 2 (e) , thin lines) with the residual and the wild bootstrap approximations   √ ' √ ' †2 ∗2 (e) − σ 2 (e) and T h σ 2 (e) , bold lines). (e) − σ ( Th σ g g h h Figure 5.1 and Figure 5.2 indicate that the residual and the wild bootstrap are weakly consistent, which corresponds to the theoretical results of Theorem 5.3 and 5.4.

119

0.0

0.1

0.2

0.3

0.4

5.4 Simulations

−3

−2

−1

0

1

2

3

0.0

0.1

0.2

0.3

0.4

Figure 5.1: Residual Bootstrap

−3

−2

−1

0

1

2

Figure 5.2: Wild Bootstrap

3

Appendix

K. Shimizu, Bootstrapping Stationary ARMA-GARCH Models, DOI 10.1007/978-3-8348-9778-7, © Vieweg+Teubner | GWV Fachverlage GmbH, Wiesbaden 2010

A Central Limit Theorems Definition A.1 (Martingale Difference Sequence) Let {Zk , k ∈ N} be a sequence of integrable random variables defined on a ( common probability space (Ω, A , P) and let {Fk , k ∈ N {0}} denote an increasing sequence of sub-σ -fields of A such that Zk is Fk -measurable for each k ∈ N. Then {Zk } is called a martingale difference sequence if E (Zk |Fk−1 ) = 0 almost surely for every k ∈ N. Theorem A.1 Let {Zk } be a martingale difference sequence with 0 < s2K :=

K

∑ E (Zk2 ) < ∞.

k=1

Then s−1 K

K

∑ Zk → N (0, 1) d

k=1

if (a) s−2 K (b) s−2 K

K

p

∑ E (Zk2 |Fk−1 ) → 1

k=1 K

∑E

and



 p Zk2 1{|Zk | ≥ ε sK } → 0 f or every ε > 0.

k=1

Proof. Brown (1971), Theorem 1.  Here we observe a triangular array {Zn,k , 1 ≤ k ≤ kn , n ∈ N} of integrable random variables defined on a common probability space (Ω, A , P). Let {Fn,k , 1 ≤ k ≤ kn , n ∈ N} be a given array of sub-σ -fields of A such that Zn,k is Fn,k -measurable and Fn,k is monotone increasing in k for every n. Theorem A.2 Suppose that for each n the sequence Zn,1 , ..., Zn,kn is independent, E (Zn,k ) = 2 ) < ∞. If the Lindeberg condition n E (Zn,k 0 and 0 < s2n := ∑kk=1 kn

1

E n→∞ ∑ s2 lim

k=1 n



  2 Zn,k 1 |Zn,k | ≥ ε sn =0

124

A Central Limit Theorems holds for all positive ε , then s−1 n

kn

−−→ N (0, 1). ∑ Zn,k −n→∞ d

k=1



Proof. Billingsley (1995), Theorem 27.2. Corollary A.1 If, under the assumptions of Theorem A.2, the Lyapounov condition kn

1

E n→∞ ∑ s2+δ lim



 |Zn,k |2+δ = 0

k=1 n

holds for some positive δ , then the Lindeberg condition holds for all positive ε. Proof. Billingsley (1995), Theorem 27.3.



B Miscellanea Proposition B.1 Given g : Rk → Rl and any sequence {bn } of k × 1 random vectors such that p

p

bn → b, where b is a k × 1 vector. If g is continuous at b, then g(bn ) → g(b). 

Proof. White (2001), Proposition 2.27. Remark B.1 For every sequence {Zt , t = 1, ..., T } of random variables with E (Zt2 ) = o(1) T we obtain ) E

1 T

T



∑

t=1

1 T = 2 ∑ E T s, t=1 =

    Zt − E Zt Ft−1

)

*2

    Zt − E Zt Ft−1



    Zs − E Zs Fs−1

  2 1 T   E Zt − E Zt Ft−1 ∑ 2 T t=1    ≤E (Zt2 )

=o(1). Chebychev’s inequality yields for every δ > 0    1 T  1 T     P  ∑ Zt − ∑ E Zt Ft−1  > δ  T t=1  T t=1 *2  )    1 1 T  Zt − E Zt Ft−1 ≤ 2E ∑ T t=1 δ =o(1), that is, according to the definition 1 T

T

p

1

T

∑ Zt → T ∑ E

t=1

t=1

    Zt Ft−1 .

*

126

B Miscellanea

Theorem B.1 Let {Zt , Ft } be a martingale difference sequence. If ∞

E (|Zt |2+p ) <∞ t 1+p t=1

∑

for some p ≥ 1, then 1 T Proof. Stout (1974, p. 154-155).

T

∑ Zt −−→ 0. a.s.

t=1



Bibliography [1] Basel Committee on Banking Supervision (1996) Amendment to the Capital Accord to incorporate market risks. Bank for International Settlements, Basel. [2] Berkes, I., Horváth, L. and Kokoszka, P. (2003) GARCH processes: structure and estimation. Bernoulli 9, 201-227. [3] Billingsley, P. (1995) Probability and Measure, 3rd Edition. Wiley, New York. [4] Bollerslev, T. (1986) Generalized autoregressive conditional heteroskedasticity. Journal of Econometrics 31, 307-327. [5] Bollerslev, T., Chou, R. Y. and Kroner, K. F. (1992) ARCH modeling in finance. Journal of Econometrics 52, 5-59. [6] Brown, B. M. (1971) Martingale Central Limit Theorems. The Annals of Mathematical Statistics 42, 59-66. [7] Chang, B. R. (2006) Applying nonlinear generalized autoregressive conditional heteroscedasticity to compensate ANFIS outputs tuned by adaptive support vector regression. Fuzzy Sets and Systems 157, 1832-1850. [8] Comte, F. and Lieberman, O. (2003) Asymptotic theory for multivariate GARCH processes. Journal of Multivariate Analysis 84, 61-84. [9] Dowd, K. (2005) Measuring market risk, 2nd Edition. Wiley, Chichester. [10] Efron, B. (1979) Bootstrap methods: another look at the jackknife. The Annals of Statistics 7, 1-26. [11] Efron, B. and Tibshirani, R. J. (1993) An Introduction to the Bootstrap. Chapman & Hall, New York. [12] Engle, R. F. (1982) Autoregressive conditional heteroscedasticity with estimates of the variance of United Kingdom inflation. Econometrica 50, 987-1007. [13] Engle, R. F. and González-Rivera, G. (1991) Semiparametric ARCH Models. Journal of Business & Economic Statistics 9, 345-359. [14] Fama, E. F. (1965) The Behavior of Stock-Market Prices. The Journal of Business 38, 34-105. [15] Fan, J. and Yao, Q. (1998) Efficient estimation of conditional variance functions in stochastic regression. Biometrika 85, 645-660. [16] Francq, C. and Zakoïan, J.-M. (2004) Maximum likelihood estimation of pure GARCH and ARMA-GARCH processes. Bernoulli 10, 605-637.

K. Shimizu, Bootstrapping Stationary ARMA-GARCH Models, DOI 10.1007/978-3-8348-9778-7, © Vieweg+Teubner | GWV Fachverlage GmbH, Wiesbaden 2010

128

Bibliography

[17] Franke, J., Kreiß, J.-P. and Mammen, E. (1997) Bootstrap of kernel smoothing in nonlinear time series. Working Paper, Technische Universität Braunschweig. [18] Franke, J., Kreiß, J.-P. and Mammen, E. (2002) Bootstrap of kernel smoothing in nonlinear time series. Bernoulli 8, 1-37. [19] Fuller, W. A. (1996) Introduction to Statistical Time Series, 2nd Edition. Wiley, New York. [20] Hartz, C., Mittnik, S. and Paolella, M. (2006) Accurate value-at-risk forecasting based on the normal-GARCH model. Computational Statistics & Data Analysis 51, 22952312. [21] Horowitz, J. L. (2003) The Bootstrap in Econometrics. Statistical Science 18, 211218. [22] Jorion, P. (2007) Value at Risk: The New Benchmark for Managing Financial Risk, 3rd Edition. McGraw-Hill, New York. [23] Kreiß, J.-P. (1997) Asymptotical Properties of Residual Bootstrap for Autoregressions. Working Paper, Technische Universität Braunschweig. [24] Kreiß, J.-P. and Franke, J. (1992) Bootstrapping stationary autoregressive movingaverage models. Journal of Time Series Analysis 13, 297-317. [25] Lee, S.-W. and Hansen, B. E. (1994) Asymptotic theory for the GARCH(1,1) quasimaximum likelihood estimator. Econometric Theory 10, 29-52. [26] Ling, S. (2007) Self-weighted and local quasi-maximum likelihood estimators for ARMA-GARCH/IGARCH models. Journal of Econometrics 140, 849-873. [27] Ling, S. and Li, W. K. (1997) On Fractionally Integrated Autoregressive MovingAverage Time Series Models With Conditional Heteroscedasticity. Journal of the American Statistical Association 92, 1184-1194. [28] Ling, S. and McAleer, M. (2003a) Asymptotic theory for a vector ARMA-GARCH model. Econometric Theory 19, 280-310. [29] Ling, S. and McAleer, M. (2003b) On Adaptive Estimation in Nonstationary ARMA Models with GARCH Errors. The Annals of Statistics 31, 642-674. [30] Liu, R. Y. (1988) Bootstrap procedures under some non-i.i.d. models. The Annals of Statistics 16, 1696-1708. [31] Lumsdaine, R. L. (1996) Consistency and Asymptotic Normality of the QuasiMaximum Likelihood Estimator in IGARCH(1, 1) and Covariance Stationary GARCH(1, 1) Models. Econometrica 64, 575-596. [32] Maercker, G. (1997) Statistical Inference in Conditional Heteroskedastic Autoregressive Models. PhD Dissertation, Technische Universität Braunschweig. [33] Mammen, E. (1992) When Does Bootstrap Work?: Asymptotic Results and Simulations, Lecture Notes in Statistics 77. Springer, New York.

Bibliography

129

[34] Mandelbrot, B. (1963) The Variation of Certain Speculative Prices. The Journal of Business 36, 394-419. [35] Nicholls, D. F. and Pagan, A. R. (1983) Heteroscedasticity in Models with Lagged Dependent Variables. Econometrica 51, 1233-1242. [36] Palkowski, F. (2005) Nichtparametrisches Bootstrap in heteroskedastischen Zeitreihen. PhD Dissertation, Technische Universität Braunschweig. [37] Pantula, S. G. (1988) Estimation of autoregressive models with ARCH errors. Sankhy¯a: The Indian Journal of Statistics B 50, 119-138. [38] Pascual, L., Romo, J. and Ruiz, E. (2005) Bootstrap prediction for returns and volatilities in GARCH models. Computational Statistics & Data Analysis 50, 2293-2312. [39] Reeves, J. J. (2005) Bootstrap prediction intervals for ARCH models. International Journal of Forecasting 21, 237-248. [40] Robio, P. O. (1999) Forecast intervals in ARCH models: bootstrap versus parametric methods. Applied Economics Letters 6, 323-327. [41] Shao, J. and Tu, D. (1996) The jackknife and bootstrap. Springer, Berlin. [42] Stout, W. F. (1974) Almost Sure Convergence. Academic Press, New York. [43] Straumann, D. (2005) Estimation in Conditionally Heteroscedastic Time Series Models, Lecture Notes in Statistics 181. Springer, Berlin. [44] Tapia, R. A. and Thompson, J. R. (1978) Nonparametric probability density estimation. Johns Hopkins University Press, Baltimore. [45] Thombs, L. A. and Schucany, W. R. (1990) Bootstrap Prediction Intervals for Autoregression. Journal of the American Statistical Association 85, 486-492. [46] Weiss, A. A. (1984) ARMA models with ARCH errors. Journal of Time Series Analysis 5, 129-143. [47] Weiss, A. A. (1986) Asymptotic theory for ARCH models: Estimation and testing. Econometric Theory 2, 107-131. [48] White, H. (2001) Asymptotic Theory for Econometricians, Revised Edition. Academic Press, San Diego.

Index AR(1)-ARCH(1) model, 10, 60 residual bootstrap, 10–12 wild bootstrap, 12–13 AR(1)-NARCH(1) model, 117 AR(p)-ARCH(q) model, 20 OLS estimation, 22–36 residual bootstrap, 37–52 two-step wild bootstrap, 57–59 wild bootstrap, 52–59 AR(p)-NARCH(1) model, 86 NW estimation, 87–98 residual bootstrap, 98–114 wild bootstrap, 115–117 ARCH(p) model, 2 ARMA(1, 1)-GARCH(1, 1) model, 3, 15, 80 ARMA(p, q)-GARCH(r, s) model, 14, 65 QML estimation, 67–68 residual bootstrap, 68–75 wild bootstrap, 76–80 bootstrap basic idea, 5–7

estimation of heteroscedasticity, 10– 13 false application, 13–16 GARCH(p, q) model, 2 geometric returns, 2 Newton-Raphson method, 70 residual bootstrap, 10–12, 37–52, 68–75, 98–114 simulation AR(1)-ARCH(1) model, 60–61 AR(1)-NARCH(1) model, 117–119 ARMA(1, 1)-GARCH(1, 1) model, 3–4, 15–16, 80–81 VaR forecast model, 15–16 VaR forecast model, 13–16 wild bootstrap, 12–13, 52–59, 76–80, 115– 117

K. Shimizu, Bootstrapping Stationary ARMA-GARCH Models, DOI 10.1007/978-3-8348-9778-7, © Vieweg+Teubner | GWV Fachverlage GmbH, Wiesbaden 2010