This action might not be possible to undo. Are you sure you want to continue?

BooksAudiobooksComicsSheet Music### Categories

### Categories

### Categories

Editors' Picks Books

Hand-picked favorites from

our editors

our editors

Editors' Picks Audiobooks

Hand-picked favorites from

our editors

our editors

Editors' Picks Comics

Hand-picked favorites from

our editors

our editors

Editors' Picks Sheet Music

Hand-picked favorites from

our editors

our editors

Top Books

What's trending, bestsellers,

award-winners & more

award-winners & more

Top Audiobooks

What's trending, bestsellers,

award-winners & more

award-winners & more

Top Comics

What's trending, bestsellers,

award-winners & more

award-winners & more

Top Sheet Music

What's trending, bestsellers,

award-winners & more

award-winners & more

Welcome to Scribd! Start your free trial and access books, documents and more.Find out more

**Volatility Impulse Response Functions for Multivariate GARCH Models
**

Christian M. Hafner and Helmut Herwartz

1 2

August 1998

Abstract

In the empirical analysis of nancial time series, multivariate GARCH models have been used in various forms. In most cases it is not well understood how the use of a restricted model has to be paid with loss of valuable information. We investigate the structural implications of the alternative models for the response of the conditional co variances to independent shocks. The impulse response analysis, adopted to volatility models, appears to be a convenient methodology to obtain information on the interaction of nancial series. We de ne volatility impulse response functions and provide an empirical analysis for a bivariate exchange rate series. For the analyzed series, the impulse response function of the correlation reveals strong discrepancies between the estimated diagonal and BEKK models. This indicates that the diagonality restriction may hide important structural properties of the series. Keywords: Multivariate GARCH, impulse response functions, exchange rate volatility JEL Classi cation: C32, G11

Institut fur Statistik und Okonometrie, Wirtschaftswissenschaftliche Fakultat, Humboldt Universitat zu Berlin, Spandauer Str.1, D-10178 Berlin, Germany, and CORE, Universit
Catholique de Louvain, e Louvain-la-Neuve, Belgium. e-mail hafner@wiwi.hu-berlin.de. 2 Institut fur Statistik und Okonometrie, Wirtschaftswissenschaftliche Fakultat, Humboldt Universitat zu Berlin, Spandauer Str.1, D-10178 Berlin, Germany. e-mail helmut@wiwi.hu-berlin.de. Financial support by the Deutsche Forschungsgemeinschaft is gratefully acknowledged. Both authors would like to thank Luc Bauwens and Helmut Lutkepohl for helpful comments. This text presents research results of the Belgian Program on Interuniversity Poles of Attraction initiated by the Belgian State, Prime Minister's O ce, Science Policy Programming. The scienti c responsability is assumed by the authors.

1

Instead of de ning news to occur in the dependent " . However. For a dynamic interpretation of the estimation results for multivariate GARCH models. introduced by Sims 1980 for the analysis of VAR models. Koop. will provide information on the impact of independent shocks on volatility. impulse responses depend on initial conditions. which the researcher has to assume a priori. the conditional expectation of volatility can be derived analytically for prespeci ed shocks. Volatility impulse response functions VIRF. This general framework may give valuable information on the structural dependencies if the model is highly complex and nonlinear. we do not have orthogonality problems. we assume a shock in the independent innovation and regard the expectation of volatility conditional on this shock. the components of " in general are contemporaneously correlated. since there are similarities between GARCH and VAR type models. Apart from being dependent on the past.1 Introduction Multivariate GARCH models have been designed to model the conditional covariance matrix of multiple time series. Adopted to parametric GARCH models. This may still work for macroeconomic systems. Pesaran and Potter 1996 de ne a generalized impulse response function using the mean of the response vector conditional on the past and a present shock. The knowledge of this volatility matrix may give valuable information on risk measures associated with a given set of nancial assets. to the analysis of shocks in volatility. one can generalize the impulse response analysis. Pesaran and Potter 1996 but rather provide a comparison of VIRF for example with the conditional moment pro les proposed by Gallant. In their framework. all of which can be regarded as random variables. but it will fail for highly interdependent systems of nancial time series. Rossi and Tauchen 1993. intermediate innovations and the model parameters. The de nition of shock or news by GRT and Engle and Ng 1993 is a perturbation in the conditionally heteroskedastic error term " . when considering orthogonalized impulse responses the order of variables in the system determines the causality scheme. no uni ed methodology is available. However. Since is composed of contemporaneously uncorrelated random variables. Thus. GRT hereafter. Some recent approaches generalize impulse response functions to nonlinear models and higher conditional moments. as de ned in our paper. we do not proceed along the lines of Koop. It has been frequently observed both in univariate and multivariate frameworks that volatility changes over time. The aim of our paper is to introduce a new de nition of volatility impulse response functions. Monte Carlo techniques are required for the computation of impulse responses. We propose a means to calculate the VIRF for the t t t t t 2 . an initial shock. in the framework of multivariate GARCH models. These conditional moment pro les are introduced as a means to uncover the dynamic relation of variables for a semi-nonparatric model. volatility pro les can be easily evaluated. As it is well known for VAR models.

q case one may allow to depend on past observations " .most general vec representation of a multivariate GARCH model. : : : . i = 1. = 0 and Var " j . quasi likelihood estimation or Bayesian methods. However. Section 4 presents the empirical study for a bivariate exchange rate series. by generalized method of moments. we present estimates for the diagonal and BEKK model for the same exchange rate series as in Bollerslev and Engle 1993. Let " denote an N dimensional random vector such that t " = 1=2 t t t N t 1 where denotes a multivariate i:i:d: random vector with mean zero and covariance matrix the identity matrix I . As in the univariate GARCHp. Section 3 de nes VIRF for general multivariate GARCH models. q model Bollerslev. = . i = 1. We nd distinct di erences for the VIRF of the correlation between DEM USD and GBP USD for the diagonal and BEKK speci cations of volatility. p.g. This indicates that the more restrictive diagonal speci cation may not capture in total the dynamic structure of the conditional correlation. because we will later de ne news to appear in the i:i:d: innovation . SV models may be more parsimonious and are sometimes found to provide better ts than GARCH models. but with the estimation period extended from 1980 to 1994. It should be mentioned that time-varying volatility of multivariate nancial time series may also be modelled by stochastic volatility SV. In an empirical study. q and on past covariance matrices . With vech: denoting the operator that stacks the lower fraction of an N N matrix into a t t 1 t t 1 t t 1 t t t i t i 3 . and Section 5 concludes. 1 may be called a strong GARCH model. Section 2 reviews representation. However. identi cation and estimation issues of multivariate GARCH models. 1986 to the multivariate case is straightforward. : : : . denoting the information set available at time t .g. 2 Multivariate GARCH Models Speci cation. One may conjecture that the computational advantage of GARCH type models is even more relevant in a general multivariate setting. . 1 implies that E " j . the advantage of GARCH type models is the straightforward estimation by quasi maximum likelihood. With . 1. According to the terminology of Drost and Nijman 1993. . Harvey. Ruiz and Shephard 1994. The paper is organized as follows. in the multivariate framework possible dependencies easily become intractable for empirical work. Estimation can be performed e. and Estimation From a theoretical point of view the generalization of the univariate GARCHp. We employ this strong GARCH form. see e. Identi cation. We show that VIRF may give a better understanding of the highly complex dynamic behavior of nancial time series.

G 11 11 11 G . Note that even in the case N = 2 and p = q = 1. potentially important dynamics between " and its history are excluded. Second. 21 parameters characterize the dynamic relationship between " and its history. however. Bollerslev. In this case the assumption that the upper left elements of A and G are greater than zero is su cient for the model parameters to be identi ed. 1988 in order to maximize the log likelihood function. A diagonal parametrization of the matrices A and G yields a restricted version of the diagonal model. For the present analysis we take K = 1 and p = q = 1.g. A t t N 11 A . The vector c accounts for time invariant variance components and contains N N + 1=2 coe cients. A and G . K p ki t i k =1 i=1 ki ki 3 In 3. in Judge et al. and . even in the case of K = 1 the BEKK representation allows for direct dependence of the conditional variance of one variable on the history of other variables within the system. : q p t i t i t i i t i i=1 i=1 i i 2 In the so called vec representation 2. I . In order to obtain tractable models for empirical work one usually has to impose parameter restrictions on the vec model 2. For the empirical part of this study we used the BHHH algorithm as described e. The so called BEKK representation in 3 copes with two issues evolving in multivariate GARCH modelling.N N + 1=2 dimensional vector one may advocate the multivariate GARCHp. such a framework clearly reduces the number of model parameters. First. . Engle and Wooldridge 1988 propose the so called diagonal model in which the matrices A and B are assumed to be diagonal such that the i. For the vec model in 2 positive de niteness of is often ensured by appropriate positivity restrictions. On the other hand. one always obtains a positive de nite matrix . irrespective of the parameters in C . + B vech . C is an upper triangular matrix and A and G are N N parameter matrices. the 11 t N 4 . "0 . K q ki t i t i k =1 i=1 ki ki X X G0 G + . " is covariance stationary if and only if the eigenvalues of the matrix 0 ik ik t t ki ki 11 11 t A 11 A +G 11 11 G 11 4 have moduli less than one. vecC 0 C : 11 1 0 0 5 The estimation of the parameters in A and G is analogous to the univariate case. For N 0. The implied unconditional covariance matrix is obtained as E vec" "0 = I 2 . j element in depends linearly on the respective elements of the matrices " . A and B denote parameter matrices each containing N N + 1=2 parameters. On the one hand. Engle and Kroner 1995 discuss the speci cation given in 3 in detail. q model: X X vech = c + A vech" . . Engle and Kroner 1995 discuss a dynamic speci cation of the following form: 2 t i i t t i t i t i t 0 t = C0 C 0 0 X X A0 " "0 A + . "0 .

"0 .g.ij t. i. 1995. let us restrict to the GARCH1. VIRF is now de ned as the expectation of volatility conditional on an initial shock and initial volatility. 6 is the quasi likelihood function. as indicated by the size. A . one may arbitrarily assume a shock in one or more components and set all other components to their expectation zero. + G vech . The conditional covariance matrix is a function of the innovations . V is an N N + 1=2 dimensional vector. if N = 2.e. and the second element of V is t 0 t 0 0 t 0 t 0 the impulse response function of the conditional covariance. For example. Because for every BEKK model there exists a unique equivalent vec speci cation see Engle and Kroner. vech = c + A vech" .ij t 0 0 0 t. V V th th t. and . Thanks to the independence of the components of .contribution of an observation " to the joint log likelihood of a sample with T observations log L = P l is obtained as: l = . N ln2 .j j 0 5 . cf. VIRF for BEKK models can be obtained by rst transforming the model to a vec model. : : : . which provides consistent estimates regardless of the underlying distribution. t T t=1 t t t 1 t t t t 3 Volatility Impulse Response Functions Let us consider model 1 with dimension N 1. = : 8 In 8. G . as indicated by the sign of the individual components. The volatility state at time t = 0 is assumed to be the steady state = . we will in the following only investigate VIRF for the vec model. and estimation is performed by quasi maximum likelihood QML. c: 1 1 1 At time t = 0 some independent news are re ected by and it is not speci ed whether the news are `good' or `bad'. t 1 t 1 t 1 1 t 1 7 with unconditional covariance matrix . the initial shock . In order to simplify the presentation. e. respectively.ii 0 t. Bollerslev and Wooldridge 1992. 1 ln j j . An impulse response function of the conditional correlation between the i and j variable is naturally de ned to be 9 C = q V . " : 6 2 2 2 If the multivariate distribution of is other than normal. . . and how important the news are. the rst and third element of V represent the impulse response function of the conditional variances of the rst and second variable. 1 "0 . 0 t 0 t 1 t 1 0 0 V = E vech j . This can be changed to any other state of interest. and vech = I . large or small volatility states.1 case.

1 0 1 1=2 0 0 1=2 1 10 and. A vech 0 + A + G . Starting at t = 1. t t 0 1 1 1 2 t 0 2 t 2 t i 2 t 0 1 @ t 1 2 0 2 0 to be the `impact of an initial shock on the conditional variance'. E vech j " . G c: The idea to consider shocks in the independent errors also appears in Bauwens and Lubrano 1997 who deal with the univariate case and the response of the conditional variance to an initial shock . . For the case of a threshold GARCH model with a threshold at zero. : : : . one for a positive. . and assuming that the process starts at time t = 0 see Nelson. the expectation is taken with respect to the intermediate innovations . 1990. .!1 V = vech = I . we compare VIRF with the methodology of GRT in a bivariate system. so that our VIRF approaches a constant the unconditional covariance matrix for multivariate GARCH rather than the zero matrix: . Next. In general.A . : : : . possibly compared with a baseline.q models can be expressed as a function of past squared innovations . . Bauwens and Lubrano propose two impulse response functions. An explicit expression for the VIRF is t V = c + A + G V . . When extensions of standard GARCH models are considered. . Bauwens and Lubrano de ne 13 E @ . c+ + A + G . one for a negative shock. it can be seen that t. GRT assume a shock in " and compare the volatility pro le relative to a baseline " that will be set to its expectation zero. in general provides a better understanding of volatility dynamics. lim . . t t 0 1 t 0 0 1 14 6 . A . Thus.G . A + G . 13 will depend on in a nontrivial way and thus we think that regarding the expectation E j . 11 This recursive formula can e ciently be used in algorithms.ij 0 th th t V = c + A vech 0 + G vech. G vech: t 0 N N +1=2 1 1 t 1 N N +1=2 1 1 t 1 1 1 t 1 1 1=2 0 0 1=2 1 1 1 1 12 Unlike GRT and Koop. threshold or exponential GARCH models. we write 1 t 1 0 0 2 t 0 2 0 t 0 t GRT = E vech j ". Pesaran and Potter 1996 we do not incorporate a baseline function into the de nition of impulse responses. since the innovations enter the volatility equation in a form other than quadratic. . For GARCH models this function does not depend on the shock . directly.g. : 0 1 1 t 1 0 V = I . As in univariate GARCHp. I . e. In 13.where V corresponds to the i row and j column of . 13 may not measure the impact of a shock on volatility correctly. for t 2.

One may imagine that this procedure becomes intractable for higher dimensions. = 0.1 0 0 2 11 11 12 12 p p + 11 12 1 2 13 12 p : 2 1 17 The di erence between 17 and 15 is obvious: In the case of zero covariance equivalent to p = 0. medium. however. we de ne a baseline innovation = 0 .e. We have V . and small shocks in both variables. is a very peculiar case. @ dg 0 0 0 0 0 1 2 1 2 t t 0 t where . In the bivariate setting. R = @ vechE ""0j . 00 and = 0 .where " = 0 . i. 0 1 15 where is the i. consider the case of a shock in the rst variable.0 2 2. This. 10. If unit shocks. GRT consider di erent scenarios corresponding to alternative combinations of large. " and R is an N N + 1=2 N matrix. the question is how to choose a realistic contemporaneous shock in the other variable. j element of the matrix . Then we obtain for the conditional variance of the rst equation 1 0 0 1 0 1 0 0 1 1=2 0 0 1=2 ij 1=2 V . When xing a shock in one variable or . since usually we deal with nancial time series that are contemporaneously closely related. j element of A . 0 . 00 and " = " + with = . = 1 .1 ij 2 11 1 . After one time period. 6= 0. 2 12 4 Empirical Analysis 4. the response of the conditional variance of the rst variable is dg"0"0 = "2 0 1. and no shock in the second variable. 00. are considered. V = p + 11. we obtain the impulse response matrix R de ned by Lin 1997. To provide a simple comparison of the implications of the di erent de nitions. For an exposition of VIRF consistent with the GRT methodology. V = A vech 0 : 16 Let p denote the i.1 Modelling a bivariate exchange rate series As in Bollerslev and Engle 1993 we analyze daily data for the Deutsche Mark DEM and British pound GBP exchange rates vis-. 17 and 15 are equal up to the scaling factor p 1 .1 0 11. If p 6= 0 the functions in 17 and 15 may exhibit quite di erent shapes. 00 and a schock = .0 0 0 0 t 1 2 GRT = 11.

To further characterize the linear dynamics of this bivariate system the second line of Figure 1 provides a graph of 7 .-vis the US Dollar USD to provide an a empirical illustration of VIRF introduced above. 1994 which amounts to 3720 available observations. The estimation period is December 31. 1979 to April 1. Plots of the respective log exchange rates are given in the rst panel of Figure 1.

Note that the improvement of 2 log L by about 20 points is signi cant with respect to the 4 distribution at any reasonable level. one should be careful when interpreting the signi cance of the computed statistics. 1987. which are both not signi cant at the 5 level. Enforcing the matrices A and G to be diagonal gives a maximum of the likelihood function of 28554. As initial values for the estimation of BEKK models. The asymptotic distribution of these test statistics is developed under the assumption of i:i:d: innovations. An LM Test on ARCH e ects Engle.the di erence between the two log rates. Prior to the analysis of conditional heteroskedasticity for the bivariate system we performed an unrestricted VAR analysis for the rst di erences of the logarithmic rates in order to obtain a residual series which is free of serial correlation. That is.30 and 23. which for arbitrage reasons resembles the log rate DEM GBP. since the largest eigenvalues are smaller than but close to one. O diagonal elements of A and G were initialized with zero. diagonal elements of the matrices A and G were taken to be the square root of the respective univariate estimates.56. Furthermore. 1991 of 85. They indicate that both parametrizations imply covariance stationarity and high persistence.82 which is signi cant at the 5 level but not at the 1 level. respectively. Taking into account the apparent conditional heteroskedasticity. the ADF statistic given for the di erence between the two rates shows that both rates presumably are not cointegrated since their di erence is nonstationary see e. Two of four estimated o -diagonal elements are found to be signi cant at the 5 level. Lutkepohl. we used the results of preliminary univariate GARCH estimates.62. 1982 yields strong evidence for conditional heteroskedasticity in both univariate residual series.g. Augmented Dickey Fuller ADF test regressions clearly support the presence of a unit root factor in the data generating process of the two series. The estimated o diagonal elements are negative for the A matrix and positive for the G matrix. Estimated o diagonal elements of the A matrix are about two times larger than those of the G matrix. introducing four additional parameters increases the log likelihood by 10 points. The estimated residual series are depicted in the lower panels of Figure 1. 11 11 11 11 0 11 11 2 11 11 11 11 8 . Engle and Granger. The bivariate residual series obtained from the VAR5 model is characterized by a multivariate portmanteau statistic with 20 lags see e. We speci ed a VAR5 model which minimized the multivariate AIC criterion.e. Allowing o diagonal coe cients di erent from zero i. The portmanteau statistics for the univariate series are 21. The estimated parameter matrices are given in Table 1.14.g. Initial conditions for the matrix C were obtained from the unconditional covariance matrix 5 and the sample means of variances and covariances. The initial model parameterized with results from the univariate analyses yields a value of the log likelihood function of 28533. The estimated eigenvalues of the competing GARCH models are also given in Table 1.

48 82. The plots on the left of Figure 2 show the VIRF for DEM USD and a shock in DEM USD .04 108. In order to visualize the functions one can proceed in two ways: 1 reduce the `shock matrix' 0 to a vector by xing one component of .98 85.0 1.0 2.023 0.04 2.4 G 1 log L . are the eigenvalues of the matrix 4 and log L is the value of the log likelihood function.974 .62 .80e-04 4. the VIRF of the diagonal model is larger than the VIRF of the BEKK model for the variances.280 14.542 for DEM USD 0 0 0 0.0 2.276 0. : : : . As is plausible.949 10.0 t 0 1.46e-04 4.37 0.950 9.14 . Not shown are the cross e ects where the size of the impact is much smaller.6 1. shocks in one rate have a large impact on the rate itself.941 0. the plots on the right of Figure 3 display the VIRF for GBP USD and a shock in GBP USD .24 0.0 1.979 28554.535 and 0.70e-04 6.969 . and Figure 3 for the correlation as in 9 and a shock in DEM USD plots on the left and GBP USD plots on the right.2 0.15e-03 4.8 -0. the plots are given in Figures 2 and 3. that is.982 28564.060 -2. In the same manner. 4.969 .26 6.000.982 for the BEKK model. For the rst procedure.28e-03 8. The left axis displays a shock . 50. Recall from Table 1 that the largest eigenvalues of the estimated models are close to one: 0.0 9 .93 6.276 0. It is obvious that shocks in either or persist very long in both GBP USD and DEM USD volatilities and in the correlation between GBP USD and DEM USD. Di erences between the BEKK and diagonal VIRF can be found mainly in the level of the VIRF.268 12. and the right axis the time horizon for t = 1. and 2 x the time t and then plot V for selected t. i = 1.36 0 i 0. say.963 i Table 1: Estimation results for the diagonal and BEKK models t statistics in parentheses.23 5.77 A 1 0. we refer to the VIRF plots given in Figures 2 to 4 for the estimation results of diagonal and BEKK models reported in Table 1.40 BEKK 1.1 t 2.947 115.0 2.Model C DIAG 1.2 VIRF for the estimated diagonal and BEKK models In the following. The reason is that the estimated unconditional variances slightly di er diagonal model: 0.974 .27 1.0 2.48 -0. the constant matrix in the volatility equation was multiplied by 10. to a constant. : : : . In all cases.017 -1. .979 for the diagonal model and 0.40e-04 7. and plot the function V over t and .967 . 4. where is xed to its mean zero.041 0.

Unlike other recent approaches we de ne news to appear in the independent innovation rather than in the conditionally heteroskedastic 10 . Regarding shocks in GBP USD. but the qualitative di erence in the functions is smaller. we also computed the VIRF for the estimation results reported in Bollerslev and Engle 1993 for a vec model and the shorter estimation period until 1985. The same VIRF for BEKK shows an inverted U shape at the rst periods. However. This may motivate the constant conditional correlation model of Bollerslev 1990. respectively. as was visible in the corresponding VIRF plots not shown. the shapes di er: For a shock in DEM USD. The di erent interpretations are obvious: regarding short term e ects. whereas it decreases the correlation for the BEKK model. However.518 and 0. depending on the model speci cation. This again explains the slightly higher overall level of the diagonal VIRF compared with the BEKK VIRF.509. For the second procedure to generate VIRF plots. To compare our results with previous work. Shocks in DEM USD are on the left axis.and GBP USD.03. t 0 1. Their results imply a relatively fast dissipation of shocks in the GBP USD. Qualitatively. as for example in the CAPM. keep in mind that so far we have only looked at an independent shock in one variable.0 0 5 Conclusions and Outlook We introduced volatility impulse response funtions VIRF for multivariate time series exhibiting conditional heteroskedasticity. which are mainly due to the o diagonal element of the matrix C . The estimated unconditional correlations are respectively 0. This may be due to the lack of events such as the 1987 crash in their sample. Considering the correlation or alternatively the covariance as an important risk measure in nancial models. For this kind of plot. keeping the other shock xed at zero.752 for the diagonal and BEKK model. Note that overall the variation in the correlation VIRF appears rather small a range of about 0. so we only show the plots for the BEKK model.0 2. Let us turn to the VIRF in Figure 3 for the conditional correlations as de ned in 9. t = 10 is selected and the plots of V shown in Figure 4. one can arrive at di erent conclusions from the empirical analysis. This may change when in the following we look at two simultaneous independent shocks of arbitrary size. no remarkable qualitative di erence between the diagonal and BEKK model was detected. large shocks in DEM USD increase the correlation with GBP USD for the diagonal model. the plots are quite similar. Note the extremely asymmetric shapes of the variance plots. If there was no interaction between the two variance series. and then turns to a U shape. one can draw similar conclusions.777 and 0. shocks in GBP USD on the right one. BEKK model: 0. we would obtain U shapes along the axes of the cross variable. the diagonal VIRF has a U shape that remains stable over time.

The impacts of independent shocks on conditional co variances of the DEM and GBP rates measured vis-. The advocated methodology is applied to a bivariate exchange rate series. For the general vec representation of multivariate GARCH models we provide the analytic expressions for VIRF.error.

Bollerslev. Review of Economics and Statistics. The VIRF methodology proposed in this paper is easily extended to multivariate threshold GARCH models. R. 1993. References Bauwens. 1993. Bayesian Option Pricing Using Asymmetric GARCH". Using the BEKK model as parametrization device we found deviations from a restricted version of the so|called diagonal model to be signi cant. Engle. Modelling the Coherence in Short-run Nominal Exchange Rates: A Multivariate Generalized ARCH Approach". 307 327. L.. 116 131. A Capital Asset Pricing Model with Time Varying Covariances". T.-vis the US Dollar are traced through time. T. bootstrap methods may be considered. However. CORE DP 9759.F. Generalized Autoregressive Conditional Heteroskedasticity". 61. 96.. 1997. Journal of Econometrics. 1986. 1990. one may investigate distributional properties of VIRF estimates.. Bollerslev. M. T. The structural dependence a of the conditional correlation is shown to depend crucially on the multivariate GARCH speci cation in use. Journal of Political Economy.. this will give directions for future research. Lubrano. Common Persistence in Conditional Variances". 498 505. Wooldridge. Characterizing volatility impulse functions implied by the multivariate GARCH model can be seen as a promising strategy to characterize dynamic relationships between conditionally heteroskedastic series whenever the multivariate GARCH model serves the purpose of approximating the underlying data generating process. Econometrica. Bollerslev. 167 186. 72.F. 11 . Louvain-la-Neuve. R. To this end. as in the univariate case the multivariate GARCH model is not able to account for asymmetric responses to shocks see e. we assumed the estimated models to be equal to the true data generating process. but extensions to multivariate exponential GARCH models will inevitably require assumptions about the innovation distribution. T. 1988. Volatility shocks are highly persistent for both series. As asymmetry is an important issue of asset price dynamics. Throughout our discussion of the empirical VIRF. Bollerslev. Engle. J. 31. Engle and Ng. Belgium. As another direction of future research.g.M.

and Nijman.F. 247 264. 1993. Engle. 251 276. Tauchen.. 143 172. Ng. Macroeconomics and Reality". Drost. C. 987 1008. Multivariate Simultaneous Generalized ARCH".-L. W. Koop. 1994. Measuring and Testing the Impact of News on Volatility". V. 1980. Lee. 12 . K.F. Lin. G. 318 334. 122 150. 1995.. Hill. Rossi. T. Journal of Business & Economic Statistics. T. 61. Econometric Reviews. G.M. R. 11.R. Granger.. Econometric Theory. Autoregressive Conditional Heteroskedasticity with Estimates of the Variance of U. Quasi Maximum Likelihood Estimation and Inference in Dynamic Models with Time varying Covariances". S. 1987. Introduction to Multiple Time Series Analysis.M.. New York. Pesaran. 11. 74. Potter. 61. Review of Economic Studies. Engle. Berlin. 1993. Ruiz.. Kroner. M. 1997. Gri ths. R. Lutkepohl 1991. Impulse Response Analysis in Nonlinear Multivariate Models".. 1988.E. 6. 1749 1778. 1993. R. The Journal of Finance.. Nelson..G. Econometrica. 119 147.E. J..H. R.. Journal of Econometrics. 48. Stationarity and Persistence in the GARCH1.Bollerslev. A. in ation". 48. Nonlinear Dynamic Structures". Sims. Springer Verlag. Econometrica. 50. Gallant.. 871 907. Econometric Theory. 55. C. 1982. Co-Integration and Error Correction: Representation. Wiley. 15. T. Multivariate Stochastic Variance Models". Econometrica. A.C.1 Model".C. Harvey. Wooldridge. 1 48. Estimation and Testing". Impulse Response Function for Conditional Volatility in GARCH Models". P.K. R. Econometrica. G. Temporal Aggregation of GARCH Processes". Engle.. 1992. F. 1990. Introduction to the Theory and Practice of Econometrics. W. 15 25. 1996. and Shepard.K.C. Engle. 909 927. H. Econometrica. D. Judge. E.. Lutkepohl.F. N. 61.

lnDEM=USD lnGBP=USD lnDEM=GBP Figure 1a: log exchange rates series Figure 1b: VAR5 residual series. GBP lower line 13 . DEM upper line.

0 14 . 0 t 2. 1. DIAG GBP USD. BEKK 1. BEKK GBP USD. The plots on the left show the response of DEM USD volatility to shocks in DEM USD .0 t Figure 2: VIRF for the estimated diagonal upper panel and BEKK models lower panel. 0 t 2.0 2.0 t DEM USD. on the right the response of GBP USD volatility to shocks in GBP USD . DIAG 1.DEM USD.

0 15 . 0 t 2. on the right to shocks in GBP USD . BEKK 1.DEM USD.0 t DEM USD.The plots on the left show the response to shocks in DEM USD . Upper panel: diagonal model. DIAG 1.0 2. lower panel: BEKK model.0 t Figure 3: VIRF for the correlation between DEM USD and GBP USD. DIAG GBP USD. 0 t 2. BEKK GBP USD. 1.

0 16 .0 2.0 2.0 Figure 4: VIRF for the BEKK model and simultaneous shocks in DEM USD and . Upper left: Response of DEM USD volatility.0 2.DEM USD GBP USD 1.0 2.0 Correlation 1. lower panel: reponse of the correlation. 1.0 1. upper right: response of GBP USD volatility.

Are you sure?

This action might not be possible to undo. Are you sure you want to continue?

We've moved you to where you read on your other device.

Get the full title to continue

Get the full title to continue listening from where you left off, or restart the preview.

scribd