In this section we review the development of multivariate GARCH models and their
wide applications. Understanding and predicting volatilities and correlations of asset
returns has been the object of much attention, since volatilities and correlations are
the two most important elements in financial activities such as asset pricing, asset
allocation decisions, portfolio management and rish assessment.
In the last few decades, so many volatility models have been put forward. The
most popular and successful models among them are the autoregressive conditional
heteroskedasticity (ARCH) model by Engle (1982) and extended to generalized ARCH
(GARCH) model by Bollerslev (1986). The ARCH/GARCH models have generated a
great spectrum of models, which have been applied and tested in many areas. Their
success stems from their ability to capture some stylized facts of the studied time
series, especially for financial time series, such as time–varying volatility and volatility
clustering. See Bollerslev, Engle, and Nelson (1994), Bera and Higgins (1993), and
Kroner and Ng (1998) for a comprehensive survey of the univariate volatility models
and their application. See also Bollerslev, Chou, and Kroner (1992) for a review of
the ARCH modeling in the financial area. Other volatility models include the vast
stochastic volatility (SV) models and etc., which we will not explore here, however.
See Taylor (1994) for a review of the univariate SV models. See Harvey, Ruiz, and
Shephard (1994) for a review of the multivariate SV models.
Although univariate ARCH/GARCH models have been proved to be very powerful
in explaining the stylized facts of univariate time series, researchers find them
unsatisfactorily incapable to examine the characteristics of multivariate time series
simultaneously. Since in reality we are more concerned about the relationships between
volatilities of several markets or assets and variance–covariance matrices of
various portfolios, univariate ARCH/GARCH models seem to be not applicable and
therefore their multivariate generalization stands out to be the better solution. There
3
are generally two directions for modeling the multivariate time series, modeling the
variance–covariance matrix directly and modeling the correlation between the time
series indirectly. Bollerslev, Engle, and Wooldridge (1988) proposed the first multivariate
GARCH model for the conditional variance–covariance matrix, namely the
VEC model, which was a successful attempt towards the first direction. However, this
model is a very general model and very difficult to implement in practice. The number
of parameters in the model is O(K4
) with respect to the dimension of the model and it
is difficult to impose the positive definiteness of the variance–covariance matrix in the
model. Thus, a portion of the subsequent literature is to try to simplify this model.
It should be noted that the advantage of this model is that we can directly interpret
the coefficients in the model. Bollerslev, Engle, and Wooldridge (1988) introduced a
simplified version of the VEC model, the Diagonal–VEC model. This model reduced
the number of parameters greatly and it is relatively easier to derive the conditions
to guarantee the positive definiteness of variance–covariance matrix. However, since
the variance or covariance in the model is only the function of its past observations,
it can not capture the interactions between different variances and covariances.
Engle and Kroner (1995) proposed the BEKK 1 model which can be viewed as
a restricted version of VEC model. BEKK model has a very good property, that is,
conditional variance–covariance matrix is positive definite by construction. But the
number of parameters in BEKK model still increase rapidly with the dimension of the
model. Another problem is that it is hard to interpret the coefficients of the model.
Further simplified models include the Diagonal–BEKK model and the Scalar–BEKK
model. Diagonal–BEKK model faces the same problem of Diagonal–VEC model,
although it reduces the number of parameters greatly. Scalar–BEKK model is too
restrictive as it imposes the same dynamics to all the variances and covariances.
Engle, Ng, and Rothschild (1990) developed another way to reduce the number of
1BEKK is the acronym of Baba, Engle, Kraft, and Kroner who initially wrote the paper.
4
parameters involved in the model by introducing several factors. The main problem
of multivariate GARCH models in most specifications is the very large number of
parameters, which rapidly makes the estimation infeasible as the number of series
increases. Those specifications which bypass this problem, on the other hand, pay
the price in terms of a severe loss of generality. Neither multivariate SV models,
although relatively more parsimonious, are able to handle more than a few number
of series because of their complexity of estimation. The key for dimensionality reduction
stands in the idea of the existence of a few latent variables,