We propose a simple new model named a Copula-based Multivariate GARCH model,
or in short C-MGARCH model, which permits modeling conditional correlation and dependence
separately and simultaneously for interested financial returns with non-elliptically
distributed dependent errors. Our approach is based on a transformation, which removes
the linear correlation from the dependent variables to form uncorrelated dependent errors.
The dependence structure is controlled by a copula while the correlation is modeled by an
MGARCH model. The C-MGARCH model can capture the dependence in the uncorrelated
errors ignored by all existing MGARCH models. For every MGARCH model, the
corresponding C-MGARCH model can be constructed.
The paper is organized as follows. Section 2 provides a brief review on MGARCH models.
Section 3 introduces the new C-MGARCH model with uncorrelated dependent errors. Monte
Carlo simulation in Section 4 illustrates how MGARCH and C-MGARCH perform under
the non-elliptical distributions, and shows the likelihood gains of the C-MGARCH models.
Section 5 conducts empirical analysis for comparison of existing MGARCH models with their
corresponding C-MGARCH models in terms of in-sample model selection criteria and outof-
sample density predictive ability. The C-MGARCH models outperform corresponding
DCC, VC and BEKK models when they are applied to a pair of the U.S. equity indices
(NASDAQ and Dow Jones) and two pairs of the foreign exchange rates (French Franc and
Deutschemark, and Japanese Yen and Deutschemark). Section 6 concludes. Section 7 is
Appendix on copulas.
1