Estimating multivariate GARCH and stochastic correlation models equation by equation

A new approach is proposed to estimate a large class of multivariate volatility models. The method is based on estimating equation-by-equation the volatility parameters of the individual returns by quasi-maximum likelihood in a first step, and estimating the correlations based on volatility-standardized returns in a second step. Instead of estimating a $d$-multivariate volatility model we thus estimate $d$ univariate GARCH-type equations plus a correlation matrix, which is generally much simpler and numerically efficient. The strong consistency and asymptotic normality of the first-step estimator is established in a very general framework. For generalized constant conditional correlation models, and also for some time-varying conditional correlation models, we obtain the asymptotic properties of the two-step estimator. Our estimator can also be used to test the restrictions imposed by a particular MGARCH specification. An application to financial series illustrates the interest of the approach.