Multivariate stochastic volatility models: Estimation and a comparison with VGARCH models

Abstract

A multivariate stochastic volatility (MSV) model is presented together with an estimation method. Estimation of the MSV model requires specialized estimation techniques since the volatility is a dynamic latent variable. The simulated maximum likelihood technique is expanded to allow for estimation of models with multiple latent variables and is applied to the MSV model. The MSV model is compared with a multivariate GARCH model. The MSV model has fewer parameters and higher likelihood values than the multivariate GARCH models. The univariate models are compared in a Monte Carlo experiment where the likelihood domination is confirmed.