Likelihood stabilization for ill-conditioned vector GARCH models (Q2430220)

Pseudo-likelihood estimation is considered for vector GARCH model parameters. It is shown that unstable behaviour of the estimates can arise due to high correlations in the components of the considered vector-valued time series or due to nearly flat likelihood near its maximum. To avoid these difficulties, the authors propose a preliminary data-driven linear transform of the time series analogous to projections on principal directions of the canonical correlations matrix. Performance of the obtained estimating procedures is assessed by simulated and real life financial data.