Bayesian inference for latent factor GARCH models

Abstract

Latent factor GARCH models are difficult to estimate using Bayesian methods because standard Markov chain Monte Carlo samplers produce slowly mixing and inefficient draws from the posterior distributions of the model parameters. This paper describes how to apply the particle Gibbs algorithm to estimate factor GARCH models efficiently. The method has two advantages over previous approaches. First, it generalises in a straightfoward way to models with multiple factors and to various members of the GARCH family. Second, it scales up well as the dimension of the o, bservation vector increases.