Efficient Parameter-Expanded Gibbs Samplers for Dynamic Factor Models

Abstract

A dynamic factor model is estimated with a parameter-expanded algorithm to accelerate standard Gibbs sampling. The proposed algorithm is easy to implement since it involves only draws from standard distributions. It also leads to substantial improvement in the Markov chain Monte Carlo (MCMC) performance compared to prevailing algorithms. Additionally, a heavy-tailed prior is adopted to ease the process of prior parameter elicitation when one has little knowledge. We also implement an efficient simulation algorithm by exploiting the computational advantage of sparse and banded matrices. The performance of the methods is illustrated with simulated data and an application to construct economic indicators.