Economics seminar: Variational Bayes inference for large vector autoregressions

Event Date: 22 January 2014

Mr Tomasz Wozniak, University of Melbourne

Time: 16.15-17.30

Dept. of Economics

Rm 2.11 Architecture Bldg

Abstract:

Variational Bayes provides an approximation to the joint posterior distribution of parameters of a model. The approximate posterior is usually accurate and of a tractable form. We show that when applied to large Bayesian Vector Autoregressions, proven to have excellent performance for forecasting of economic variables, Variational Bayes allows for fast and accurate computations of posterior distributions. The algorithms for the Variational Bayes estimation of VAR models with a variety of prior distributions, including hierarchical prior structures are derived. A procedure of choosing the optimal hyper-parameters of the prior distributions with respect to a Variational Bayes measure of the fit in sample is also proposed. Finally, a new estimator of the marginal data density based on the output from both MCMC and Variational Bayes estimation is shown to have good properties.

Published: 23 January 2014