Bayesian multivariate autoregressive models with structured priors

A variational Bayesian (VB) learning algorithm for parameter estimation and model-order selection in multivariate autoregressive (MAR) models is described. The use of structured priors in which subsets of coefficients are grouped together and constrained to be of a similar magnitude is explored. This allows MAR models to be more readily applied to high-dimensional data and to data with greater temporal complexity. The VB model order selection criterion is compared with the minimum description length approach. Results are presented on synthetic and electroencephalogram data.