Strong selection consistency of Bayesian vector autoregressive models based on a pseudo-likelihood approach (Q820793)

In contrast to Bayesian variable selection for high-dimensional linear regression models with independent and identically distributed observations (where a rich literature exists), a detailed theoretical investigation of Bayesian variable selection procedures for the VAR (vector autoregressive) parameters in high-dimensional settings has not been undertaken. The VAR model, although related to linear regression, leads to a more complex setting. It aims to capture linear temporal interdependencies among multiple time series. And even though the estimation of the VAR model parameters can be formulated as a linear regression problem, there are significant differences from a technical point of view. In the paper under review, a pseudo-likelihood based Bayesian approach for consistent variable selection in high-dimensional VAR models is proposed, which is based on considering hierarchical normal priors on the autoregressive coefficients, as well as on the model space. The authors prove that the posterior probability assigned to the true underlying VAR model converges to 1 under high-dimensional scaling where the dimension of the VAR system grows nearly exponentially with the sample size. A notable feature of this strong selection consistency result is the presented application to linear regression with serially correlated predictors and errors.