Sparse Graphical Vector Autoregression: A Bayesian Approach

In high-dimensional vector autoregressive (VAR) models, it is natural to have large number of predictors relative to the number of observations, and a lack of efficiency in estimation and forecasting. In this context, model selection is a difficult issue and standard procedures may often be inefficient. In this paper we aim to provide a solution to these problems. We introduce sparsity on the structure of temporal dependence of a graphical VAR and develop an efficient model selection approach. We follow a Bayesian approach and introduce prior restrictions to control the maximal number of explanatory variables for VAR models. We discuss the joint inference of the temporal dependence, the maximum lag order and the parameters of the model, and provide an efficient Markov chain Monte Carlo procedure. The efficiency of the proposed approach is showed on simulated experiments and real data to model and forecast selected US macroeconomic variables with many predictors.