Forecasting in VAR models with large datasets

This paper deals with model selection and forecasting in vector autoregressions (VARs) in situations where the set of available predictors is inconveniently large to accommodate with methods and diagnostics used in traditional small-scale models. Available information over this large dataset can be summarized into a considerably smaller set of variables through factors estimated by the dynamic factor model. Even in the case of reducing the dimension of the data, the number of significant factors may still be large. To solve this problem Bayesian model selection methods are introduced to the VAR framework. Model estimation and selection of predictors is incorporated through a stochastic search variable selection (SSVS) algorithm which is easily implemented within the MCMC framework. We apply this method to a forecasting VAR with 128 potential predictors and show how variable/factor selection can save degrees of freedom and improve out of sample fit in high dimensional specifications that otherwise would suffer from the proliferation of parameters.