Evaluation of Forecasting Performance Using Bayesian Stochastic Search Variable Selection in a Vector Autoregression

This paper examines forecasting performance of a vector autoregressive (VAR) model by a Bayesian stochastic search variable selection (SSVS) method. We use several artificially generated data sets to evaluate forecasting performance using a direct multiperiod forecasting method with a recursive forecasting exercise. We find that implementing SSVS prior in a VAR improves forecasting performance over unrestricted VAR models for either non-stationary or stationary data. As an illustration of a VAR model with SSVS prior, we investigate US macroeconomic data sets with three variables using a VAR with lag length of ten, and find that the SSVS restrictions on insignificant parameters alleviates over-parameterized problem of VAR and thus offers an appreciable improvement in forecast performance