Evaluating CPB's published GDP growth forecasts; a comparison with individual and pooled VAR based forecasts

We employ a data driven methodology for selecting variables to include in our VAR models and we find that a randomly selected classical VAR model performs worse in most cases than the Bayesian equivalent, which performs worse than our published forecasts in most cases. However, when we pool forecasts across many VARs we can produce more accurate forecasts than we published. A review of the literature suggests that forecast accuracy is likely irrelevant for the non-forecasting activities the model is used for at CPB because they are fundamentally different activities.