Point and density forecasts for the Euro area using many predictors : are large BVARs really superior?

Forecast models with large cross-sections are often subject to overparameterization leading to unstable parameter estimates and hence inaccurate forecasts. Recent articles suggest that a large Bayesian vector autoregression (BVAR) with sufficient prior information dominates competing approaches. In this paper we evaluate the forecast performance of large BVAR in comparison to its most natural competitors, i.e. averaging of small-scale BVARs and factor augmented BVARs with and without shrinkage. We derive point and density forecasts for euro area real GDP growth and HICP inflation conditional on an information set which is appropriate for all approaches and find no consistent outperformance of the large BVAR. While it produces good point forecasts, the performance is poor when density forecasts are used to evaluate predictive ability. Moreover, the ranking of the different approaches depends inter alia on the target variable, the forecast horizon, the state of the business cycle, and on the size of the dataset. Overall, we find that a factor augmented BVAR with shrinkage is competitive in all setups.