The mechanics of VAR forecast pooling—A DSGE model based Monte Carlo study

This paper analyzes the mechanics of VAR forecast pooling and quantifies the forecast performance under varying conditions. To fill the gap between empirical and purely theoretical research we run a Monte Carlo study and simulate the data from different New Keynesian DSGE models. We find that equally pooling VAR forecasts outperforms single predictions in general and that the gains are substantial for sample sizes relevant in practice. In contrast, the estimation of theoretically optimal weights or model selection is advisable only for very large data sets hardly available in practice. Notably, equally pooling forecasts from small-scale VARs can even dominate forecasts from large VARs including all relevant variables. Given our results, we advocate the use of equally pooled predictions from parsimonious VARs as an easy to implement and competitive forecast approach.

► We highlight under which circumstances pooling of VAR forecasts outperforms other approaches.
► We employ Monte Carlo techniques and simulate the data from anestimated DSGE model for the U.S.
► Hence, any VAR used for forecasting suffers from misspecification or mis-estimation.
► It is a risky strategy to rely on optimal pooling schemes or model selection procedures.
► For some target variables pooling small-scale VARs dominates large VARs.