The Virtues of VAR Forecast Pooling: A DSGE Model Based Monte Carlo Study

Since the seminal article of Bates and Granger (1969), a large number of theoretical and empirical studies have shown that pooling different forecasts of the same event tends to outperform individual forecasts in terms of forecast accuracy. However, the results remain heterogenous regarding the size of gains. As there are numerous sources for the large variation of the resulting gains, it is difficult to estimate the improvement in accuracy based on empirical findings. Consequently, we use Monte Carlo techniques which enable us to identify the gains of pooling from VAR forecasts under lab conditions. In particular, the results are allowed to vary with respect to sample size, forecast horizon, number of pooled forecasts, weighting scheme and structure of the model economy. Given strict lab conditions, our setup of the experiment yields a quantification of the virtues that can be obtained in almost any forecast situation. The analysis shows that pooling leads to a substantial reduction of MSE of about 20%, which is comparable to the elimination of estimation uncertainty. Most notably, this reduction is already obtained with an average of about four different forecasts.