Working Paper: CEPR ID: DP3313
Authors: Massimiliano Marcellino
Abstract: It is rather common to have several competing forecasts for the same variable, and many methods have been suggested to pick up the best, on the basis of their past forecasting performance. As an alternative, the forecasts can be combined to obtain a pooled forecast, and several options are available to select what forecasts should be pooled, and how to determine their relative weights. In this Paper we compare the relative performance of alternative pooling methods, using a very large dataset of about 500 macroeconomic variables for the countries in the European Monetary Union. In this case the forecasting exercise is further complicated by the short time span available, due to the need of collecting a homogeneous dataset. For each variable in the dataset, we consider 58 forecasts produced by a range of linear, time-varying and non-linear models, plus 16 pooled forecasts. Our results indicate that on average combination methods work well. Yet, a more disaggregate analysis reveals that single non-linear models can outperform combination forecasts for several series, even though they perform rather badly for other series so that on average their performance is not as good as that of pooled forecasts. Similar results are obtained for a subset of unstable series, the pooled forecasts behave only slightly better, and for three key macroeconomic variables, namely, industrial production, unemployment and inflation.
Keywords: European Monetary Union; Forecast Pooling; Nonlinear Models; Time-Varying Models
JEL Codes: C20; C53; E30
Edges that are evidenced by causal inference methods are in orange, and the rest are in light blue.
| Cause | Effect |
|---|---|
| Pooling forecasts (G17) | Lower MSFE (C51) |
| Instability in series (C62) | Performance of pooling methods (C52) |
| Type of model (nonlinear vs. pooled) (C52) | Forecast accuracy (C53) |