Working Paper: CEPR ID: DP9096
Authors: Cristina Conflitti; Christine De Mol; Domenico Giannone
Abstract: We consider the problem of optimally combining individual forecasts of gross domestic product (GDP) and inflation from theSurvey of Professional Forecasters (SPF) dataset for the Euro Area. Contrary to the common practice of using equal combination weights, we compute optimal weights which minimize the mean square forecast error (MSFE) in the case of point forecasts and maximize a logarithmic score in the case of density forecasts. We show that this is a viable strategy even when the number of forecasts to combine gets large, provided we constrain these weights to be positive and to sum to one. Indeed, this enforces a form of shrinkage on the weights which ensures good out-of-sample performance of the combined forecasts.
Keywords: Forecast Combination; Forecast Evaluation; High-Dimensional Data; Real-Time Data; Shrinkage; Survey of Professional Forecasters
JEL Codes: C22; C53
Edges that are evidenced by causal inference methods are in orange, and the rest are in light blue.
| Cause | Effect |
|---|---|
| Constraints on weights (D20) | Improved forecast accuracy (C53) |
| Optimal weights derived from minimizing MSFE (C51) | Better out-of-sample performance for GDP forecasts (E17) |
| Optimal weights derived from minimizing MSFE (C51) | Better out-of-sample performance for inflation forecasts (E27) |
| Optimal weights (H21) | Decrease in forecast error (C53) |
| Equal weighting schemes (C46) | Higher forecast error (C53) |