Working Paper: CEPR ID: DP4333
Authors: Atsushi Inoue; Lutz Kilian
Abstract: A common problem in out-of-sample prediction is that there are potentially many relevant predictors that individually have only weak explanatory power. We propose bootstrap aggregation of pre-test predictors (or bagging for short) as a means of constructing forecasts from multiple regression models with local-to-zero regression parameters and errors subject to possible serial correlation or conditional heteroskedasticity. Bagging is designed for situations in which the number of predictors (M) is moderately large relative to the sample size (T). We show how to implement bagging in the dynamic multiple regression model and provide asymptotic justification for the bagging predictor. A simulation study shows that bagging tends to produce large reductions in the out-of-sample prediction mean squared error and provides a useful alternative to forecasting from factor models when M is large, but much smaller than T. We also find that bagging indicators of real economic activity greatly reduces the prediction mean squared error of forecasts of US CPI inflation at horizons of one month and one year.
Keywords: Bootstrap aggregation; Forecasting; Model selection; Pretesting
JEL Codes: C22; C52; C53
Edges that are evidenced by causal inference methods are in orange, and the rest are in light blue.
| Cause | Effect |
|---|---|
| bagging (C78) | improved forecast accuracy (C53) |
| bagging (C78) | reduced variance of forecasts (G17) |
| bagging (C78) | outperform unrestricted models (C52) |
| bagging (C78) | perform better than traditional forecasting methods (C53) |
| heterogeneity among predictors (C21) | effectiveness of bagging (C52) |