Bagging or Combining (or Both)? An Analysis Based on Forecasting U.S. Employment Growth

Abstract

Forecasting a macroeconomic variable is challenging in an environment with many potential predictors whose predictive ability can vary over time. We compare two approaches to forecasting U.S. employment growth in this type of environment. The first approach applies bootstrap aggregating (bagging) to a general-to-specific procedure based on a general dynamic linear regression model with 30 potential predictors. The second approach considers several methods for combining forecasts from 30 individual autoregressive distributed lag (ARDL) models, where each individual ARDL model contains a potential predictor. We analyze bagging and combination forecasts at multiple horizons over four different out-of-sample periods using a mean square forecast error (MSFE) criterion and forecast encompassing tests. We find that bagging forecasts often deliver the lowest MSFE. Interestingly, we also find that incorporating information from both bagging and combination forecasts based on principal components often leads to further gains in forecast accuracy.