Forecasting Financial and Macroeconomic Variables Using Data Reduction Methods: New Empirical Evidence

Abstract

In this paper, we empirically assess the predictive accuracy of a large group of models based on the use of principle components and other shrinkage methods, including Bayesian model averaging and various bagging, boosting, LASSO and related methods. Our results suggest that model averaging does not dominate other well designed prediction model specification methods, and that using a combination of factor and other shrinkage methods often yields superior predictions. For example, when using recursive estimation windows, which dominate other “windowing” approaches in our experiments, prediction models constructed using pure principal component type models combined with shrinkage methods yield mean square forecast error “best” models around 70% of the time, when used to predict 11 key macroeconomic indicators at various forecast horizons. Baseline linear models (which “win” around 5% of the time) and model averaging methods (which win around 25% of the time) fare substantially worse than our sophisticated nonlinear models. Ancillary findings based on our forecasting experiments underscore the advantages of using recursive estimation strategies, and provide new evidence of the usefulness of yield and yield-spread variables in nonlinear prediction specification.