Forecasting financial and macroeconomic variables using data reduction methods: New empirical evidence

In this paper, we empirically assess the predictive accuracy of a large group of models that are specified using principle components and other shrinkage techniques, including Bayesian model averaging and various bagging, boosting, least angle regression and related methods. Our results suggest that model averaging does not dominate other well designed prediction model specification methods, and that using “hybrid” combination factor/shrinkage methods often yields superior predictions. More specifically, when using recursive estimation windows, which dominate other “windowing” approaches, “hybrid” models are mean square forecast error “best” around 1/3 of the time, when used to predict 11 key macroeconomic indicators at various forecast horizons. Baseline linear (factor) models also “win” around 1/3 of the time, as do model averaging methods. Interestingly, these broad findings change noticeably when considering different sub-samples. For example, when used to predict only recessionary periods, “hybrid” models “win” in 7 of 11 cases, when condensing findings across all “windowing” approaches, estimation methods, and models, while model averaging does not “win” in a single case. However, in expansions, and during the 1990s, model averaging wins almost 1/2 of the time. Overall, combination factor/shrinkage methods “win” approximately 1/2 of the time in 4 of 6 different sample periods. 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 model specification.