COMBINING INFORMATION FOR PROBABILITY FORECASTS: APPLICATION TO US RECESSIONS

Abstract

There are two popular approaches for pooling information sets to produce forecasts: combination of information sets in a single model and combination of forecasts from different models. In this paper, I compare the pseudo out-of-sample accuracy of both approaches using probit models for US recession probability forecasts. For combining information sets, I consider combining estimated common factors underlying a set of macroeconomic and financial variables in addition to directly using the variables. Combination forecasts are derived from several univariate model forecasts with simple averaging, Bates-Granger (1969) weights, and Bayesian model averaging. Results indicate that combining the variables in a single model outperform the combinations, except when the former is not specified with variable lags selected on information criterion. However, the differences in accuracy are not statistically significant and the combination forecasts can still be reliably used as benchmarks, particularly in cases of possible model mis-specification.