Forecasting recessions with time-varying models

Abstract

This study presents a flexible recession forecast model where predictive variables and model coefficients can vary over time. In an application to US recession forecasting using pseudo real-time data, we find that time-varying logit models lead to large improvements in forecast performance, beating the individual best predictors as well as other popular alternative methods. Through these results, we also demonstrate the following features of the forecast models: (i) substituting roles between the two key features of predictor switching and coefficient change, (ii) considerable variations in the model size (i.e., the number of predictors used) over time, and (iii) substantial changes in the role/importance of major individual predictors over business cycles.