Similarity-based path forecasting of US recession periods

Abstract We develop a nonparametric similarity-based approach for binary time series that exploits recurring historical patterns to construct probability forecasts for all feasible multi-period outcome sequences. In contrast to conventional horizon-specific parametric models, our path forecasts are obtained simultaneously for all the horizons and remain internally consistent across them. Simulation experiments demonstrate that our method delivers accurate and robust performance in realistic sample sizes. In an empirical application to US business cycle data, our approach successfully anticipates the onset of the past three recessions about one year in advance and provides informative predictions of their expected duration.