Comparison of regime switching, probit and logit models in dating and forecasting US business cycles

Abstract

Three non-linear model specifications are tested for their efficacy in dating and forecasting US business cycles, viz. a probit specification, a logit specification — both binomial and multinomial alternatives — and a markov, regime-switching specification. The models employ leading indicators compiled by the Economic Cycle Research Institute as putative explanators. They are tested within sample to determine their relative abilities to produce a business cycle chronology similar to the official NBER chronology. They are also tested in a post-sample context to test their relative abilities in anticipating future turning points with the result that the regime-switching model with time-varying transition probabilities performs the best.