Abstract:
Recent research provides controversial evidence on the stability of yield-curve based binary probit models for forecasting U.S. recessions. This paper reviews so far applied specifications and presents new procedures for examining the stability of selected probit models. It finds that a yield-curve based probit model that treats the binary response (a recession dummy) as a nonhomogeneous Markov chain produces superior in-sample and out-of-sample probability forecasts for U.S. recessions and that this model specification is stable over time. Thus, the failure of yieldcurve based forecasts to signal the 1990-1991 and 2001 recessions should not be attributed to parameter instability, instead the evidence suggests that these events were inherently uncertain.