Une lecture probabiliste du cycle d'affaires américain

This article explores 35 years of the U.S. business cycle with a multivariate hidden Markov model using monthly data. It identifies ten U.S. time series offering particularly reliable information to detect recessions. It also assesses the performances of different and complementary ?recession models? based on Markov processes and draws two main conclusions: (1) simple univariate models are decisive to monitor the business cycle providing that the series are shown to be highly reliable; (2) models adding a multivariate dimension are useful but work only marginally better than a simple summary. The primary determinant of model quality appears to be the variables?information content. The author introduces a new reading of the business cycle using a preferred recession model and concludes by discussing the limitations of leading indicators and ?real-time detection.?