Detecting Turning Points with Many Predictors through Hidden Markov Models

This paper explores the American business cycle with the Hidden Markov Model (HMM) as a monitoring tool using monthly data. It exhibits ten US time series which offer reliable information to detect recessions in real time. It also proposes and assesses the performances of different and complementary “recession models” based on Markovian processes, discusses the most efficient and easiest way of encompassing information through these models and draws three main conclusions: simple HMM are decisive to monitor the business cycle and some series are proved highly reliable; more sophisticated models such as the Dynamic Factor with Markov Switching (DFMS) model or Stock and Watson’s Experimental Recession Index seem not to be more powerful than simple (univariate or pseudo-multivariate) Hidden Markov Models, which remain far more parsimonious; combining information in temporal space seems to work marginally better than in probability space for high frequency data. We conclude about leading and “real time detection” properties related to HMM and give some hints for further research.