Evaluating the predicting power of ordered probit models for multiple business cycle phases in the U.S. and Japan

We investigate the probability forecasting performance of a three-regime dynamic ordered probit model framework suitable to forecast recessions, low growth periods and accelerations for the U.S. and Japan. In a first step, we apply a non-parametric dating algorithm for the identification of these three phases. We compare the pseudo-out-of-sample forecasting skills of an otherwise standard binary dynamic probit model with a three-regime dynamic ordered probit framework by means of a rolling-window exercise combined with time-varying indicator selection. Based on a set of monthly macroeconomic and financial leading indicators, the results show the superiority of the ordered probit framework to forecast all three business cycle phases up to six months ahead under real-time conditions. Apart from standard probability forecast evaluation measures, receiver-operating curves and related summarizing statistics are computed.