In this paper we identify and try to predict the turning points of the Japanese business cycle. As a measure of the business cycle we use a composite economic indicator (CEI). This indicator is endowed with nonlinear dynamics to capture the asymmetries between different cyclical phases. Two types of nonlinear dynamics are considered : Markov switching and smooth transition autoregression (STAR). The performance of these models in terms of forecasting the business cycle turns is compared. Both types of models produce statistically equivalent in-sample forecasting results, whilst the CEI with exponential STAR tends to outperform the CEI with Markov-switching and logistic STAR in the out-of-sample prediction.