In this paper, a large amount of different financial and macroeconomic variables are used to predict the U.S. recession periods. We propose a new cost-sensitive extension to the gradient boosting model which can take into account the class imbalance problem of the binary response variable. The class imbalance, caused by the scarcity of recession periods in our application, is a problem that is emphasized with high-dimensional datasets. Our empirical results show that the introduced cost-sensitive extension outperforms the traditional gradient boosting model in both in-sample and out-of-sample forecasting. Among the large set of candidate predictors, different types of interest rate spreads turn out to be the most important predictors when forecasting U.S. recession periods.

Keywords: recession forecasting, business cycle, machine learning, gradient boosting, class imbalance

JEL Classification: C22, C25, C53, C55, E32