Machine Learning Classification in Economics: Applications for the Business Cycle and Monetary Policy

Abstract

We explore applications of machine learning for key economic questions. We fit generative and discriminative classification models (naive Bayes and regularized logistic regression, respectively) to the problems of recession dating and monetary policy timing. Substantial attention is devoted to exploring the expected and realized performance of these data science techniques in the context of economic applications. Both generative and discriminative models perform surprisingly well in the context of anticipating NBER recession dating decisions. We can correctly identify the state of the economy 90-95% of the time, with a ten month lead time over the final NBER determination. Optimization-driven dimensionality reduction techniques, namely regularization, help identify the key economic features that indicate a change in the economic state. Classification of Federal Open Market Committee (FOMC) policy decisions is less effective, which highlights the need for careful application of theory to guide the specification of classification models. We consider the implications of the lackluster FOMC results for further efforts to integrate machine learning techniques into the econometric repertoire.