Explainable Machine Learning for U.S. Recessions Nowcasting: the Shapley Approach

Much of the recession forecasting literature is focused on producing a model for predicting current and future economic states. Nevertheless, the question of why a model predicts a given class (recession or expansion), that is, which features are mostly responsible for the classification outcome, remains unaddressed. The main goal of our paper is to fill this gap. To address this, we apply Shapley values to provide model transparency. Using high-quality macroeconomic data from multiple U.S. government agencies, we use competing linear and non-linear models. We then compute Shapley values across 154 feature candidates to nowcast US recessions. Our findings offer novel insights into feature importance and model decision-making, which allows rationalizing model predictions. This is useful, for example, in the context of monetary policy, where the central banker needs, not only a good prediction of whether or not we are in a recession, but also a sense of why that prediction was made.