Explainable Performance: Measuring the Driving Forces of Predictive Performance

We introduce the XPER (eXplainable PERformance) methodology to measure the specific contribution of the input features to the predictive performance of a model. Our methodology is theoretically grounded on Shapley values and is both model-agnostic and performance metric-agnostic. Furthermore, XPER can be implemented either at the model level or at the individual level. We demonstrate that XPER has as a special case the standard explainability method in machine learning (SHAP). In a loan default forecasting application, we show how XPER can be used to deal with heterogeneity issues and significantly boost out-of-sample performance. To do so, we build homogeneous groups of individuals by clustering them based on their individual XPER values. We find that estimating group-specific models yields a much higher predictive accuracy than with a one-fits-all model.