SME default prediction: A systematic methods evaluation

This study evaluates the performance of various methodologies used in the literature to predict failures in small- and medium-sized enterprises (SMEs) using a data set of U.S. SMEs. By evaluating over 6,100 models and data subsample combinations, we find that the light gradient boosting machine (LightGBM) exhibits the best out-of-sample predictive performance, closely followed by extreme gradient boosting (XGBoost). These two estimation methods perform best using their built-in feature-selection mechanisms and do not require sample rebalancing. However, for most other estimation methods, feature selection and sample rebalancing are critical. For example, logistic regression (Logit) performs significantly better with appropriate feature selection and sample rebalancing. We also provide an overview of the importance of various features in predicting failures.