Machine Learning, Corporate Bankruptcies and Variable Selection

Abstract

I employ a variety of machine learning techniques to predict corporate bankruptcies. I compare machine learning techniques' predictions with the ones of reduced-form regressions and structural models. To assess the performances of different models, I compute a range of scores both in-sample and out-of-sample. I show that neural networks produce better predictions than other machine learning methods, reduced-form regressions and structural models. I provide evidence that a small set of variables consistently predict bankruptcy for firms of different dimensions, over different periods and in different industries.