Bankruptcy Prediction of Privately Held SMEs Using Feature Selection Methods

Abstract

We test alternative feature selection methods for bankruptcy prediction and illustrate their superiority versus popular models used in the literature. We apply these methods to a comprehensive dataset of more than one million financial statements covering the entire universe of privately held Norwegian SMEs in 2006-2017. We find that input variables chosen by an embedded least absolute shrinkage and selection operator (LASSO) method yield the best in-sample fit, out-of-sample performance, and stability. This finding holds across different time periods and is robust to using either discrete hazard models with logistic regression or a deep artificial neural network in the estimation. We also contribute to the literature by illustrating the superiority of LASSO in a real-world simulation of a competitive credit market. Specifically, using the variable set chosen by LASSO generates bank profits that are about 50% higher than those resulting from using the second best performing set of bankruptcy predictors. Finally, we show that model performance can be further improved by running feature selection methods on sub-sets of the company universe, such as for example within-industries.

Keywords: Bankruptcy prediction, Feature selection methods, LASSO, Deep learning, Bank profitability

JEL Classification: C25, G17, G33, M41