Machine Learning in Bankruptcy Prediction

Abstract

Bankruptcy prediction is a well researched and important topic that can be important in many decision making processes. In this work we ontribute by analysing the predictive strength of payment behaviour data, and applying machine learning techniques on this data in a realistic setting. First, we review the topic of bankruptcy prediction. Second, we introduce the novel type of data available to a pension fund, along with the challenges in structuring the data appropriately. We then develop Logistic Regression, Neural Network, Random Forest, and Decision Tree models. We show that models based on payment behaviour are very well capable of predicting bankruptcy in the next 12 months. We conclude that Random Forests outperform the other techniques, while Logistic Regression models in conjunction with the Elasticnet regularization technique closely follow.