Machine learning in bankruptcy prediction: A literature review

Abstract

Substantial research efforts have focused on the topic of bankruptcy prediction. Researchers have analyze bankruptcy and default events using various statistical and machine learning techniques for risk management. Academics have also employed various data sources and processing techniques. In this thesis, various proposed models since 2017 in the literature are evaluated and compared. A literature review is conducted to compare the use of data and processing techniques. The models are then implemented and compared in a uniform testing environment to determine the most optimal method. The data from 2009 to 2021 on US firms from the Compustat database is used for the experiment. Several features sets, both from the literature and from widely used feature selection methods, are generated and applied in the experiment to determine the most suitable set of features for predicting bankruptcy.

The experiment results have highlighted several insights in relation to the application of machine learning models in bankruptcy prediction. The models trained on Altman’s original features set outperforms those of the the other sets, including recently proposed features sets. This may be due to the diverse set of firms in the experiment, which are from various industries with varying financial conditions. Regarding machine learning models, ensemble methods, random forest, categorical boosting, and gradient boosted decision tree, outperform the other techniques in almost every evaluation metric.

Most of the recently proposed methods show lackluster performances compared to previously employed models. The results encourage future research in a more focused manner, which focuses on firms in a single field or scope, to avoid introducing noises and affecting classifying ability. Furthermore, more research on the interpretability of the models would be beneficial to professionals in the field.