The Applicability of Machine Learning Models in Corporate Bankruptcy Prediction: Multi-dimensional Prediction of Bankruptcies in Denmark

Abstract

This thesis investigates the applicability of machine learning algorithms for bankruptcy prediction, focusing on the predictive performance and practical considerations associated with implementing Logistic Regression, Random Forest, and Artificial Neural Network models. First, the research evaluates their respective predictive performance on both a raw and a resampled dataset in terms of F1-score and AUC metrics. Second, through expert interviews and a literature review, it investigates the benefits and limitations of applying each model in a practical setting. The study finds that i) the Random Forest-based model displayed superior predictive performance on both datasets, demonstrating its predictive power and insensitivity to the underlying data quality. Next, it was found that ii) while the Logistic Regression model exhibited inferior predictive performance to the other models, it is the most widely utilized amongst the surveyed experts due to its high level of interpretability. It was then found that iii) interpretability is a central concern when implementing machine learning algorithms for bankruptcy prediction. Through this, the study highlights that the use of non-linear machine learning models displays great potential through enhanced bankruptcy prediction accuracy. However, their interpretability causes practical implications which must be carefully considered when applied in a real-world context.