Machine learning approaches to corporate failure prediction

Angelina Schmitt (Student)

Student thesis: Master's Thesis

Abstract

Corporate failure prediction is a central topic in financial literature due to the substantial costs associated with firm failures. At the same time, advances in machine learning have introduced powerful new tools for addressing this challenge. This thesis examines the predictive perfor mance of traditional and machine learning models in forecasting corporate failures, with an evaluation across different prediction horizons. Two iterations of datasets were employed: one based on the established Campbell (Campbell et al., 2008) variables and another extended dataset incorporating accounting, market, and macroe conomic information. Firms were observed over consecutive months until disappearance, fol lowing a hazard model framework (Shumway, 2001). Logistic regression served as the baseline and was compared against random forest and extreme gradient boosting (XGBoost). Model performance was assessed using AUC and other evaluation metrics and feature importance was analyzed to shed light on the main drivers of corporate failure risk. The results indicate that Logistic Regression with Campbell variables performs well over short horizons (one month). However, expanding the feature set and applying machine learning models, particularly ensemble models, substantially improves predictive accuracy and ensures greater stability across longer horizons. These findings underscore the importance of combining richer datasets with advanced algorithms, offering both theoretical contributions to the literature and practical implications for early warning systems, risk management, and corporate monitor ing.