Predicting Corporate Bankruptcy Using Deep Learning Algorithms

Predicting a company's bankruptcy is a non-trivial task, complicated by the fact that company insolvency cases are rare, thus, unevenly distributed. In this research work, we propose a Transformer-based architecture for bankruptcy prediction using a three-year temporal window of firm-level accounting data. We train the Transformer Encoder with class-balancing, EMA-based validation, and an automated hyperparameter search, and further ensemble the models to enhance stability. Experiments on a public U.S. stock market bankruptcy dataset demonstrate state-of-the-art performance. Our ensemble achieves a test AUC of 0.903, surpassing the previous best 0.847 on the same benchmark. Notably, this improvement is attained with a shorter input window (3 years vs. 4 years), which indicates that the model captures relevant time dependencies more effectively. These results highlight the power of Transformer-based models for financial risk predictions.