Corporate failure prediction: An evaluation of deep learning vs discrete hazard models

In recent years, deep learning has emerged as a dominant machine learning method used in a variety of applications, including robotics (such as self-driving cars), speech recognition, text analysis and natural language processing, fraud detection, earthquake prediction, medical image analysis just to mention a few applications. In this paper, we propose an optimum deep learning model using a panel data structure to predict corporate failure. We compare deep learning with the more traditional discrete hazard model which has been widely applied in the finance literature in panel data applications (such as bankruptcy prediction). Based on a sample of 641,667 firm-month observations of North American listed companies between 2001 and 2018 and including many financial and market-based feature variables, our deep learning model can predict corporate failure with 93.71% accuracy. This is significantly more accurate than the discrete hazard model which predicted corporate failure with only 86.95% accuracy. Not only has our deep learning model proven highly effective in corporate failure prediction but it can be potentially applied to many other classification problems in finance involving panel data structures.