Corporate Bankruptcy Prediction via Mechanics-Grounded Feature Engineering and LightGBM

Abstract

Predicting corporate bankruptcy is key to managing financial risks. This paper analyzes financial data of Taiwanese listed companies from the Taiwan Economic Journal (1999-2009), and covers the entire prediction process from exploratory analysis to model evaluation. This study considers a dataset of 6,819 instances with 220 positive and 6,599 negative bankruptcy cases, resulting in an imbalanced class of approximately 1:30. Instead of traditional metrics, this paper compute PR-AUC (area under the precision-recall curve), and with no fancy techniques and on a baseline, LightGBM achieved a PR-AUC of 0.487, which is a better result than logistic regression, decision trees and a random forest. The contribution to the field is the introduction of 21 new features based on the bankruptcy risk theory about profitability, pressure from debt, growth, and efficiency and liquidity, which resulted in a PR-AUC of 0.527. The combination of standardized, strict preprocessing, the SMOTE technique for oversampling the minority class, and a strict search for optimal model hyperparameters resulted in a PR-AUC of 0.542 and a ROC-AUC of 0.956. In the final model, the feature based on the bankruptcy risk theory ranked the sixth most important. This order of importance confirms that the introduction of features based on bankruptcy risk theory, rather than the choice of model or resampling techniques, achieved the best results.