Exploring Ensemble Methods for Predicting Corporate Bankruptcy

Corporate bankruptcy prediction is an important issue in finance and risk management. The critical challenges in corporate bankruptcy prediction are addressed in this study using machine learning approaches with ensemble methods. This research aims to evaluate and compare the effectiveness of different ensemble methods in predicting corporate bankruptcy, especially in predicting the bankrupt class. Eight ensemble models-Gradient Boosting, Ada Boost, XG Boost, Cat Boost, Voting Classifier, Bagging, Random Forest, and LightGBM-were implemented using a dataset from Taiwan Economic Journal for the years 1999–2009, which includes company financial data for the period 1999 to 2009. The dataset consists of 95 attributes covering various financial ratios, asset growth, liquidity and leverage. The experimental results show that all models have high accuracy in detecting non-bankrupt companies, with an average accuracy above 96%. Among all models tested, AdaBoost demonstrated the best performance in the bankruptcy prediction with the highest recall of 24% for bankruptcy class, although the overall accuracy remained at 96%, making it a strong candidate for identifying potential bankruptcy risks.