Improving the Accuracy of Financial Bankruptcy Prediction Using Ensemble Learning Techniques

Anthonia Oluchukwu Njoku, Olushina Olawale Awe, Berthine Nyunga Mpinda

Abstract: Financial institutions have been seeking ways to improve their bankruptcy prediction capabilities to mitigate the disruptive effects of future bankruptcies. One such way is using machine learning models. However, financial datasets are often imbalanced, posing a significant challenge for building effective predictive models. In this work, three resampling techniques are used to produce the datasets that were used for model building: oversampling, undersampling, and hybrid sampling. We evaluate the effectiveness of these sampling techniques on five machine learning models (Logistic Regression, Bagging, Random Forest, Support Vector Machine, Neural Networks) in predicting financial bankruptcies. We also investigate the impact of ensembling on model performance by stacking the high-performing individual models using a logistic regression meta-classifier. Our results show that hybrid sampling provides a better balance of accuracy and accountability for the minority (bankrupt) class, which makes it a suitable balancing technique for imbalanced financial datasets. Additionally, ensembling the models using stacking improved the performance of the models, resulting in a better performance for predicting bankruptcies. Remarkably, our proposed model demonstrated an outstanding accuracy of 99.75% while models from existing literature, and previous studies reported accuracies ranging from 83% to 98% for similar ensemble stacking tasks. Results from this study will be useful for practitioners in the finance sphere in making informed decisions, managing risks and choosing the right models for bankruptcy prediction.