Assessing Data Imbalance in Financial Distress Prediction: A Comparative Approach of Machine Learning and Economic Models

This study aims to compare the effectiveness of machine learning models and economic models in predicting corporate bankruptcy, with a focus on addressing the issue of data imbalance. In this context, the number of companies experiencing financial difficulties is significantly smaller than that of healthy companies, which can lead to bias in predictions. The method used is an experiment with various data handling techniques and involves several classification models, namely Decision Tree, Neural Network (NN), K-Nearest Neighbors (KNN), Case-Based Reasoning (CBR), Support Vector Machine (SVM), and Merton Structural Model, which are tested on several data scenarios with resampling techniques, including Random Oversampling (ROS), Random Undersampling (RUS), and a combination of both. The evaluation results show that the Decision Tree, excluding ROA variables, and the Neural Network provide the best performance, with the Decision Tree achieving 86% accuracy and an AUC of 77.75, and the Neural Network achieving 86.76% accuracy and an AUC of 90.5. Other models, such as KNN and SVM, exhibit lower performance, achieving around 80% accuracy and a lower AUC. Based on these results, Decision Tree without ROA and Neural Networks are the best choices for predicting corporate bankruptcy. This study also demonstrates that financial models, such as the Merton Structural Model, are not significantly affected by data imbalance. The ultimate goal of this study is to provide recommendations for more reliable prediction models that enable financial institutions, investors, and companies to make more informed strategic decisions, as well as reduce financial risks through the early detection of companies at risk of failure.