Early Warning System for Debt Group Migration: The Case of One Commercial Bank in Vietnam

Abstract

This study utilizes machine learning models, including Logistic Regression, Support Vector Machine, Decision Tree, and Random Forest, in the early warning system for debt group migration in a Vietnamese commercial bank. In predicting customers’ overdue debt migration (B Score), the RF model achieves the highest accuracy of 81.84%. However, if the priority is to reduce Type I errors, SVM performs better with a recall of 91.48%, although the accuracy drops to 46.62%. When predicting customers’ debt group improvement (C Score), SVM proves to be the optimal model in terms of both accuracy and criteria based on Type II errors, with an accuracy of 71.6% and precision of 62.3%. When applied to new datasets, the evaluation criteria decrease, but SVM remains the most optimal model for both B Score and C Score. Additionally, the research results demonstrate that tuning the model parameters leads to a significant improvement in accuracy compared to the default parameters.

Keywords

• machine learning models
• debt group migration
• b score
• c score
• model parameters tuning
• e50
• e51
• g33
• g21
• g24