Comparative analysis of boosting algorithms for predicting personal default

Accurately predicting personal default risk is crucial for financial institutions to manage credit risk effectively. This study conducts a comparative analysis of the performance of boosting algorithms, including AdaBoost, XGBoost, LightGBM, and CatBoost, in predicting personal defaults. The dataset used in the study comprises 7,542 individual customers collected from Vietnamese commercial banks and financial institutions between 2014 and 2022, with 12 features related to the financial and demographic characteristics of the borrowers. All customer-related information is fully anonymized and encrypted during the data collection process to ensure compliance with research ethics. The predictive models are evaluated based on six criteria: Accuracy, Precision, Sensitivity, Specificity, F1 score, and AUC. The results indicate that the LightGBM model has the best performance, demonstrating the ability to efficiently handle large and complex datasets. Additionally, the study identifies the five most significant factors influencing personal default risk: Monthly Liability, Credit Balance, Credit History Length, Max Credit Limit, and Yearly Income. However, the study’s limitations in the size and scope of the dataset may reduce the generalizability of the results when applied to other regions. These findings provide valuable insights that help financial institutions enhance their strategies for managing credit risk effectively.