1- , ma.mahmood@iau.ac.ir
Abstract: (14 Views)
Despite the growing use of machine learning in credit scoring, many domestic studies still rely mainly on traditional statistical models and static borrower characteristics, while limited attention has been paid to the role of real behavioral, transactional, and repayment-performance data in post-disbursement credit risk monitoring. To address this research gap, this study compares the performance of four stepwise logistic regression models and the LightGBM algorithm in predicting credit default, using data from 119,050 loan facilities granted to individual customers of Resalat Qard al-Hasan Bank during the period between 26 March 2022 and 18 March 20240. The target variable was defined based on repayment delays of more than 90 days, and model performance was evaluated using AUC, Accuracy, Recall, F1-Score, and Balanced Accuracy. The knowledge contribution of this study lies in providing empirical evidence on the effectiveness of real banking data, focusing on behavioral and transactional variables, comparing a classical statistical model with a machine learning algorithm, and assessing model performance in identifying the minority class under imbalanced credit data. The results indicate that repayment-related variables, particularly the number of overdue installments and outstanding debt balance, are the most important predictors of default. Although the fourth logistic regression model achieved a high overall AUC of 0.98, it performed poorly in identifying high-risk customers, with a Recall of only 0.12%. In contrast, LightGBM identified 92.2% of high-risk customers and outperformed logistic regression on imbalance-sensitive evaluation metrics. These findings suggest that, in imbalanced credit datasets, relying solely on AUC and Accuracy can be misleading, while Recall, F1-Score, and Balanced Accuracy are more informative for assessing a model’s ability to detect high-risk borrowers. Therefore, in the post-disbursement monitoring scenario, machine learning algorithms based on behavioral and transactional data can provide a more accurate and reliable framework for credit risk management in Iranian banks.
Keywords: Credit risk, Default, LightGBM, Logistic regression, Behavioral and transactional data, machine learning algorithm
Send email to the article author
| Rights and permissions | |
 | This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License. |
