Credit operations in commercial banks improving through an integrated credit scoring framework based on financial indicators, alternative data and machine learning

The credit processing at the commercial banks of Uzbekistan is currently conducted manually and by using conventional analysis methods, which are characterized by a low level of efficiency and a high number of defaulted loans. The goal of the research is to investigate the performance of automated models of scores to increase the quality and speed of credit provision. For this purpose, four machine learning algorithms are applied to 8,500 loan records of three commercial banks of Uzbekistan for the period of 2020-2024. In order to resolve the problem of class imbalance, the method of SMOTE was employed, and the performance was measured using accuracy, sensitivity, AUC-ROC, and cost-benefit analysis. It was found out that gradient boosting (XGBoost) gave the highest performance, with the value of AUC equal to 0.87 and the sensitivity equal to 0.78. As opposed to expert decisions concerning 350 common files, the machine-learning model detected 13 additional defaults and increased the sensitivity from 57% to 78%. Cost-benefit analysis has confirmed 23.8% decline in the expected cost. Moreover, the time required to process each file declined from 14.8 hours to 20 seconds. Alternative data, including transaction patterns, have raised AUC to 0.90, which is very important for customers lacking credit history. From this analysis, it is clear that machine learning in the credit process provides more than just improving the accuracy of predictions and making processes more efficient; it also brings about significant economic benefits by reducing default costs and risk assessment. The model suggested here, where traditional financial data is used along with digital data, presents a practical solution to commercial banks operating in emerging economies in the digital age.

Hali tarjima qilinmagan