Predictive Modeling of Digital Credit Risk in Commercial Banks Using Machine Learning Algorithms
Isaiah N. Barasa *
Department of Mathematics, Kibabii University, P.O. Box 1699-50200, Bungoma, Kenya.
Samson W. Wanyonyi
Department of Mathematics and Computer Science, Pwani University, Kilifi, Kenya.
*Author to whom correspondence should be addressed.
Abstract
Normally, financial institutions issue loans to customers aiming to recover the repayments within the scheduled time. However defaults, common with digital credit, cause them significant financial risks. Advancement in technology has created digital loan products. Unfortunately,in developing economies such as Kenya,the growth rate of such products has outpaced development of sound credit risk assessment systems. These factors, along with the limitations of conventional risk models, which are static, necessitate the adoption of dynamic models capable of capturing complex and non linear patterns among variables. This study addressed this gap by developing predictive models using machine learning algorithms. Using 6,000 simulated loan records, necessitated by restricted access to real world data, the study evaluated performance of the two models, among which, random forest emerged as the most robust. Even though random forests model initially achieved an Area Under Curve score of 1.0, under simulated conditions, a 10-fold cross validation model produced a more realistic. mean AUC of 0.90. This reflects strong discriminative ability and high predictive accuracy. These findings demonstrate that machine learning algorithms can enhance risk assessment framework in commercial banks thereby minimizing financial
Keywords: SHAP, digital credit, random forest, XGBoost, temporal dimension, simulated, loan to income ratio