Application of Logistic Regression in Enhancing Digital Credit Risk Management in Commercial Banks
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.
M.M Kololi
Department of Mathematics, Kibabii University, P.O. Box 1699-50200, Bungoma, Kenya.
*Author to whom correspondence should be addressed.
Abstract
For a long time, credit risk has been and so remains, the main cause of concern among lenders, whether digital or conventional. The greatest among bank officers’ key performance indicators is effective management of the quality of loan portfolio. The officers are therefore required at all times, to be vigilant and get to know which borrower will fail to honor the contract by failing to repay the borrowed amounts as and when it falls due. Failure by the borrower to repay the loan results in a credit risk where the bank writes down the value of its assets through impairment. At the same time, the bank’s revenue diminishes as a result of reduced income from the expected interests. The bank is therefore interested in designing a model that will robustly predict the probabilities of default accurately. The study aims to provide an assessment of credit risk on digital loans in commercial banks using logistic regression model.The study employed logistic regression, a widely used statistical method in credit risk modeling, due to its effectiveness in predicting binary outcomes such as loan default versus repayment, and its ability to provide easily interpretable results for key stakeholders.The model was fitted with simulated data obtained from commercial banks’ digital loan portfolio. The dataset consisted of twelve independent variables representing income to loan ratio, debt to income ratio, risk taking behavior, past loan behavior, gender, employment status, credit score, multiple loan applications, application timing and the dependent variable, being the digital loans probability of default. Analysis was done using R software and the parameters estimated using Maximum Likelihood method. The results from the study showed that income to loan ratio and credit score were critical variables in predicting digital loan default. However, variables such as debt to income ratio and past loan behavior did not show significance in determining success or failure of the digital loan outcome in the model.The performance of the model was measured with two mathematical approaches. First, the Area Under the Receiver Operating Characteristic Curve (AUC-ROC) was scored at 0.9579, which implies the model has been very good in predicting the defaults and the robust model in separating the defaulting borrowers from the non-defaulting ones. Second, the Hosmer-Lemeshow GOF test yielded the p value equals to 0.8475 which means that there are no significant differences between observed and predicted probability of the data and so the model is appropriate to the data.
Keywords: Digital credit provider, logistic regression, maximum likelihood, loan-to-income ratio