Credit Scoring in Digital Lending: A Stepwise Logistic Regression and Weight of Evidence Approach in Kenya

Mobile money lending has grown rapidly in recent years, with one of the major challenges being the determination of borrowers' creditworthiness, a classification problem complicated by credit risk and the choice of appropriate scoring models. Credit scoring, which uses statistical analysis of historical borrower data to estimate repayment likelihood, remains a critical tool for financial institutions in assessing lending risk. This study evaluates borrower creditworthiness using historical data from Mobipesa Limited, a digital lending company based in Nairobi, Kenya. Data were collected from 495 borrowers over one year, comprising 33 predictor variables (32 numeric and one categorical). Stepwise logistic regression was used to reduce variables and automatically obtain an optimal model. The dataset was split into 70% training and 30% testing sets. Using R, the training data were binned and transformed using the weight-of-evidence (WoE) transformation, after which probabilities of default (PD) for individual borrowers were computed from the cumulative distribution function of the standard logistic model. The PDs represent the likelihood of repayment for each borrower. The developed model achieved an area under the Receiver Operating Characteristic (ROC) curve of 0.825, reflecting 82.5% accuracy in distinguishing reliable borrowers from potential defaulters on the Mobipesa platform. These findings enabled the company to refine its borrower database, inform the public on credit approval thresholds, and strengthen credit policy. Moreover, other firms in digital lending ecosystems can draw on these insights to improve credit scoring models and variable-selection approaches, thereby enhancing overall lending efficiency and risk management