DEVELOPMENT OF RETAIL LOAN SCORECARD USING MACHINE LEARNING

B a c k g r o u n d. Banks use credit scoring to track loan performance, manage provisions, and adjust lending policies. Thismethod assigns points based on a borrower's loan history and unique characteristics, enabling lenders to predict default risk and improve credit conditions for low-risk borrowers. With increased data access and computing power, it is possible to do credit scoring via new methods with possibly better predictive power. This study aims to develop a scorecard for Ukrainian retail borrowers using Credit Registry data, exploring the effectiveness of logistic regression and Support Vector Machine (SVM) methods. Key research questions address the potential for binning data to improve model interpretability, the accuracy probability of default estimates, and differences in decision thresholds across scorecards built with logistic and SVM models. M e t h o d s. The study applies Weight of Evidence (WoE) binning, a technique that transforms variables to establish a monotonic relationship with default risk, thereby improving interpretability and model robustness. Using this binned data, the paper constructs scorecards with logistic regression and SVM. Each scorecard uses predictor variables such as Debt Service to Income (DSTI), age, interest rates, and days overdue to assess default likelihood. Scores are assigned based on each variable's impact on default probability. R e s u l t s. Findings indicate that it is possible to develop a scorecard based on Credit Registry data. Logistic regression and SVM models yield similar score distributions, with high predictive accuracy and robustness as measured by accuracy and F1-score. The scorecard approach provides transparency and interpretability; for instance, borrowers with a DSTI exceeding 40% receive lower scores, indicating higher risk. C o n c l u s i o n s . Banks may use both logistic and SVM models for real-time credit assessments, leveraging accessible borrower characteristics to streamline decision-making. For regulators, the scorecards can support policy frameworks that restrict lending based on borrower risk bins, thus mitigating risks arising from specific retail lending segments.