Credit Default Risk Scoring with Explainable ML and Profit-Based Loan Approval

Every decline comes with its reasons, the cut-off is set by profit, and the outcomes are audited for fairness. A probability-of-default model is only half of a lending decision, so this project turns default probabilities into decisions a bank could defend: (1) adverse-action reason codes for every declined applicant - up to four plain-word reasons from the classical scorecard and from the machine-learning model (SHAP), checked against each other; (2) a fairness audit by sex, age band and education (approval gaps, disparate-impact ratio, equal-opportunity gap) for the attribute-blind model and for the same model trained with the attributes; (3) a profit-maximising approval cut-off from an assumed profit per good account and loss per defaulted account, compared with the common 'approve when PD < 0.5' rule, with a sensitivity analysis. A classical WoE scorecard is compared with LightGBM, monotone LightGBM and XGBoost on the public UCI Default of Credit Card Clients data (30,000 Taiwanese card holders, 2005). A small local LLM drafts approval / decline letters behind a rubric gate with a template fallback.

Category: Banking / Fintech

Master of Business Administration Project MBA05 from ProTechSC Academy, with graphical abstract, video demonstration and project guidance.