A Predictive Risk Framework for P2P Micro-Lending Default Prediction with Anomaly Detection

Machine-learning credit assessment is often discussed as a substitute for manual underwriting, yet high-stakes lending decisions require a more careful design in which predictive models augment human judgement. This paper presents a supervised learning framework for peer-to-peer micro-lending default prediction using a reproducible pipeline built around LightGBM classification, engineered affordability features, class-imbalance handling, Isolation Forest anomaly detection, and SHAP explainability. The dataset is processed through exploratory analysis, feature construction, stratified train-test splitting, median and modal imputation, robust scaling, categorical encoding, SMOTE rebalancing, model training, diagnostic evaluation, business-impact estimation, and global and local explanation. The experimental run reports an AUC-ROC of 0.7459, average precision of 0.3042, macro-F1 of 0.5614, and default-class F1 of 0.3332. At the operating threshold used in the pipeline, the model identifies 4,081 defaults while missing 1,850 defaults, representing $236,020,901.20 in potential principal at risk. SHAP analysis ranks age, interest rate, months employed, credit-to-income, dependents, and co-signer status among the strongest drivers, revealing both useful affordability signals and governance-sensitive proxy variables. The findings support an augmentation thesis: AI can improve default-risk triage, anomaly surfacing, and explanation consistency, but false positives, missed defaults, proxy-feature risks, and highvalue edge cases require human-in-the-loop underwriting, fairness review, and accountable lending policy.