AI-DRIVEN LOAN ELIGIBILITY PREDICTION SYSTEM WITH EXPLAINABLE AI AND BANK TRANSACTION INTEGRATION

Abstract

Automated loan eligibility assessment is critical for financial institutions seeking to balance credit access with risk management. This paper presents a complete end-to-end system that combines a LightGBM binary classifier with a real-time, MongoDB-backed two-day rolling application window, live bank ledger synchronisation, and SHAP-based post-hoc explainability. The model achieves 81.81% cross-validated accuracy (5-fold, std=2.64%) on the standard Loan Prediction dataset with 13 features including two engineered ratios: TotalIncome and LoanToIncome. An operating threshold of 0.50 yields 82.6% hold-out accuracy (ROC-AUC=0.91). A Streamlit dashboard surfaces approval probabilities, SHAPranked reasons, a what-if simulator, and combinatorial actionable recommendations for borderline applicants.