Real-Time Credit-Card Approval with AID-Score: An AI-Driven Credit-Scoring and Policy-Fusion System
Abstract
Rapid digitalization and tighter regulation expose the limits of traditional scorecards and rule engines for credit-card approval, especially in calibration, explainability, and low-latency operation. This paper introduces AID-Score - an AI-Driven Credit Scoring system for real-time credit card approval in the banking system, that produces calibrated probabilistic scores, human-readable explanations, and auditable decisions. The approach fuses ensemble supervised learners (XGBoost, LightGBM, calibrated Multi-Layer Perceptron (MLP) with Platt scaling for probability calibration, SHapley Additive exPlanations (SHAP) for local explanations, and a Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS)-inspired multi-criteria fusion layer to enforce policy constraints; thresholds are chosen by expected-loss minimization under stratified temporal evaluation. Key results on benchmark datasets: AUC-ROC = 0.994, AUPRC = 0.974, Brier = 0.050, Expected Calibration Error (ECE) = 0.135, p95 inference latency ≈ 164 ms. In business terms, at a representative operating point, AID-Score reduces portfolio expected loss (EL) to ≈ ₹53.75 crore versus ≈ ₹59.59 crore for a LightGBM baseline, a reduction of ≈ ₹5.84 crore (≈ 9.8%). Integrating AI scoring, Multi-Criteria Decision Making (MCDM) justification, and deployment optimizations yields an explainable, audit-friendly decision engine ready for real-time credit-card approval with measurable portfolio benefits.