Inclusive and Transparent Loan Prediction: A Cost-Sensitive Stacking Model for Financial Analytics

Loan eligibility prediction is critical in the financial sector, directly influencing credit access and risk management. This study proposes a cost-sensitive stacking ensemble model that integrates fairness-aware preprocessing, explainable AI techniques, and a custom loss function to address key challenges in loan prediction. The model combines Logistic Regression, Random Forest, and XGBoost as base learners with Ridge Regression as the meta-classifier, optimizing predictive accuracy, fairness, and interpretability. Experimental results on a real-world dataset demonstrate that the proposed model outperforms baseline methods, achieving higher accuracy (84.55%), F1-Score (90.85%), and lower log loss (0.3821). SHapley Additive exPlanations (SHAP) provide feature-level insights, enhancing transparency and regulatory compliance. Additionally, fairness-aware preprocessing mitigates biases, promoting equitable credit access. The study contributes a practical, interpretable, and ethical solution for automating loan decisions, and addressing financial and societal objectives. Future work includes exploring dynamic data integration, broader fairness metrics, and scalability optimization.