A HYBRID GRADIENT BOOSTING AND NEURAL NETWORK ARCHITECTURE WITH DUAL EXPLANATION FOR CREDIT RISK PREDICTION WITH LLM

This research presents a hybrid approach combining gradient boosting algorithms with neural networks for credit risk prediction. By integrating various models including XGBoost, LightGBM, CatBoost, and residual neural networks, we developed a stacking ensemble architecture that achieved significant predictive accuracy on an imbalanced credit-risk dataset, with a test ROC-AUC of 0.96 and PR-AUC of 0.91. To ensure transparency and regulatory compliance in financial decisions, we introduced a dual-explanation framework using SHAP and LIME to provide quantitative feature importance. Furthermore, a Large Language Model (LLM), Phi-2, is utilized to generate human-readable, non-technical explanations of the model’s predictions, bridging the gap between high-performance machine learning and accessible, explainable AI. The proposed system demonstrates how advanced ensemble methods and LLMs can enhance the interpretability and trustworthiness of credit risk models, making them suitable for deployment in regulated financial environments.