Detecting fraud in the digital financial system poses challenges as the types of fraud, imbalanced data, and lack of transparency of machine learning models are not handled properly. The static rule-based methods do not adapt to fraud types, whereas the machine learning models act like a black- box. This paper introduces a hybrid, adaptive and explainable fraud detection framework by combining ensemble learning, behavioral feature engineering and rule-based decision making process. Random Forest, AdaBoost and LightGBM are utilized for prediction probability generation, which is enriched using a hybrid risk-based decision framework. Class imbalance is tackled with SMOTE and an adaptive threshold technique is used for dynamic control of precision and recall, without any retraining. Feature-level explainability is achieved using SHAP and the LLM generates human understandable explanations. The framework is put to use using FastAPI, which provides near real time processing and experimental evaluation results prove its significance for BFSI real-world applications by providing accuracy of 95.15% and AUC of 0.9926.

Index Terms—Fraud Detection, Ensemble Learning, Explain- able Artificial Intelligence (XAI), SHAP, Large Language Mod- els (LLM), SMOTE, Adaptive Thresholding, Hybrid Decision Framework, Behavioral Feature Engineering, BFSI, Real-Time Systems.