A Hybrid Stacking Ensemble Model with Temporal Validation and SHAP Explainability for Intelligent Financial Fraud Detection

Financial fraud detection in digital transactions remains a critical challenge for modern financial institutions. This study proposes an intelligent hybrid stacking ensemble that integrates LightGBM, HistGradientBoosting, and Logistic Regression, optimized using Optuna and evaluated through time-based cross-validation. Across five temporal folds, the proposed model achieved outstanding predictive performance, with F1-scores ranging from 0.96 to 1.00 and ROC AUC scores approaching 1.00. When tested on unseen future data, it maintained an F1-score of 0.94 for the minority (fraudulent) class and an overall ROC AUC of 0.9999, confirming strong generalization capability. SHAP-based explainability revealed that features such as transaction amount ratios and balance differences were the dominant factors influencing predictions, aligning well with domain intuition. Compared with benchmark models including Autoencoder, LSTM, and Isolation Forest, the proposed ensemble demonstrated superior accuracy, interpretability, and robustness highlighting its practical value for real-time fraud detection in financial systems.