FinEdge: A Resource-Aware Stacking Ensemble for Financial Fraud Detection on Edge Devices

Financial fraud detection requires accurate models that can operate efficiently on resource-constrained edge devices. This study proposes FinEdge, a probability-level stacking framework combining Random Under-Sampling (RUS) with Extreme Gradient Boosting (XGBoost), Synthetic Minority Over-sampling Technique (SMOTE) with Optimized Random Forest, and Optimized Light Gradient Boosting Machine (LightGBM) through a class-balanced Logistic Regression meta-classifier. Independent F1-optimal thresholds are selected using held-out validation data. On the original PaySim split, FinEdge achieved a Recall of 0.8896, an F1-score of 0.8973, and an area under the precision–recall curve (PR-AUC) of 0.9605. Across five independent runs, FinEdge obtained the highest mean PR-AUC (0.9646), whereas Random Forest achieved a slightly higher mean F1-score (0.8894 versus 0.8857). Paired bootstrap analysis supported FinEdge’s PR-AUC advantage, while the F1-score difference remained inconclusive. Excluding post-transaction balance features reduced FinEdge’s F1-score to 0.8346 and PR-AUC to 0.9006. On a Raspberry Pi 4, FinEdge achieved an amortized inference time of 0.426 ms per transaction under batched prediction. These findings demonstrate the ranking-performance benefits of stacking while highlighting feature-availability constraints and the computational overhead associated with edge deployment.