Benchmarking Machine Learning Models for Real-Time Fraud Detection in Digital Banking Transactions

Abstract

Purpose- The rapid expansion of digital banking has intensified the need for fraud detection systems that are not only accurate but also efficient and cost-effective in real-time environments. This study addresses this challenge by evaluating the trade-offs between predictive performance, computational latency, and operational costs across state-of-the-art machine learning models. Methods-We benchmarked a range of algorithms including Random Forest, XGBoost, LightGBM, LSTM, TCN, and Transformer, on two widely used fraud detection datasets (ULB and PaySim). Model evaluation was conducted across predictive performance metrics (precision, recall, F1-score, ROC-AUC, PR-AUC), real-time inference metrics (average and 95th percentile latency, throughput), and cost-related indicators. Findings- The results demonstrate that deep learning models, particularly LSTM and Transformer, achieved the highest detection accuracy but incurred higher computational overhead and inference latency. Gradient-boosted tree ensembles, while slightly less accurate, delivered superior scalability and real-time responsiveness, making them more practical for high-throughput fraud detection scenarios. Cost–benefit analysis further revealed that minimizing false positives and latency offered significant operational savings and improved customer experience, underscoring that accuracy alone is not sufficient for deployment in live banking systems. Conclusion: The study highlights the need for context-aware model deployment strategies. A hybrid layered pipeline (employing lightweight models for rapid initial screening and deep models for secondary verification) emerges as an effective approach to balance detection accuracy, latency, and cost. Beyond benchmarking, the reproducible experimental framework introduced in this work provides a foundation for future research on adaptive, scalable, and collaborative fraud detection strategies in digital banking.