Machine Learning-Based Real-Time Fraud Detection and Risk Prediction in Secure Banking Software Systems

This study investigates machine learning-based real-time fraud detection and risk prediction for secure banking software systems. We comparatively evaluated Logistic Regression, Random Forest, XGBoost, LightGBM, and Neural Network models using a highly imbalanced credit-card transaction dataset. The models were assessed using accuracy, precision, recall, F1-score, ROC-AUC, false-positive rate, and detection latency. The comparative results indicate that ensemble boosting approaches provide the most effective balance between fraud detection performance and operational efficiency. XGBoost achieved the strongest overall performance across the evaluated criteria, while LightGBM demonstrated competitive predictive capability with lower computational latency. Based on these findings, we propose an integrated risk-scoring architecture in which the machine learning model assigns transaction-level fraud risk and supports differentiated responses, including normal approval, step-up authentication, and fraud investigation or blocking. The study further highlights the importance of model governance, explainability, continuous monitoring, concept-drift detection, and secure integration when deploying machine learning within banking environments. Although the experimental evaluation is based on a publicly available benchmark dataset and therefore does not directly represent production-scale U.S. banking data, the findings demonstrate the potential of machine learning, particularly gradient-boosting methods, as a foundation for scalable and adaptive real-time fraud detection systems.