Deep Learning for Real-Time Payment Fraud Detection: Economic and Computational Perspectives

Abstract

This research examines the role of deep learning in enhancing real-time payment fraud detection by integrating economic and computational perspectives. It aims to evaluate how advanced deep learning models improve detection accuracy, reduce fraud-related losses, and support efficient decision-making in high-speed payment environments. The research adopts a systematic literature review approach, analyzing recent national and international scholarly works on deep learning–based fraud detection. The review focuses on studies that employ architectures such as Long Short-Term Memory networks, Convolutional Neural Networks, hybrid models, and generative approaches in real-time payment systems, with attention to both technical performance and economic assessment. The findings indicate that deep learning models consistently outperform traditional rule-based and classical machine learning methods in detecting complex and evolving fraud patterns. These models demonstrate higher precision, recall, and F1-scores while maintaining low latency suitable for real-time processing. From an economic standpoint, the literature highlights significant reductions in fraud losses, false positives, and operational costs, despite increased computational requirements. The results imply that financial institutions can achieve substantial efficiency gains and improved risk management by adopting deep learning–based fraud detection systems, provided that computational optimization and regulatory compliance are addressed. This research contributes by synthesizing economic impact analysis with computational considerations, offering an integrated perspective on real-time fraud detection using deep learning.