Real-Time Hybrid Optimization Models for Edge-Based Financial Risk Assessment: Integrating Deep Learning with Adaptive Regression for Low-Latency Decision Making

Abstract

Financial institutions in the United States face mounting pressure to detect fraud and evaluate transaction risk in real time across highly distributed payment infrastructures, including mobile banking, point-of-sale devices, and ATM networks operating on resource-constrained hardware. Deep learning models deliver strong predictive accuracy for fraud detection but often exceed strict latency budgets when executed at the edge. Conversely, lightweight regression systems provide rapid decision-making but sacrifice accuracy under nonlinear transaction behaviors common in U.S. financial environments. This study develops a real-time hybrid optimization framework that dynamically integrates deep neural inference with adaptive regression, guided by an online controller that monitors latency, compute utilization, and confidence thresholds per transaction. Using a U.S. credit card fraud dataset structured as a streaming financial workload, we benchmark hybrid performance against standalone deep learning and regression baselines under simulated edge CPU and memory constraints. Experiments show that hybrid routing reduces inference latency by up to 55 percent compared to deep learning alone, while preserving high recall on fraudulent cases and improving transaction-level risk detection without overwhelming edge hardware. A latency-accuracy Pareto analysis highlights the system’s ability to maintain regulatory-aligned response times without destabilizing detection performance, demonstrating practical readiness for deployment in payment terminals and digital banking infrastructure. These findings suggest that real-time model optimization can significantly enhance operational compliance, fraud resilience, and customer experience across U.S. financial systems, which are increasingly dependent on edge-based decision intelligence.