Deep Learning vs. Financial Fraud Real-Time Detection in High-FrequencyTrading

Abstract

HFT systems are sensitive to microseconds, and generate orderbook streams that present novel challenges to the detection of fraud related behaviors like spoofing and layering. Current algebraic based surveillance strategies are ineffective in describing the nonlinear temporal patterns and subtle manipulation schemes that exist in contemporary financial markets since these strategies are typically solely rule-based. This paper presents an exploratory investigation of DL architectures to support real-time fraud detection in HFT, albeit with very low costs in terms of latency and in spite of predictions with a high level of accuracy. We test temporal convolutional networks (TCNs) and lightweight Transformers as well as machine learning- and rule-based baselines under a mix of historical data collected on a limit-order-book (LOB) and simulator-generated manipulation scenarios. We combine latency-aware acceleration techniques, including quantization, pruning, and micro-batching within a streaming design that can complete inference in sub-5 Ms. Experimental outcomes support the claim that DL models can perform detection better at extremely low false-positive rates even as operational service-level objectives are met. In addition to benchmarking, we cite difficulties of distributional robustness, deployment tradeoffs and explain ability, providing a reproducible framework and methodological improvements to applying deep learning to real time fraud detection in high frequency financial settings.