Financial Fraud Detection in Algorithmic Trading Systems Using BERT Variants and Time-Series Embedding

Algorithmic trading systems have revolutionized financial markets by enabling rapid, automated trade execution. However, their complexity and speed have also made them susceptible to sophisticated financial fraud, such as spoofing, layering, and quote stuffing. Existing fraud detection methods often rely on either statistical models or isolated analysis of textual or time-series data, lacking the ability to capture nuanced interactions between contextual and temporal patterns. To address these limitations, this study proposes a novel framework called Temporal-Aware Dual-Stream Transformer (TADST), which integrates BERT variants for semantic understanding of financial logs with Transformer-based time-series embedding to model sequential market behavior. The dual-stream architecture employs a fusion attention mechanism to align insights from both textual and temporal modalities, thereby enabling the detection of fraudulent activity in real-time. The TADST framework is applied to high-frequency trading datasets, combining order book data with trading sentiment and execution logs. It effectively uncovers hidden patterns indicative of fraudulent strategies. Experimental results demonstrate a significant improvement in efficiency by 98.7%, detection accuracy by 97.4%, robustness against noise by 98.8%, and early identification of manipulative behaviors by 96.2% compared to existing benchmarks.

Ҳали таржима қилинмаган