Adaptive Machine Learning Framework for High-Frequency Trading Detection

High-Frequency trading (HFT) plays a critical role in today’s financial markets, offering increased liquidity, but also introducing concerns about fairness, transparency, and regulation. Existing HFT detection techniques often rely on static datasets, lack real-time adaptability, or struggle with class imbalance. This study proposes a machine learning-based framework for accurate and interpretable HFT detection using historical and real-time data. The model combines microstructural order-level features from the BEDOFIH dataset with dynamic attributes derived from real-time EUR/USD tick data via the OANDA API. A RUSBoost classifier is used to address class imbalance, improving the identification of minority HFT cases. To support regulatory transparency, a regression tree is reverse-engineered from the ensemble model, providing interpretable decision logic without compromising predictive performance. Empirical evaluations show that the proposed system achieves high accuracy, precision, and recall in both historical and live trading environments. The hybrid design ensures robustness, real-time responsiveness, and interpretability, making it suitable for practical deployment in algorithmic trading surveillance systems. In general, this work contributes to a scalable and adaptive solution for HFT detection, addressing key limitations of current methodologies and supporting data-driven decision making in high-speed financial markets.