Explainable machine learning for high frequency trading dynamics discovery

High-frequency trading (HFT) plays an essential role in the financial market. However, discovering and revealing trading dynamics remains a challenge in Fintech. In this study, we propose a novel explainable machine learning approach: Feature-Interpolation-based Dimension Reduction SCAN (FIDR-SCAN) to address the challenge by creating a trading map. The trading map deciphers an HFT security's trading dynamics by marking the status of each transaction, grouping transactions in clusters, and identifying the trading markers. The proposed method presents new feature interpolation techniques to build a more informative and explainable feature space, unveiling hidden trading behaviors. It mines HFT data in their low-dimensional embedding to seek exceptional trading markers and classify the statuses of transactions. We validate the meaningfulness and effectiveness of the trading markers discovered by FIDR-SCAN in trading as well as examining its special characteristics. Additionally, we apply the proposed algorithm to cryptocurrency data and achieve reliable performance. We design AI trading algorithms by reusing trading markers identified during explainable trading dynamics discovery, applying them to HFT stock and cryptocurrency markets, besides constructing trading machines using identified trading markers. To the best of our knowledge, this study is the first to use interpretable machine learning to reveal HFT trading dynamics.