Classifying High-Frequency Trading with an Interpretable Machine-Learning Model

Summary

The document summarizes a study that develops a probabilistic classifier to identify high-frequency trading activity from intraday order data. Using French BEDOFIH market records, the researchers engineered features describing orders, including their prices, lifetimes, depth, and distance from the best quotes. They evaluated several methods and selected RUSBoost to address class imbalance, then approximated the ensemble with a regression tree to make its decisions easier to interpret.

On the reported test data, the model’s average true positive rate was 99.86% and its average true negative rate was 73.69%. Order lifetime, order status, account type, and trading-group information were among the most useful distinguishing features. The summary describes short-lived orders and a preference for liquid securities as typical signals. It presents public order-book data as a basis for real-time identification, with possible use in research and oversight. The evidence is limited to the French dataset and a single day in 2017, so performance and transferability to other markets require further study.

Key ideas

• The study uses order-level features to classify high-frequency trading activity.
• RUSBoost was chosen to handle imbalanced classes, and a regression tree was used to explain its decisions.
• The reported test results show a very high true positive rate and a lower true negative rate.
• Order lifetime, order status, account type, and trading-group characteristics helped distinguish HFT.
• The findings come from one French market dataset and day, so broader generalization remains unverified.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.