High Frequency Trading Strategy for the 30 Year Treasury Bond

Abstract

We construct new features based on order book data and separate them into three groups, e.g., time-insensitive features, time-sensitive features and cointegration features. For time-insensitive features, we applied serval transformation on imbalance in different levels, and some other features based on order book data. For time-sensitive features, we constructed features with historic information. Besides, we extracted information about deleting and adding order book to construct features on it. For cointegration, we applied linear regression, online regression and Kalman filter to both the treasury data and the corresponding futures data to construct cash and futures cointegration features separately. Then, we predicted the fair-price for each quote given each single feature and combination of features. At last, we designed two smart algorithms to trade 30 Year Treasury Bond given the predicted fair-price. We found that combination features from different groups can help to reduce transaction cost by 95% compared with one tick-size. We believe that the new features we constructed can extract more information from order book, and can be very effective for trading strategies.