Pair trading of Commodity Futures in China through the lens of intraday data A Machine Learning Framework and Empirical Analysis

In statistical arbitrage, paired trading, as a market-neutral strategy, is widely used because of its simple method and easy implementation. This paper constructs a machine learning framework for commodity futures matching trading from four aspects: systematic feature extraction, unsupervised learning, pairing filtering rules and algorithmatic trading based on spread prediction. This paper selects 47 highly liquid commodities in China's commodity futures market and establishes trading strategies using a continuous commodity index constructed from 5-minute high-frequency data within the 2019-2021 intraday. The results show that the machine learning framework selects potential matching combinations that cannot be identified by traditional commodity classification from all possible matching combinations. After taking transaction costs into account, the results of the out-of-sample experiment confirm that the machine learning framework's pair selection program can produce a more profitable and robust commodity arbitrage portfolio, with 1.8 Sharp and 13.8% annualized return, but the in-sample traditional commodity classification arbitrage is more profitable. When deep learning prediction model is introduced to compare with non-prediction and traditional linear prediction model, the traditional threshold model with no prediction of spread is superior to the prediction model, and LSTM prediction is superior to ARIMA prediction. The conclusion of this paper can provide important enlightenment for designing better statistical arbitrage strategies of commodity futures and monitoring abnormal changes of futures market spreads