Automatic Optimization of Trading Strategies Based on Reinforcement Learning

At present, investment transactions in domestic financial markets are gradually shifting from subjective transactions relying on traditional technical analysis to programmatic quantitative strategy transactions. Although a large number of quantitative strategies have been studied in the stock market, there is still insufficient research on quantitative trading strategies in the futures market, and there is still room for improvement in the investment return and risk control of existing strategies in intraday high-frequency trading. In order to improve the profitability and risk management of the high-frequency quantitative strategy of futures, this paper designs a futures trading environment, uses the 1-minute granularity of high-frequency K-line as the environmental state, and constructs the position state and the action space and algorithm of trading operations suitable for futures trading. Lstm-dueling DQN, a deep reinforcement learning model based on LSTM, is adopted to better deal with the state space of sequence input and significantly accelerate the learning speed of the model. Through the experimental comparison of three benchmark models DQN, Double DQN and Dueling DQN (FF-Dueling DQN) based on fully connected neural network, the results show that the cumulative return rate of the proposed strategy in the trading of four black series commodity futures can reach 43%, and the annual return rate can reach 153%. The maximum retracement was controlled within 10.7%. The experimental results show that this strategy can achieve excess returns beyond the performance benchmark in both volatile and trend markets.