Analysis Of Intraday High-frequency Quantitative Trading Based On GRU-DDQN

With the continuous development of computer software and hardware technology,the application of artificial intelligence in all walks of life is becoming more and more extensive,and the financial market is no exception.Taking the stock market as an example,when trading stocks,a huge amount of data will be generated,and traditional traders can often only take a part of it when facing these massive data for trading judgments,and cannot cover all aspects.And when traders trade,they are very susceptible to their own emotions,so as to make wrong judgments about trading.Nowadays,a variety of powerful data processing algorithms have emerged,and major investment banking institutions have also begun to conduct relevant research on financial quantitative transactions.Based on the current development status of the stock market,this paper introduces the deep reinforcement learning algorithm into the quantitative trading of ETF(referred to as ETF),hoping to tap the trading rules of the ETF market through the combination of recurrent neural network and DDQN(Double Deep Q-learning Network,referred to as DDQN)algorithm,so as to carry out effective trading and enhance the income of investors.The main research content of this paper is divided into three parts: First,according to the existing relevant literature at home and abroad,the shortcomings of the existing research are summarized.Based on the actual situation of the ETF market,a modeling method combining recurrent neural network and deep reinforcement learning is adopted,and it is applied to the quantitative trading of 5-minute data of intraday ETFs.Secondly,on the basis of the original GRU-DQN(Gate Recurrent Unit-Deep Q-learning Network,referred to as GRU-DQN),by improving its network structure and connection methods,reward functions,and action space,it is optimized to GRU-DDQN(Gate Recurrent Unit-Double Deep Q-learning Network,referred to as GRU-DDQN).and apply the GRU-DDQN model to the stock market.It not only solves the problem that DQN(Deep Q-learning Network)networks are prone to overestimation of target action Q values,but also greatly improves the convergence speed of financial time series data when calculating Q values with the help of GRU.Finally,it was compared with the same structure of LSTM-DDQN(Long Short Term Memory-Double Deep Q-learning Network)and other models in multiple sets of experimental data to draw conclusions.The main contributions of this article are as follows:1.When constructing the state space,this paper further processes the data metrics by introducing the Auto-Encoder autoencoder and the GRU neural network.The original 24-dimensional stock technical index is reduced to 12 dimensional by AutoEncoder autoencoder,which greatly reduces the data redundancy and noise between the indicator data,and after decoding,the mean squared difference and the average absolute error between the original data are maintained at a small value.GRU neural network is used for long-term monitoring of stock data,by entering the stock historical opening price,closing price,high price,low price to obtain the closing price forecast for the next period of time,thereby reducing the model’s trading risk,maximum drawdown rate,and improving the final yield.2.Some improvements have also been made in the action space and reward function within the model.For example,in the action space,this article uses more actions to choose from,adding the action of trading 500 shares and trading 1000 shares to the original trading of 100 shares.In terms of the reward function,this article refers to the ETF grid trading method,using the total value as the reward function,and through the final trading process analysis,the model does implement an operation process similar to the ETF grid trading method.3.Recurrent neural networks have inherent advantages over time-series data processing,so this paper introduces the GRU network to calculate the Q value of all actions in the model,whether it is compared to LSTM neural networks or ordinary fully connected neural networks.The GRU network not only ensures the model fitting effect,but also greatly improves the training speed of the model.After comparing the models under multiple sets of experimental data,the GRUDDQN model in this paper can not only achieve a higher final yield,but also achieve a lower maximum drawdown rate through long-term monitoring of stock prices,which means that investors take less risk while obtaining higher returns.Compared with the same structure of LSTM-DDQN,GRU-DDQN has a faster model training speed,and in the field of quantitative trading,training speed is also one of the factors that cannot be ignored.Therefore,GRU-DDQN’s quantitative trading strategy on ETF intraday high-frequency trading data is successful.