| The stock market plays an important role in the financial market.How to obtain higher stock returns is a hot topic for financial industry researchers and investors for a long time.With the rapid development of big data and artificial intelligence,the amount of stock data has also increased explosively.The traditional fundamental analysis of stocks is becoming more and more difficult.Experts and scholars in the financial industry also try to apply machine learning technology to quantitative trading.This paper mainly completes the following work:Firstly,the traditional quantitative trading method has some shortcomings in the representation of financial signals.This paper proposes to preprocess the stock data by denoising and optimizing the stock data with Grubbs method.In the data preprocessing,this paper uses the Grubbs method to eliminate the stock noise and improve the agent’s representation ability of stock data,so as to improve the profitability.Through experiments,it is proved that the profit level of the transaction agent optimized by Grubbs has been improved compared with the traditional quantitative transaction agent.Secondly,according to the characteristics of stock trading,combined with the theory of deep reinforcement learning,an algorithm based on T-DQN is proposed,and a trading agent model based on T-DQN algorithm is constructed.The traditional DQN algorithm is difficult to deal with the problem of time series characteristics,so the long-term and short-term memory network is introduced to deal with the time series characteristics of stocks.Moreover,the traditional DQN agent can not immediately obtain feedback rewards from the environment.This paper constructs a temporary memory pool to store the current time state,action and next time state.When going through a series of decisions until making the selling decision,it can obtain the income or loss corresponding to a series of decisions.Finally,the trading agent model based on T-DQN algorithm is constructed with this algorithm,and the performance comparison experiment is carried out with the traditional DQN agent model.Good decision-making results are obtained in the experiment.When the stock trend fluctuates greatly,the excess return can still be obtained,and the maximum rate of return reaches 280%.In the research of this paper,an improved T-DQN algorithm is proposed,and the stock trading agent is realized based on T-DQN algorithm.Good trading results are achieved in the experiment,which has certain guiding significance for the further research of this problem in the future. |
