Since the 21st century,software and hardware technologies related to the computer field have flourished,laying a good foundation for the rise of artificial intelligence technology,and quantitative trading has attracted more investors due to its efficient and stable performance.In this paper,we explore the potential of deep reinforcement learning and meta-learning in quantitative trading,and propose an agent based on LSTM deep neural network to learn the time model in stock data;and utilizes the policy gradient training method to trade automatically,which is based on current market condition and historical data.1.Aiming at the shortcomings of traditional quantitative trading methods in characterizing financial signals,an optimized method for de-noising Grubbs is proposed to process financial data,and combining them with quantitative trading systems can capture the characteristics of financial market accurately.The experimental results show that the optimized process of Grubbs de-noising can greatly improve the ability of agent to represent stock data,thereby surpass the profitability of traditional methods.2.Because traditional quantitative trading methods have difficulty in processing high-dimensional data and it is not easy to mine the historical significance of stock data,this paper proposes a quantitative trading method based on deep reinforcement learning.The agent utilizes LSTM deep neural network,which can explore the optimal trading policy while memorizing the time-series relationships in stock data.The experiment proves that the performance of the trading agent based on the LSTM network is better than the agent based on the fully connected neural network,and the highest rate of return of the former is 280%.3.Due to the large amount of noise and redundant data in financial data,although financial indicators can characterize the internal relationship and trends of financial data,there are many financial indicators and a large number of financial indicators have conflicts and passivation issues in different situations.To characterize the market state reasonably and achieve the goal of maximizing accumulated profits,we select and combine the characteristics of stock indicators.The stock trading system proposed in this paper has been tested retrospectively on the Chinese stock market:in most cases,trading agents can benefit from excess and have good performance.4.Due to the excessive dependence of deep reinforcement learning on a large amount of training data,the trading agent could not quickly adapt to the complex environment.In order to improve the learning ability and the generalization ability of the agent,model-agnostic meta-learning is combined with deep reinforcement learning,and the agent's learning on new task is guided by the previous experience.The comparison of experimental results shows that the trading agent can quickly adapt to changes in the environment,which can balance the benefits and risks,and can also explore better trading strategies.This paper applies deep reinforcement learning and meta-learning methods to quantitative trading in the stock field,so that when faced with intricate financial market,the trading algorithm can reasonably characterize the market state,mine real-time hidden patterns of high-dimensional stock data,and select the optimal policy to obtain accumulated profits.So,this method has practical value. |