Research On Deep Reinforcement Learning In Stock Trading Model Combining Sentiment Analysis

Deep reinforcement learning as a method to assist intelligence in decision-making has been validated by numerous studies and practices,and is widely used in all aspects of social and economic life.This thesis is dedicated to applying deep reinforcement learning to the financial investment field to build algorithmic trading models and help investors make stock trading decisions,with a focus on the investment model combining sentiment analysis of investor groups and deep reinforcement learning.The main research work and results of this thesis are as follows.(1)Deep reinforcement learning is applied to the field of quantitative stock trading to improve investment returns by learning from massive stock market trading data and continuously optimizing it,using deep reinforcement learning’s ability in perception and decision making.In this thesis,we construct a deep reinforcement learning stock trading model based on deep reinforcement learning theory and related knowledge in the field of quantitative trading,which includes DQN and DRQN.(2)Investor sentiment and stock price trend affect each other,and studying the change of investor sentiment can help investors to make scientific decisions,improve trading win rate and reduce trading risk.The text constructs a sentiment analysis model based on sentiment dictionary and machine learning,collects stock forum posts,quantitatively analyzes the sentiment tendency of the posts,and calculates the daily investor sentiment score as part of the data in the trading model.Among them,the sentiment dictionary is the core of the sentiment analysis method based on sentiment dictionary.The posts made by investors in online forums contain professional words and internet phrases with either positive or negative emotions,reflecting the emotional characteristics of investors.In this thesis,we construct a sentiment dictionary containing two parts,the first part is a general seed sentiment dictionary,and the second part is a self-made financial sentiment dictionary for quantitative calculation of investor group sentiment.(3)Deep reinforcement learning model incorporating sentiment analysis is applied to stock trading modeling.The investor sentiment obtained by sentiment analysis is then combined with the use of deep reinforcement learning to build a stock trading model,and finally five stock trading models are experimentally compared.The experimental results show that the deep reinforcement learning model combined with sentiment analysis proposed in this thesis obtains the highest total return of35.5% in the 2-year comparison experiment,which is 27.1% higher than the average rise of 30 stocks participating in the experiment,and improves the two main indicators of win rate and total return by 6% and 6.5%,respectively,than the deep reinforcement learning model without considering sentiment.The experimental results demonstrate the practical value of the approach used in this thesis to achieve better returns in stock trading.