Automated Stock Trading with Deep Reinforcement Learning

Chiu, Derek Toh Hui (2023) Automated Stock Trading with Deep Reinforcement Learning. Final Year Project (Bachelor), Tunku Abdul Rahman University of Management and Technology.

Text RDS_Derek Chiu Toh Hui_FullText.pdf Restricted to Registered users only Download (2MB)

Abstract

Making passive income from the stock market is kind of famous nowadays but there are only a few people who can make money from the stock market. The reason for losing money can be news influence, personal emotion issues, economic effect and so on. Learning to invest in the stock market may cost a lot of time so building a machine learning model that can remove those external factors, buying and selling stocks will be a good idea. This project will use different tools such as sentiment analysis for news headlines sentiment analysis, web scraping for data collection, feature engineering and reinforcement learning for the stock trading process. For data collection, web scraping will be used to collect the information needed such as news headlines, stock prices and so on rather than using an API which will be limited to the number of users per day. It will get that information from the HTML code of the specified webpage. For news headlines sentiment analysis, the pipeline will be used for sentiment analysis to predict whether it is positive, negative and neural. Preprocessing of the text will also be done by removing characters, stopwords and lemmatization. For feature engineering, external features will be created such as moving average, rate of change and so on to have more information for reinforcement learning agents. PPO, TRPO, A2C and RecurrentPPO agents will choose and compare. PPO will be the best one based on the test dataset which will be the stocks information for a certain period of time and the result will be the profit the agents earn, mean rewards and standard deviation. The strengths of this project are it is free and removes external factors. The weakness will be that four of the models cannot successfully earn profit from the test dataset.