Deep learning and reinforcement learning for trading financial assets

• Main Author: Mundhra, Divyesh
• Other Authors: Ng Wee Keong
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2022
• Subjects: Business::Finance::Assets; Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: https://hdl.handle.net/10356/156541
• Institution: Nanyang Technological University

The emergence and advancements in Deep learning and Artificial Intelligence have been disruptive for the field of financial markets. Prediction of financial asset prices and trading them based on historical prices have been an area of research for a long time and none of the methods till now have been able to give a 100% accuracy for this task. Hence, it remains an unexplored territory to a significant depth as it involves decisions which can result in huge profits for individuals as well as business investors. In this era of big data, these advancements help the user to process massive amounts of historical data and current market conditions to make predictions that influence their investment decision. In this report, we design customized neural network frameworks which can be used for financial asset predictions. We compare multiple deep learning methods and provide our analysis on the results obtained by them. After appropriate scaling, the results obtained RMSE of as low as 0.08 which indicates that some models, implemented in this report, have remarkable accuracies. Moreover, we comprehensively describe our implementation of a trading agent which can be optimized to take buy and sell decisions based on the current stock price. This was achieved by using Deep Q Learning which is one of the algorithms used to solve Reinforcement Learning problems. Additionally, the research led to the development of two web applications which can be used by individuals to get an overview of the short-term future financial market by providing them with predictions of financial asset prices. These web applications use various deep learning models for their predictive tasks. The backend is combined with a user-friendly front end which allows general public to use it for their wealth management decisions. Lastly, the project achievements and future works have been discussed. In addition to contributing to the existing methods in this field, this project also aims to set the ground for future research in numerous directions.