Big Data Finance: Trading strategy creation using Deep Reinforcement Learning.

Abstract

Deep Reinforcement Learning has been successfully used to control an agent in a high-dimensional space. For this reason, an important question is whether these powerful capabilities can be applied to highly dynamic spaces such as the stock market. This thesis explores how multimodal financial data can be processed to obtain low-dimensional features representing the economic state of a given country and exploited by a trading system to maximize wealth and minimize risks. Allocating resources is crucial when investing in the stock market, given that errors can be very costly. In particular, it is essential that investors navigate the dangerous waters of the collapsing markets unscathed during an economic crisis. With recent advances in ML, it might be possible to do so by helping investors make better decisions when allocating their wealth in a way that maximizes wealth while minimizing market risks. In order to achieve these goals, custom loss functions that control the maximum amount of cash allocated to a specific asset while selecting those shares that maximize wealth are combined with a Deep Reinforcement Learning (DRL) system trained to learn stock trading and evaluated using a set of financial metrics, including ROI and the Sharpe ratio. Two loss functions are used: a barrier method that limits the cash allocation to a maximum value of 35% and the same barrier method combined with a penalty method that punishes the trading agent when the action did not add up to one. When comparing these loss functions to a baseline, they earned higher ROI and obtained better financial metrics. They outperformed the benchmark (S&P 500) and the baseline. In particular, during periods of economic turmoil, they lost significantly less wealth than the benchmark and the baseline models. Conversely, during periods of economic growth, they earned more wealth than the benchmark and the baseline models. Our models developed a diversification strategy that allocated almost an equal amount of resources to each asset in the S&P 500. Based on the analysis of portfolios, this diversification strategy demonstrated that it could earn higher ROI than other strategies and the benchmark and increase the initial capital 12-fold.

Date of Award	31 Dec 2021
Original language	English
Awarding Institution	The University of Manchester
Supervisor	John Keane (Supervisor) & Xiaojun Zeng (Supervisor)