- Published on
Ensemble of Reinforcement Learned Agents for Stock Trading
Final Project Report Link
Link It may take a few seconds to load.
Motivation
Reinforcement learning agents might be able to learn policies to trade stocks using input market data better than human heuristics or intuition- in particular, an ensemble of reinforcement learning methods may be able to outperform other market strategies.
Abstract
Stock trading and portfolio management are massive, lucrative fields within the broader financial services industry. With the advent of machine learning and improved algorithmic technologies, there has arisen a significant interest in the potential for automating such tasks. If this were done well, one could hope for a more “objective” method of making investment decisions with better risk-adjusted performance than intuition derived human strategies. However, designing an automated model with consistently strong performance is difficult to do in the highly complicated, multi-state environment of the stock market. In this paper, we propose a method for training a diverse pool of specialized reinforcement learned agents (“experts”) and compare two distinct strategies for combining these experts into a functional ensemble, with the first ensemble strategy being an online learning technique and the second being a deep learning one. Both of our expert ensembles were given $1, 000, 000 to manage over the 210 day period from 1 February 2022 to 29 August 2022, and could buy and sell 28 stocks from the Dow Jones Industrial Average index. The performance of our proposed deep ensemble strategy and the research ideas that it inspired has lead us to conclude that the application of deep machine learning to finance is a promising but difficult task.