Portfolio Management Via Deep Reinforcement Learning

Abstract

The stock market is a data filled environment that cannot physically be iterated through by a human in order to find the best opportunities. Therefore, this paper investigates the strategies that utilize deep reinforcement learning, specifically Deep Q Networks (DQN), to not only optimally trade one stock but all of the constituents of the S&P 100 index for just under a five year period (January 6, 2017 to November 29,2021). The models for single stock analysis were capable of having a higher return than that of the stock they followed, returning over 100% regardless of the starting hyper-parameter, and the combined model was able to beat the S&P 100's growth with an unbounded model, returning 362.23%, and compete with the S&P 100 when limitations were placed upon it, returning 89.55%.