Agente de aprendizado por reforço tabular para negociação de ações

Abstract

Supervised learning models applied in the context of financial asset trading have been proposed and studied for more than two decades. Although many studies have succeeded in demonstrating good results in terms of financial yields and risk, this approach suffers from important limitations such as the need for constant retraining, especially in the presence of large market fluctuations and the difficulty in converting a good model in terms of prediction accuracy into a system that generates high financial yields. These limitations can be overcome with the use of Reinforcement Learning techniques. In this approach, an agent can learn to trade financial assets so as to maximize total gain or minimize risk through its own interaction with the market. In addition, it is also able to keep itself updated with each modification of the environment, eliminating the need for retraining since the agent is always learning. To obtain evidence of these properties, a reinforcement learning agent was proposed and developed using a tabular SARSA algorithm modeling. Afterwards the agent was applied to a set of stocks with varying trend patterns in order to observe how the agent behaves in terms of its strategy in each trend situation. In addition, a financial trading agent based on supervised learning was also developed using an LSTM neural network to compare its performance with that of the proposed reinforcement learning agent. Both agents were applied to a set of 10 stocks from the brazilian stock market Bovespa in the year 2018 and its performance were assessed in terms of financial yield, risk and accuracy. The experimental results provided evidence not only of the limitations of the proposed supervised learning agent, but also of the aforementioned properties of the reinforcement learning agent in adapting to changes in the market in order to produce financial gains with less accumulated financial losses.