Aprendizagem por reforço direto aplicado em séries temporais

Abstract

This paper presents the development of a neural network model using direct, recurrent, and deep reinforcement learning, as well as obtaining and pre-processing the data. To extract the input parameters of the network from the raw data, formulas were used that are also used by humans to identify patterns in time series when it comes to stock market shares, such as the Rate of Price Change. The algorithms to build the model applied in this paper are presented according to the development sequence. The work aims to demonstrate the effectiveness of using this model to predict complex time series, as in the case of stock price series. For its construction, the Tensorflow framework was applied in conjunction with the Python programming language, which was chosen for its ability to be used on various platforms, and thus can be used for numerous purposes. Finally, the results of the tests are presented, where some of them outperform the stock price by 300%, besides proposing strategies for future optimizations.