Um estudo de redes neurais recorrentes no contexto de previsões no mercado financeiro

Abstract

Financial time series forecasting is one of the most researched artificial intelligence applications by financial market analysts, both in the academic and corporate world. Within this area, there is a great emphasis on the prediction of stock values, mainly due to the possibility of profitability. One of the most used techniques to make this type of prediction is the application of artificial neural networks. This work is focused on one of the classes of artificial neural networks, the recurrent neural networks. Three types of architecture in this area are used: the Vanilla Recurrent Neural Network (VRNN) model, the Long Short-Term Memory (LSTM) model and the Gated Recurrent Unit (GRU) model. Throughout this paper, models of each of these architectures are created for three different stocks of the Brazilian stock exchange, namely the stocks from Grupo Fleury (FLRY3), Petrobras (PETR4) and Vale (VALE3). The models created for each of the stocks make a prediction of the value of those stocks, and the results are compared to each other with the intention of evaluating which model predicts the values closest to the real values. The accuracy measures used in this work are the mean absolute error and the mean square error. In the end, it was concluded that the GRU and LSTM models obtain similar results, with a slightly better performance from GRU, and VRNN was able to detect patterns in the data, but not as accurately as the other two.