Previsão de preços de ações no setor bancário brasileiro por meio do uso de redes neurais recorrentes de tipo LSTM

Abstract

From the conception of the Adaptive Markets Hypothesis (AMH), studies applying machine learning techniques have emerged to identify how it is possible to achieve financial gains through predictions. The aim of this study is to make predictions with models based on Long Short-Term Memory (LSTM) neural networks regarding the closing price of five major banks operating in the Brazilian market, using macroeconomic variables and indicators from the CAMELS system. The results demonstrated that, overall, the proposed models were unable to generalize new data and produce better forecasts than the theoretical Random Walk (RW) model. Diebold-Mariano tests were conducted comparing the metrics produced by the LSTM models to those produced by the theoretical model. Due to the observed statistical significance, there are indications that the analyzed assets follow the efficient market theory. However, specific scenarios showed that the model was able to produce superior forecasts to the Random Walk model, opening paths and opportunities for new studies to deepen the discussions.