Algoritmos de machine learning para previsão de ações da B3

Abstract

The prediction of movement in stock market prices has been an interesting area in the research of machine learning algorithms in recent years, due to its complex and dynamic content. In addition, typical stock market volatility makes forecasting more difficult. Thus, this work aims to propose a methodology that predicts the trend of the shares of three Brazilian companies traded on the Brazilian stock exchange. Therefore, we want to facilitate the recognition of movements in the market, in addition to assisting decision-making for investors. Furthermore, this work proposes to compare the predictions of the assets using three powerful machine learning algorithms, known as Random Forest, Artificial Neural Networks and Support Vector Machines. In addition, we propose to use a function (kernel-SVM) -- used for image recognition -- to solve the prediction problem. Technical indicators (such as the Relative Strength Index, the stochastic oscillator, among others) are used as inputs to train the models. Finally, we evaluate perfomance indicators such as accuracy, precision, recall and specificity. The efficiency of the models depends on the technique of separating the training and test samples used, varying between 18% and 76% of accuracy for the temporal separation technique against 71\% and 97\% of the random separation technique.