Modelos de previsão de ativos utilizando aprendizagem profunda e linguagem natural

Abstract

Predicting the future is something that humanity craves for, in order to be able to prepare for possible eventualities. Today, combining facility of access to information through news with the use of programs that “read” the feelings of users on the Internet, it is possible to use computer language to even predict market behavior. In the scope of economy, this prediction enables new ways of intervening in the economic market, providing a glimpse of different scenarios, therefore facilitating the decision-making about investments and also about the economic future of entrepreneurs, companies and governments. In this sense, artificial intelligence tools have become a great bridge between forecasting techniques and engineering techniques, useful in that approach, presenting quantitative treatments and limitations analysis. The present work aims to develop a prediction model for the Brazilian Stock Exchange, developed in the Python programming language, and to verify its predictive capacity through the submission of historical series data in Long Short Term Memory (LSTM) and Deep Neural Network (DNN). For this, news were researched on Twitter from trusted newspapers, and data from classifiers based on Naïve Bayes and Support Vector Machine (SVM) were incorporated into the historical series of the companies Banco do Brasil and Petrobras to be used in the aforementioned neural networks. The results showed that the classifiers had good results, mainly the SVM classifier, presenting an accuracy of 82% of correctly classified news for Petrobras. The neural networks had a reasonable predictive capacity, with RSME of 26.4% using the DNN network, and 35.1% in the LSTM network for the Banco do Brasil. Although the database used was relatively small, which may have influenced the results of this study, papers show that, in general, it is possible to relate news with the rising or falling of the stock market. In this regard, there are still many questions to be clarified in order to find realistic predictions.