Brazilian Stock Market Forecast with Heterogeneous Data Integration for a Set of Stocks

The significant growth of the Brazilian stock market, coupled with the increase in investors in riskier assets, has generated a demand for automated forecasting tools. This research investigated the behavior and movement of stocks in the Brazilian market by integrating historical price series and textual data extracted from sentiments in X old Twitter messages and news collected from Google News. The analysis used natural language processing techniques for sentiment analysis, enabling an efficient fusion between numerical and textual information. Experiments were carried out with the assets PETR4, VALE3, BBDC4, and ITUB4, applying the Long Short Term Memory, Deep Neural Network, and Linear Regression models to predict the behavior of these assets. The results indicated that the LSTM models, especially Model 2, presented the best performance in terms of predictive capacity, with the lowest values of RMSE 0.0171 and high values of coefficint of determination ranging from 0.9707 to 0.9873. The study concludes that integrating numerical and textual data, combined with deep learning techniques, offers a promising approach to stock market forecasting, increasing forecasting gains.