Um modelo preditivo com base na integração de dados numéricos e textuais: um estudo de caso no mercado acionário brasileiro

Abstract

The analysis of movements and prices in the Brazilian stock market has been widely studied, with a recent increase in the use of Artificial Intelligence for this purpose. Traditionally, predictive approaches rely on historical numerical data, with an emphasis on graphical analysis. However, these techniques have not fully explored the potential of fundamental data extracted from technical reports and financial statements, nor have they taken advantage of the vast amount of real-time information available through social media and news portals. This study aimed to identify the most effective approach to improving the accuracy of stock price predictions by integrating numerical and textual data, applied to a set of assets in the Brazilian stock market. Various deep learning techniques and models were employed, and the literature review revealed gaps in integrating heterogeneous data. To address these limitations, an approach was proposed that combines numerical and textual data, assessing the impact of this integration on stock price and movement predictions. The textual data includes financial statement information, posts on X (formerly Twitter), and financial and economic news published online. The numerical data consists of historical stock price and volume series, macroeconomic variables, and the Google Trends search index. The proposed model allows for an evaluation of advancements in the processing and integration of numerical and textual data to identify stock price movements in the Brazilian market. Studies were conducted to explore the behavior of numerical and textual data. Additionally, experiments implementing the proposed approach demonstrated a percentage gain in prediction accuracy compared to purely numerical analysis. The results revealed that the inclusion of tweets, news (Google News), and technical indicators, along with stock price and volume data, improved forecasting accuracy. When comparing the tested models, the LSTM outperformed the DNN. The collected RMSE values were: PETR4 (0.0114; 0.0111; 0.0210), VALE3 (0.0106; 0.0128; 0.0452), BBDC4 (0.0119; 0.0112; 0.0234), and ITUB4 (0.0117; 0.0119). It is concluded that the integration of heterogeneous data can significantly enhance stock price predictions, contributing to the development of more effective strategies in the financial market.