Otimização de portfólio de ativos do mercado financeiro brasileiro: integrando notícias e indicadores fundamentalistas com aprendizado por reforço profundo

Abstract

This study investigates the impact of incorporating fundamental and sentiment-based indicators, extracted from Portuguese news, into deep reinforcement learning (DRL) algorithms for portfolio optimization in the Brazilian financial market. The research involves collecting news articles, historical data, and financial indicators of assets, with sentiment and key entities extracted from the news using the Gemini Pro Large Language Model. To refine sentiment indicators, entity- and topic-based filtering techniques are applied to reduce informational noise. Statistical analyses using correlation coefficients show that entity-based filtering enhances the relationship between sentiment indicators and daily asset returns. Subsequently, sentiment and fundamental indicators are integrated into five DRL algorithms, tested across four distinct scenarios: (1) prices only, (2) prices and sentiment, (3) prices and fundamental indicators, and (4) all combined. A total of 44 samples were generated. Although the Kruskal-Wallis test did not reveal statistically significant differences between the scenarios, all DRL-based models outperformed baseline strategies such as Buy and Hold and the Ibovespa index in terms of Sharpe Ratio, indicating higher returns with better risk control. These findings suggest that the use of DRL algorithms can contribute to more effective risk management in portfolios composed of Brazilian financial market assets.