The Brazilian capital market presents unique challenges for portfolio optimization due to its volatility and information asymmetries. This paper proposes a Deep Reinforcement Learning (DRL) framework that integrates fundamental indicators, Portuguese-language news sentiment (via Gemini Pro), and market data (prices, volume). Five DRL algorithms (A2C, PPO, DDPG, TD3, SAC) were trained and evaluated across three feature scenarios using performance metrics such as Sharpe Ratio, Annual Return, and Maximum Drawdown. News sentiment classification and entity extraction were performed using Gemini Pro LLM. Statistical validation over 44 executions, including Shapiro-Wilk and Kruskal-Wallis tests, showed no significant differences among DRL methods. However, all DRL approaches outperformed the Ibovespa index and uniform Buy-and-Hold benchmark, highlighting the value of combining DRL with localized and diverse data sources for portfolio optimization in emerging markets.

In this work, we explore machine learning to ob-tain sentiment indicators from financial market text messages collected from Twitter and news in Portuguese. A statistical analysis was carried out with Sperman’s correlation coefficient between sentiment indicators and actual variables in the Brazilian financial market. For sentiment analysis, the MaxEnt and CNN models provided F1-scores of 85% and 96%, for tweets and news, respectively. A result shown a moderate correlation (Cohen’s scale) between some variables such as, amount of tweets and sentiment of the news; market variation and tweet sentiment; amount of retweets published and sentiment of the news; trading volume and tweet sentiment. Moreover, a very large correlation was identified between the amount of negative tweets and retweets, leading to the belief that pessimism is often propagated.