Previsão de tendência de preços com random forest: aplicação a cinco ações de diferentes setores da B3

Abstract

Forecasting trends in the financial market is a complex challenge that involves multiple variables and temporal patterns. This study uses the Random Forest machine learning technique to predict stock price direction based on technical indicators (RSI, Stochastic Oscillator, Williams %R, MACD, PROC, and OBV) extracted from smoothed prices. Five companies listed on the B3 were analyzed: Ambev S.A., Itaú Unibanco S.A., Magazine Luiza S/A, Petróleo Brasileiro S.A., and Vale S.A., with different validation methodologies, including Time Series Cross-Validation (TSCV) Expanding Window and TSCV Sliding Window. The results show that the choice of validation methodology significantly impacts the model's accuracy. The random split approach yields inflated accuracies due to the risk of overfitting, while more robust methods, such as TSCV Expanding Window and Sliding Window, provide lower and more realistic estimates. Among the companies analyzed, Magazine Luiza S/A achieved the best performance with an accuracy of 77.94% for the 90-day prediction window, while the worst accuracy was 48.80% for the 1-day period with Itaú Unibanco S.A. It was observed that as the temporal window increased, the accuracies also grew. The results demonstrate the power of Random Forest in capturing price trends, outperforming traditional methods and offering a reliable alternative in a high-volatility environment. Throughout the article, the implications of these findings are discussed, highlighting both practical applications for investors and directions for future research.