Revisiting Boi Gordo Index Futures: Long-Run Daily Data, Structural Breaks, and a Comparative Evaluation of Classical and Machine Learning Time-Series Models

We study one of the world’s largest cattle markets by revisiting and extending previous work on the forecasting of Brazil’s Boi Gordo Index (BGI). Using an updated daily dataset (July 2006–September 2025, inflation-adjusted), we evaluate classical and machine learning (ML) approaches for price prediction. Methods include Exponential Smoothing (Simple, Holt, and Holt–Winters), ARMA/ARIMA/SARIMA, GARMA variants, GARCH, Theta, Prophet, and XGBoost; models are compared under a strictly chronological 90/10 holdout (~476 test days) using RMSE, MAE, and MSE, with the AIC guiding within-family selection. Results show that, for the full out-of-sample window, GARMA delivers the best overall accuracy, with ARMA and Holt–Winters close behind, while Prophet and XGBoost perform comparatively worse in this volatile setting. Performance is horizon-dependent: in the first 180 test days, prior to the late-2024 level shift, Holt attains the lowest RMSE/MSE, and XGBoost achieves the lowest MAE. No method anticipates the October–November 2024 exogenous jump and subsequent correction, highlighting the difficulty of structural breaks and the need for timely re-specification. We conclude that GARMA is a robust default for long, turbulent windows, whereas smoothing and ML methods can be competitive on shorter horizons. These findings inform risk measurement and risk mitigation strategies in Brazil’s cattle futures market.