DOES HAVING MORE VARIABLES TRANSLATE TO BETTER PREDICTIONS IN TROPICAL GRAZING SYSTEMS? A CASE STUDY CONSIDERING ENVIRONMENTAL AND OPERATIONAL DATA

Machine learning models have been increasingly applied to precision livestock farming; however, their performance under real commercial production conditions remains poorly understood. This study evaluated the contribution of animal, environmental, and management variables to predicting final body weight and average daily gain (ADG) of Nellore cattle in a tropical grazing system using real-world field data. Data from 26 animals managed under continuous grazing for 307 days were used to develop predictive models based on LASSO, Random Forest, and CatBoost algorithms. Initial body weight showed strong predictive relevance, with the univariate model achieving high performance (R² = 0.977). The inclusion of environmental and management variables produced only marginal improvements, even for the best-performing model (CatBoost, R² = 0.9816). In contrast, removing initial body weight substantially reduced predictive performance (R² = 0.37–0.47). Models developed for ADG prediction also showed low predictive capacity (R² = 0.32–0.38). The results suggest that predictive performance was more strongly associated with data structure and information content than with algorithm complexity. Furthermore, the study highlights important limitations related to low temporal resolution, irregular sampling intervals, and aggregated environmental data under real production conditions. These findings reinforce the importance of data-centric approaches and ecologically valid datasets for the practical application of artificial intelligence in tropical livestock systems.