Multi-Horizon Herd-Based Cattle Live-Weight Forecasting Using Irregular Automated Weighing Data

Forecasting individual cattle live weight in herd-managed grazing systems remains challenging because automated weighing observations are irregular and environmental conditions vary seasonally. Evidence remains limited for live-weight forecasting across multiple forecasting periods under commercial grazing conditions. This study developed a machine learning (ML) framework for forecasting the live weight of individual cattle using automated observations from a managed herd. The framework incorporated demographic variables, historical live weights, and climatic lag predictors. The framework compared monthly, weekly, and rolling-window aggregations across forecasting periods of 1, 2, and 3 months. Quality control retained 494 of 1140 cattle (43.3%), yielding 4069 monthly aggregated records. The resulting dataset was more suitable for modelling cattle with regular voluntary weighing records. The respective forecasting datasets contained 3048, 2558, and 2068 records for one-, two-, and three-month periods. Gradient Boosting achieved the strongest testing performance. The corresponding coefficients of determination (R2) were 0.950, 0.935, and 0.902. Monthly aggregation achieved higher entropy retention, greater variance preservation, and stronger forecasting performance than alternative aggregation approaches. Feature-importance analysis identified animal age and historical live weight as the most important predictors across all forecasting periods. Lagged rainfall and temperature variables provided complementary predictive information for medium-term forecasting. The findings demonstrate that automated livestock monitoring, climatic information, and ML can support accurate live-weight forecasting. The framework produced forecasts across multiple periods for individual cattle in herd-managed grazing systems.