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

Area of Science:

• Agricultural Science
• Animal Science
• Data Science

Background:

• Forecasting individual cattle live weight in grazing systems is difficult due to irregular weighing and seasonal environmental changes.
• Limited evidence exists for live-weight forecasting across multiple periods in commercial grazing settings.

Purpose of the Study:

• To develop a machine learning (ML) framework for forecasting individual cattle live weight.
• To evaluate the framework's performance using automated observations and climatic predictors across various forecasting periods and aggregation methods.

Main Methods:

• Developed an ML framework incorporating demographic variables, historical live weights, and lagged climatic predictors.
• Compared monthly, weekly, and rolling-window aggregations for 1, 2, and 3-month forecasting periods.
• Utilized Gradient Boosting for model training and performance evaluation.

Main Results:

• Gradient Boosting achieved high predictive accuracy, with R² values of 0.950, 0.935, and 0.902 for 1, 2, and 3-month forecasts, respectively.
• Monthly aggregation outperformed other methods in entropy retention, variance preservation, and forecasting performance.
• Animal age and historical live weight were key predictors, complemented by lagged rainfall and temperature for medium-term forecasts.

Conclusions:

• The developed ML framework effectively forecasts individual cattle live weight in managed grazing systems.
• Automated livestock monitoring, climate data, and ML integration support accurate live-weight predictions.
• Monthly data aggregation is optimal for enhancing forecasting accuracy in this context.