Machine learning models to predict hot carcass weight of Angus feedlot cattle at induction in commercial conditions

Ning Han, Thomas F.A. Bishop, Luciano A. Gonzalez,
Machine learning models to predict hot carcass weight of Angus feedlot cattle at induction in commercial conditions,
Smart Agricultural Technology,
Volume 12,
2025,
101496,
ISSN 2772-3755,
https://doi.org/10.1016/j.atech.2025.101496.
(https://www.sciencedirect.com/science/article/pii/S2772375525007270)
Abstract: In modern feedlot management, the ability to accurately predict hot carcass weight at induction is critical for optimising operations and profitability. This study utilised a dataset of 40,310 records of Angus cattle from a commercial feedlot in Central Victoria (Australia) to forecast hot carcass weight from induction data including induction weight, month of induction, sex, and origin, and four variables collected throughout the production cycle: days on feed, disease (yes or no), and weather-related variables (number of days with rainfall above 5 mm and number of days with temperatures exceeding 35 °C between induction and slaughter dates). Four statistical and machine learning models were used: linear regression with stepwise variable selection, support vector machines, random forest, and cubist model. Cubist model presented the best precision being slightly better with weather data (R2 = 0.70, MAE = 20.02 kg, RSME = 27.94 kg) compared to no weather data (R2 = 0.67, MAE=21.00 kg, RSME=29.16 kg) but Cubist model showed the lowest accuracy as measured by mean bias (-0.56 kg) and slope between observed and predicted values (0.92 ± 0.006). In addition, random forest showed slightly lower precision (R2 = 0.69) than cubist model in all scenarios but greater accuracy as measured by mean bias (-0.41), intercept (-2.39 ± 2.26) and slope (1.01 ± 0.007), Linear model with stepwise variable selection and support vector machines shower lower precision (R2 = 0.64) but greater accuracy (mean bias = -0.47; intercept = 2.81 ± 2.50; slope = 0.99 ± 0.007) than cubist model. Further research is required to improve the accuracy of predicting hot carcass weight potentially including more predictors such as genetic information, diet composition, or data collected throughout the production cycle, particularly with technologies collecting data in real-time in the pen. However, the present research laid the groundwork for machine learning technologies to boost operational efficiencies and decision-making processes in the feedlot industry.
Keywords: Machine learning; Artificial intelligence; Feedlot management; Hot carcass weight; Model performance