Bayesian comparison of models for precision feeding and management in growing-finishing pigs

Precision feeding and management of growing-finishing pigs typically require mathe-matical models to forecast individual pig performance from past data. The current ap-proaches, namely double exponential smoothing (DES) and dynamic linear regression are likely to have some limitations in their applicability since they: (1) assume that responses can be forecasted linearly, which only holds in the short-term, and (2) often take insuffi-cient account of uncertainty and correlations in the estimated traits. We developed and evaluated alternative approaches to forecasting individual growth or intake responses based on nonlinear models (allometric, monomolecular, rational) and Bayesian method-ology to fit models to the data and generate probabilistic forecasts. We applied these ap-proaches to individual data from two distinct pig populations, to parameterise the models (fitting based on a training dataset) and forecast performance (forecast horizons: 1-30 d tested on a validation dataset). We found that good fitting did not guarantee accurate forecasting, which is quantitatively relevant in the medium-to-long term. Forecasts from nonlinear models were more accurate compared to those from benchmark linear models, with the allometric model being more accurate for most pigs across considered forecast horizons. While DES was the best model at fitting, it was also the least accurate at fore -casting for all forecast horizons. These results enhance the understanding of how under-lying biological growth responses could be approximated using straightforward mathematical relationships. The approach could be utilised to formulate optimised feeding strategies and inform management decisions, including pen allocation or end-weight prediction. (c) 2021 Published by Elsevier Ltd on behalf of IAgrE.

Automated summary

The study found that using nonlinear models for forecasting individual pig growth or intake responses is more accurate than using benchmark linear models, with the allometric model performing the best across most forecast horizons. While double exponential smoothing was the best model at fitting, it was also the least accurate at forecasting for all forecast horizons.