Precision feeding of growing-finishing pigs require a mathematical model to forecast individual responses to different nutrient supplies. The two most common models assume linear growth trends: (1) double exponential smoothing, a type of moving average of past observations; and (2) dynamic linear regression, where the regression parameters can vary in time. The statistical approaches that are currently used to fit these models (i.e. maximum likelihood) give limited information on parameter uncertainty and correlations, which could limit the usefulness of the forecasts. Here, we developed and evaluated alternative models of growth response to nutrient supply that are non-linear (allometric, monomolecular and rational). We utilised a Bayesian inference approach to account for the uncertainty and correlations in parameter estimates, which outputs distributions rather than point estimates. We fitted the current linear and the alternative non-linear models to individual pig data from two distinct populations for different estimation scenarios: (1) in-sample (using all available individual data); and (2) out-of-sample (using the training subset to infer parameters and a testing subset to validate the forecasts). We found that: (1) all models gave similar in-sample goodness-of-fit; (2) forecast of future growth differed between the models, with the allometric model generating the most reliable forecasts across individual pigs, especially when forecasts were made over more than 1-2 days ahead. The results of this study could help to forecast individual nutrient requirements through a robust estimation of their average requirements and a suitable estimation of their uncertainty given the known past performance of the animal. The approach could also be utilised to inform management strategies, such as pen allocation and slaughter weight prediction.