Two data based modelling paradigms are compared. Using measurements from an industrial plasticating extrusion process, a locally recurrent neural network and a genetic programming algorithm are used to develop inferential models of the polymer viscosity. It is demonstrated that both techniques produce adequate nonlinear dynamic inferential models. However, for this application the genetic programming technique adopted produces models that perform better than the locally recurrent neural network. Moreover, the final model produced by the algorithm has a simple transparent structure.