Design of data based models for nonlinear systems using universal approximators has become a standard issue for which several tool boxes exist. The validation of such models is usually done using data of the same range and distribution as the identification data. This tends to hide the fact that the choice of the basis function used for the approximation is decisive in terms of model quality, and in particular of its extrapolation qualities. To this end, this paper compares a subspace identification procedure for a class of nonlinear systems with standard neural networks. As the results shown confirm, the subspace identification procedure in its simple form is not able to yield consistent estimates, but after a suitable robustification proves clearly superior to the ANN both in terms of performance (in terms of VAF) and of complexity (in terms of the number of parameters).