This paper describes a data-based approach to the identification and estimation of non-linear dynamic systems which exploits the concept of a state dependent parameter (SDP) model structure. The major attractive features of the proposed approach are: (1) the initial non-parametric identification of the non-linear system structure using an SDP algorithm based on recursive fixed interval smoothing; (2) a compact parameterization of this initially identified model structure via a linear wavelet functional approximation; and (3) final optimized model structure selection using the predicted residual sums of squares (PRESS) statistic, prior to final parametric optimization using this optimized, parsimonious structure. Two simulation examples are used to demonstrate the proposed approach.