The purpose of this contribution is to point out and exploit the kinship between identification of linear dynamic systems and classical curve fitting. For curve fitting we discuss both global and local parametric methods as well as non-parametric ones, such as local polynomial methods. We view system identification as the estimation of the frequency function curve. The empirical transfer function estimate is taken as the "observations" of this curve. In particular we discuss how this could be done for multi-variable systems. Local and non-parametric curve fitting methods lead to variants of classical spectral analysis, while the prediction error/maximum likelihood framework corresponds to global parametric methods. The role of the noise model in dynamic models is also illuminated from this perspective.