Estimation focus in system identification: prefiltering, noise models, and prediction

We review some features related to the use of prefiltering data for identification. In addition to the well known interplay between noise models and prefilters we discuss how to find a compromise between the need for a noise model, at the same time as having control of the approximation properties of the model. The statistical paradigm tells us to use high order models so that the bias distribution aspect of the prefilter can be neglected. For real data this may however be infeasible. The discussion is illustrated with both simulated and real data.