Use of PRESS residuals in dynamic system identification

Cross-validation, whereby the data is split into an estimation set and a prediction set, is often used for model structure selection in dynamic system identification. A related approach that simulates cross-validation without requiring two sets of data is to compute the weighted residuals of the least-squares estimated model that represent the true prediction errors. The sum of the squared true prediction errors is defined as the PRESS. This paper will show that the computation of the PRESS is straightforward when using an orthogonal decomposition algorithm to obtain the least squares model parameter estimates. A benchmark example of Ljung is used to illustrate how the PRESS is applied for model structure selection.