In order to transfer a calibration model, techniques such as piecewise direct standardization (PDS) or direct standardization (DS) are used. This implies that the same subset of samples must be measured in both situations. However, when a model is transferred from an off‐line application to an on‐line application, it is not possible to remeasure the same subset of samples. In this paper, however, a method is presented which can transfer a multivariate calibration model from an off‐line application to an on‐line application without remeasuring the same subset. For this purpose a calibration model has been calculated with a subset of spectral variables. The variables, which are sensitive to the differences between the off‐line and the on‐line application, are rejected. The method is compared with the calculation of a hybrid calibration model in which off‐line and on‐line spectral data are combined into one calibration set. The methods are applied to a calibration model which is used to predict the shrinkage after heat treatment of poly(ethylene terephthalate) (PET) yarns from the Raman spectra measured both off‐line and on‐line. The Raman spectra of a calibration set have been measured off‐line and a smaller set has been measured on‐line during the production of PET yarns. It has been found that it is possible to make a universal calibration model for off‐line and on‐line predictions of PET yarn shrinkage by Raman spectroscopy. Copyright © 1999 John Wiley & Sons, Ltd.