In 1996 the National Physical Laboratory (NPL) in the UK conducted a project to determine the agreement of colorimetric measurements between different European laboratories. The results suggested significant measurement differences between instruments even though they were of similar design and built by the same manufacturer.Modeling random and systematic spectrophotometric errors and the use of multiple regression analysis improved successfully the agreement between instruments since the early 1980s. This was in particular the case if they were regressed on each wavelength. Recently a new model based on band pass error was developed. It was said to outperform other models inherent of more errors in the equation. This paper shows that this was hardly the case.Furthermore, it was evident that all models performed best when training and testing samples were the same. This showed clearly the dependency of the model on the physical properties of the samples used for training. Using other materials for testing resulted in just little improvement.It was then of interest to determine which and how many samples were needed for training the model while maintaining a good performance. A method was found to reduce the number of training samples from a larger population of the Munsell Color Book. This resulted in a training set of 20 samples compared to 245 colour samples for modeling the correlation between two instruments with similar results.