An idealised setup Suppose we have an NIR calibration whose prediction errors have a bias b and variance s, both of these parameters being unknown. Prediction error here means the difference between the prediction from the calibration and the corresponding reference value. The variance quantifies the random errors, the bias measures the systematic error. The difference between the two is that if we repeatedly measured the same sample, both by NIR and by the reference method, and averaged the prediction errors, the random error would tend to zero, but the bias would remain. If the bias is the same for all samples, which for the moment we assume to be true, then a plot of reference versus predicted will show an offset b from the ideal 45° line. In this situation the average squared prediction error (MSEP) will be variance plus squared bias