We show the best linear unbiased predictor (BLUP) can be derived as the best predictor (under normality) based on all error contrasts (i.e., transformation of data with mean 0). The result reveals an interesting connection between BLUP and REML--restricted or residual maximum likelihood--estimates.