User models for email filtering should be developed from appropriate training and test sets. A k-fold cross-validation is commonly presented in the literature as a method of mixing old and new messages to produce these data sets. We show that this results in overly optimistic estimates of the email filter’s accuracy in classifying future messages because the test set has a higher probability of containing messages that are similar to those in the training set. We propose the k-fold chronological cross-validation method that preserves the chronology of the email messages in the test set.