This article illustrates the problem with over-fitting a model even when some data is withheld for testing. This is a trap that one can fall into when using training and testing sets.