I think when your model consistently gives results indiscernable from random noise in its strongest application (cf. blog.okcupid.com), you can safely conclude that it's of no more value than a model selected at random from all possible models of similar complexity.
If you're interested in this problem, this is a great reference: http://plato.stanford.edu/entries/pseudo-science/