A Prediction Divergence Criterion for Model Selection