Statistical Theory of Overtraining - Is Cross-Validation Asymptotically Effective?