A key issue in statistics and machine learning is to automatically select the "right" model complexity, e.g., the number of neighbors to be averaged over in k nearest neighbor () regression or the polynomial degree in regression with polynomials. We suggest a novel principle-the Loss Rank...