When bad models are good When can a misspecified model achieve lower generalization error than a correctly specified model?
