5103c3584b063c431bd1268e9b5e76fb-AuthorFeedback.pdf

This is obviously true and is noted in Section 5 of the manuscript.8 As for locally adaptive split conformal prediction, we argue in our paper that scaling residuals by their variance is22 alsosuboptimal, evenifitslimitations arelesssevere. In the ideal setting where the conditional distribution is known,CQR would construct exact prediction intervals28 reflecting the intrinsic predictive uncertainty,while achieving any desiredcoveragelevel.