Tibshirani, Barber & Ramdas reframe conformal prediction via hypothesis testing, deriving universality and optimality

_onionesque · x · 2026-09-22

A new arXiv paper by Ryan J. Tibshirani, Rina Foygel Barber and Aaditya Ramdas recasts conformal prediction as inverting a permutation test for exchangeability of the joint distribution of n+1 samples. This hypothesis-testing lens reproduces known universality and impossibility results via classical theory (Neyman, Lehmann, Scheffé, Kraft, Le Cam) and yields a Neyman–Pearson-based optimality result for prediction sets under any joint distribution and sample size.

Original post →

More from Research

Research channel →