Abstract
Conformal prediction is a technique for constructing prediction intervalsthat attain valid coverage in finite samples, without making distributionalassumptions. Despite this appeal, existing conformal methods can beunnecessarily conservative because they form intervals of constant or weaklyvarying length across the input space. In this paper we propose a new methodthat is fully adaptive to heteroscedasticity. It combines conformal predictionwith classical quantile regression, inheriting the advantages of both. Weestablish a theoretical guarantee of valid coverage, supplemented by extensiveexperiments on popular regression datasets. We compare the efficiency ofconformalized quantile regression to other conformal methods, showing that ourmethod tends to produce shorter intervals.