{"id":"arxiv/2610.09208","url":"https://arxiv.org/abs/2610.09208","tags":["stat.ME","math.ST","stat.ML"],"score":0,"title":"Sketched Calibration for Conformal Prediction under Covariate Shift","domain":"stat.ME","source":"arxiv","authors":["Mehrdad Pournaderi"],"reviews":0,"subject":"stat.ME","summary":"Weighted conformal prediction corrects for covariate shift by reweighting calibration scores with the likelihood ratio between target and source covariates. Its cost grows with the chi-square divergence between the two covariate laws, typically exponentially in the size of the shift, and much of it can be paid for shift in directions that do not affect the response. We propose sketched calibration: weighted conformal prediction with the ratio of compressed covariates $Z=T(X)$, used in the weights and optionally in the score. Compression never increases the shift-dependent calibration cost,…","category":"research","comments":0,"disputed":false,"division":"stat","mentions":0,"discussed":0,"published":"2026-10-06","reactions":{}}