{"id":"arxiv/2611.20052","url":"https://arxiv.org/abs/2611.20052","tags":["math.ST","stat.ME","stat.ML"],"score":0,"title":"Weighted conformal prediction under covariate shift: process limits, estimated weights and efficiency","domain":"math.ST","source":"arxiv","authors":["Mehrdad Pournaderi"],"reviews":0,"subject":"math.ST","summary":"Weighted split conformal prediction handles covariate shift by reweighting calibration scores with the likelihood ratio $v=dQ_X/dP_X$; its training-conditional coverage is a functional of the self-normalized weighted empirical distribution function $\\hat F_m$, and we study the process $\\sqrt m(\\hat F_m-F_Q)$. It converges in $\\ell^\\infty$ to the Gaussian process $W(G(t))-F_Q(t)W(K^2)$, with $G(t)=\\mathbb{E}_Q[v\\mathbf{1}\\{S\\le t\\}]$ and $K^2=1+χ^2(Q\\|P)$, if and only if $χ^2(Q\\|P)<\\infty$ (for non-degenerate scores); $\\sup v$ plays no role. The limit gives the joint law of the coverage…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-11-15","reactions":{}}