{"id":"arxiv/2606.31600","url":"https://arxiv.org/abs/2606.31600","tags":["math.ST","cs.LG","stat.ML"],"score":0,"title":"On Optimal Data Splitting for Split Conformal Prediction","domain":"math.ST","source":"arxiv","authors":["Sayan Das","Bahram Yaghooti","Todd A. Kuffner","Soumendra N. Lahiri"],"reviews":0,"subject":"math.ST","summary":"Conformal prediction and its variants, including the split conformal prediction, provide a distribution-free framework for uncertainty quantification by constructing prediction intervals or sets with finite-sample coverage guarantees. The statistical efficiency of these intervals depends critically on how the data are split into training and calibration samples. Despite its practical importance, a principled characterization of the training-calibration split that minimizes prediction interval length while maintaining coverage has remained largely unresolved. In this paper, we develop a…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-06-30","reactions":{}}