Beyond Marginal Validity: Finite-Sample Guarantees for Localized Conformal Prediction
The paper gives finite-sample local guarantees for RLCP’s realized neighborhoods.
It proves high-probability bounds for both conditional-coverage error and prediction-set length error relative to an oracle. The result separates localization bias, stated as `O(h^β)`, from a calibration term that shrinks with more calibration data. It also extends the analysis to learned scores, including pivotal-score targets such as conformalized quantile regression. ArXiv · AI/CL/LG's note
It proves high-probability bounds for both conditional-coverage error and prediction-set length error relative to an oracle. The result separates localization bias, stated as `O(h^β)`, from a calibration term that shrinks with more calibration data. It also extends the analysis to learned scores, including pivotal-score targets such as conformalized quantile regression. ArXiv · AI/CL/LG's note
score 4