Isotonic Conformal Prediction slashes calibration costs while preserving coverage

A new paper from Daniel Bensimon, Sean Xiang Yu, Eric D. Kolaczyk, and Archer Y. Yang introduces Isotonic Conformal Prediction (ICP), a framework designed to address a key limitation in uncertainty quantification: the computational expense of Self-Calibrating Conformal Prediction (SC-CP) for continuous outcomes. SC-CP requires refitting its calibrator for every candidate outcome, making it prohibitively slow. ICP decouples calibration from prediction-set construction by fitting a single isotonic recalibration map, then constructs prediction intervals within strata of similar recalibrated predictions.

Within ICP, the authors develop two procedures. Split Isotonic Conformal Prediction (SICP) attains prediction-conditional validity in finite samples and self-calibration asymptotically, at the computational cost of standard split conformal prediction. Transductive Isotonic Conformal Prediction (TICP) achieves both objectives exactly in finite samples through a per-test-point inner loop that avoids refitting the isotonic calibrator. On synthetic heteroscedastic regression and a real-world healthcare-utilization dataset, both procedures match the coverage of SC-CP at substantially lower computational cost, making them practical for large-scale continuous prediction tasks.