Adaptive Conformal Inference

Overview

Adaptive conformal inference studies how to update prediction thresholds as outcomes arrive, helping prediction intervals or sets maintain target error rates when data drift or are selected online. Most methods use coverage feedback to adjust a quantile or significance level; variants add sliding windows, recency or density weighting, regime-specific calibration, and score correction. Researchers assess both coverage and efficiency, such as interval width, in settings including time-series forecasting, anomaly detection, and safety-critical control, alongside analyses using online learning and game theory. The central challenge is balancing narrow predictions and responsive adaptation against reliable guarantees: exchangeability supports finite-sample marginal validity for classical conformal methods, while distribution shift and sparse calibration data can cause undercoverage, and stronger conditional guarantees often require tradeoffs or extra assumptions.

Latest papers

CardsList

32 papers4 papers in the last four weeks, level with the four weeks before. 0.0% of all new papers.

  1. Adaptive Conformal Inference through the Lens of Blackwell Approachability

    Oct 17, 2025Guillaume Principato, Gilles StoltzBlackwell ApproachabilityConformal Prediction

  2. Scalable and adaptive prediction bands with kernel sum-of-squares

    May 27, 2025Louis Allain, S{é}bastien da Veiga, Brian Staberhilbert-schmidt independence criterionConformal Prediction