Conformal Prediction

Conformal prediction uses examples held out for calibration to turn a model’s scores into prediction sets or intervals with finite-sample coverage guarantees, usually when calibration and future data are exchangeable. Split and full conformal methods are common foundations; variants such as conformalized quantile regression, Mondrian calibration, weighted conformal prediction, and adaptive conformal inference target efficiency, subgroup coverage, distribution shift, or streaming data. These methods are used for classification and regression as well as time-series forecasting, structured outputs, language-model abstention, and safety-critical decisions. A central challenge is that marginal coverage can hide failures for particular groups or difficult inputs, while methods that improve local reliability or handle shift often need more data or stronger assumptions and can produce larger prediction sets.