Weighted conformal prediction improves coverage under shifting data
Sharp training-conditional coverage for conformal prediction under covariate shift
Machine Learning
Summary
When data changes between training and testing, it can be hard to trust predictions. The authors study a method that adjusts predictions to maintain correct coverage even when the data shifts. They provide clear mathematical guarantees about how well this adjustment works and identify the key factors influencing its accuracy. Their work also compares different techniques and offers practical ways to use estimated data distributions, making predictions more reliable.
What this means in practice
- •For machine learning engineers: Get reliable uncertainty estimates for models deployed in environments where input data distributions differ from training data.
- •For data scientists in finance: Produce valid prediction sets for financial risk models when market conditions change the underlying data features.
Authors
Mehrdad Pournaderi
Abstract
Weighted split conformal prediction reweights calibration scores by the likelihood ratio between the test and training covariate distributions and guarantees marginal coverage under covariate shift. We study its coverage conditional on the calibration data. An elementary argument, based on a single concentration inequality at a fixed population quantile, gives explicit training-conditional bounds without unspecified constants, and shows that the relevant scale is not the supremum of the likelihood ratio but a variance proxy built from the chi-squared divergence of the shift and from the average of the ratio over the part of the test population, of probability equal to the miscoverage level, where it is largest. A two-point lower bound shows that the root-m rate and the chi-squared contribution are intrinsic to the shift. Run at an explicitly inflated level, the weighted quantile becomes a deterministic PAC prediction set. We compare it with randomized rejection sampling and with importance-weighted learn-then-test and, through a certified choice of a clipping level for the likelihood ratio, map the regime in which each gives the narrower valid set. The analysis extends to estimated likelihood ratios and to tail functionals estimated from an unlabeled source sample, which yields a fully finite-sample certificate.