Spectral Adaptive Conformal Prediction for Structured Non-Exchangeable Data
Researchers have introduced Spectral Adaptive Conformal Prediction, a novel method designed to provide reliable prediction intervals for time-indexed datasets that are not exchangeable. This technique utilizes local spectral similarity to form weighted conformal quantiles and updates the miscoverage level online, adapting to changing uncertainty over time. The method aims to improve upon existing spectral weighting approaches and has been demonstrated through simulations and real-world U.S. data examples. AI