Two-Timescale Adaptive Conformal Prediction Under Distribution Shift

While Adaptive Conformal Inference (ACI) guarantees coverage under distribution shift, it often achieves this through a "pyrrhic victory": inflating prediction intervals until they are valid but carry little practical information. Our research identifies the root cause as the conflation of adaptation timescales. By introducing Two-Timescale (TTS) conformal prediction, we decouple immediate calibration from long-term model refitting. The result is a method that maintains safety guarantees while producing significantly tighter intervals—up to 83% shorter in real-world climate applications—without requiring complex state models.