When Does a Frozen Time-Series Foundation Model's Uncertainty Fail, and What Actually Fixes It?

Abstract

Open time-series foundation models (TSFMs) are deployed zero-shot with their native prediction intervals read as calibrated. Using frozen Chronos models on public benchmarks we show: (1) native interval coverage collapses under regime shift (86% to 72% at nominal 80%; 82% to 50% at true 90%), driven by shift, not horizon; (2) no static post-hoc conformal wrapper repairs conditional coverage, while an oracle does, proving it is fixable miscalibration; (3) online conformal (Adaptive Conformal Inference) is the deployable fix, substantially recovering coverage under drift; and (4) pretraining leakage does not inflate coverage (unlike accuracy), tested on Chronos's own documented train/zero-shot split. Under drift, a frozen TSFM's interval can be less trustworthy than a seasonal-naive baseline. Fully reproducible on a single laptop with public data at zero API cost. Code and results are included in this record.