Standard forecast uncertainty bands broke down across the COVID shock. An adaptive method brought coverage back toward its promise, though still short of it, for the models it covers.
What this shows: The chart plots Nominal, Split-conformal (empirical) and Adaptive conformal (valid under distribution shift) on a vertical axis of Coverage (%).
Conditions and method
Status: descriptive (the split-conformal series loses its guarantee across the structural break and measures that breakdown; the adaptive series stays valid under the shift).
Calibration ends in 2019, before COVID, so the split-conformal guarantee is void for a test window that spans the COVID structural break, and the split-conformal bars are not a valid 90% interval.
The adaptive series updates as each quarter arrives and is valid under distribution shift, so it is the coverage to compare across the break.