February temperature improved Liestal bloom interval scores in both test periods from 1988 to 2025

We tested whether February mean temperature improves 80% bloom prediction intervals beyond a time trend. We used fixed Liestal and Washington bloom records and nearby GHCN-Daily stations. Models used only earlier records. In 1988-2006 and 2007-2025, interval score changes at Liestal were -13.06 and -5.84 days. Both four-test block intervals were below zero. Mean widths fell by 6 to 7 days. Washington changes were -3.97 and -3.93, with main intervals that included zero. Temperature interval coverage ranged from 73.7% to 94.1% across the four tests. Thus a score gain did not establish correct 80% coverage. Shorter training and recent-error intervals gave broad support at Liestal, but not every sensitivity interval excluded zero. Small selected samples, changing observations, station mismatch, and approximate bootstrap intervals limit inference. This is a test of forecast uncertainty for ecological timing, not a causal climate or pollution test. The package includes code, source hashes, annual score evidence, all estimates, a figure, and reproduction checks.

phenology · cherry bloom · prediction intervals · temperature · observation bias · temporal validation