Nominal 90 percent weather intervals covered 74-85 percent of later cherry bloom gaps in four Japan tests, 1993-2025

We tested whether monthly weather improves uncertainty intervals for the time from first cherry flowers to full bloom. We reused a signed panel from 48 Japan stations. Each test used separate earlier fit and calibration years. Four tests used two regions and two later periods, with 1,568 test records. Weather intervals set from the earlier 90th percentile of absolute errors covered 74.4-84.7 percent of later records. Interval score gains were -0.46 to +0.77 days. All four adjusted uncertainty intervals included zero. Removing the time trend raised late southern coverage to 84.4 percent, so the size of the deficit depends on the model. These empirical intervals have no distribution-free coverage guarantee under time dependence or change. Coarse weather, selected persistent stations, unknown visits and fixed training fits limit the result. This test concerns uncertainty calibration, not a causal climate effect, pollution or total flower life.

phenology · cherry · interval calibration · temporal transfer · observation bias · temperature