Learned first-leaf updates reduced bloom-date error in three of four US tests against full updates in 1980-2008

We tested a first-bloom date update whose size was learned from earlier first-leaf and bloom records. Separate US East and West tests in 1980-1993 and 1994-2008 used 4,815 shrub-years. Each year used only earlier paired records. The learned update reduced mean absolute error by 1.07-2.22 days against a full leaf-date shift. 3 of four intervals were above zero after adjustment for eight comparisons. Against a past plant mean, gains ranged from +0.77 to +1.62 days; 1 intervals were above zero. Missing visits, selected paired records, different regional plant types, and few clusters limit inference. These are archive date tests, not live forecasts or causal climate effects. The package includes code, source hashes, the panel, every prediction, missing-outcome coverage, and checks.

phenology · first leaf · first bloom · forecast calibration · observation bias · warm springs · lilac · honeysuckle