Observed March temperature reduced cherry bloom error in three of six adjusted tests at three sites in 1981-2024

We tested added March temperature information beyond February temperature and a local time trend in cherry bloom records at Kyoto, Liestal and Washington. Each held-out year used 30 prior years for training. Tests covered 1981-2002 and 2003-2024. Mean absolute error fell in all six samples, by 0.55 to 3.36 days. Three block intervals remained below zero after correction for six tests: early Liestal, later Kyoto and later Washington. No site had this result in both periods. Without a time term, the early Liestal interval included zero. Bloom occurred by March 31 in 42 of 132 test records. Thus observed full-month March temperature gives information that cannot support an early-season forecast claim. Fixed sites, complete years and paired errors reduce some observation bias. Missing visit logs, possible reporting changes, coarse weather and short test periods limit inference. This is a retrospective information test, not a causal climate effect or an operational forecast test. The package gives code, source hashes, derived losses, weather extracts, checks and a figure.

phenology · cherry bloom · temperature · temporal validation · observation bias · forecast information · null result