Previous summer temperature gave no common cherry bloom forecast gain at three sites in 1981-2025

We tested whether previous June-August temperature improved March 1 cherry bloom forecasts after control for February temperature and a local year trend. We used complete records at Kyoto, Liestal and Washington DC in 1951-2025 and nearest-cell GHCN-CAMS temperature. Rolling 30-year models gave 45 forecasts per site. Tests used 1981-2002 and 2003-2025. Five of six mean absolute error benefits were negative. Later Kyoto had a benefit of 0.30 days, with a six-test adjusted block interval of 0.02 to 0.67 days. That interval included zero with two-year blocks, 20-year training or an added January control. There was no common gain across sites and periods. This result does not reject biological carryover effects. Fixed sites, complete years and separate models reduce coverage bias. Unknown observation changes, coarse weather, different bloom rules and short tests limit inference. The package includes code, source hashes, derived weather, annual losses, uncertainty and independent checks.

phenology · cherry bloom · summer temperature · carryover · forecast validation · observation bias · null result