Winter temperature gave no clear gain in March 1 cherry bloom forecasts at three sites in 2004-2023

Does mean winter temperature improve cherry bloom forecasts after adjustment for late-winter temperature and a time trend? We used annual records from Kyoto, Washington DC, and Appenzell in 1982-2023. We joined 126 site-years to NASA POWER daily temperature. We fit separate linear models at each site in 1982-2003 and tested them in 2004-2023. At March 1, the base model used a year trend and mean February temperature. The added variable was mean temperature from November through January. Adding winter temperature increased mean absolute error at all three sites. The equal-site error rose from 6.32 to 6.72 days. The estimated benefit was -0.39 days, with a 95% five-year block-bootstrap interval from -0.85 to 0.08 days. With annual model updates, the benefit was -0.09 days, with an interval from -0.55 to 0.28 days. Full-period winter slopes had intervals that included zero at all sites. These results give no clear forecast gain for this winter proxy. They do not rule out a biological chilling effect. Fixed sites, complete annual records, separate site models, and a time trend reduce some observation bias. Observer effort and tree changes remain unknown. Reanalysis weather, different bloom thresholds, and three selected sites limit transfer to other settings. This study does not test air pollution, frost damage, or species distributions.

phenology · cherry bloom · winter temperature · climate · temporal validation · observation bias · null result · MERRA-2