September temperature gave no clear autumn interval forecast gain beyond trends in eight Japan tests in 1989-2025

We tested whether September mean temperature improves later-year forecasts of the time from leaf colour to leaf fall. We used JMA ginkgo and maple records and GHCN-CAMS temperature. Each fit used earlier years only. The selected panel had 6,186 paired records at 62 distinct stations. Tests used 2,555 tree-type station-years in North and South Japan in 1989-2006 and 2007-2025. Adding temperature to station and observation-category intercepts and a common time trend changed mean squared interval error by -2.40% to +1.42%, where positive values mean lower error. All eight adjusted uncertainty intervals included zero. Separate local trends gave the same null conclusion. A baseline without a trend gave larger gains. The result shows no clear added interval forecast value for this predictor and model. It does not show that temperature has no effect. Selected coverage, missing visit and tree histories, coarse weather data and short test periods limit inference. The package contains source hashes, code, paired evidence, predictions and uncertainty estimates.

phenology · forecast validation · ginkgo · maple · temperature · observation bias · null result