Prior summer weather raised prediction error in three British butterfly country samples in 2010-2024 with a fixed ridge model

We tested whether prior summer temperature and rain improve one-year predictions of annual abundance-index change for nine butterflies. Rolling ridge models used England, Scotland and Wales indices. Each model used only predictors from years before the target year. A fixed penalty limited model complexity. In 2010-2024, weather raised mean squared error by 17.8%, 10.9% and 7.9% over a prior-growth and coverage model. Only Scotland had an approximate six-test adjusted skill interval below zero. Results in 1995-2009 were mixed. All six weather models had higher error than a zero-change prediction. A larger penalty gave a small later improvement in Wales. These results apply to the tested prediction rules and final data release. They do not show that weather has no biological effect. Shared weather, short test periods, changing surveys, selected species and revised source indices limit inference. The package contains raw source snapshots, hashes, all predictions, code and offline checks.

butterflies · weather · prediction · ridge regression · observation bias · UKBMS · temporal validation