Past water temperature did not improve tested fish-share forecasts at Maizuru Bay in 2012–2023

We tested whether past water temperature improves forecasts of fish species shares two scheduled surveys ahead at Maizuru Bay, Japan. We used 113 species and 528 surveys in 2002–2023. Model fitting and penalty selection used targets through 2011. Tests used 288 targets in 2012–2023. The base model used season and each species’ recent square-root share. Adding surface and bottom temperatures increased mean squared Hellinger loss from 0.263227 to 0.265429, or 0.84%. The temperature gain was −0.84%. A 95% range from three-year test blocks was −1.25% to −0.31%. When training years were also sampled and penalties selected again, the range was −3.90% to +0.05%. Recent fish shares reduced loss by 12.81% relative to season alone. Exploratory controls used the same penalties in both models or separate penalties for fish and temperature. They also gave no test gain from temperature. The result applies to these fixed additive models at one survey area. It does not exclude temperature effects or gains from nonlinear models. The package includes source data, provenance, all predictions, model fits, resampling weights, code, figures, and checks. An independent solver agreed, and a separate-directory repeat matched all 22 numerical outputs.

fish community · ecological forecasting · Maizuru Bay · water temperature · species composition · chronological validation · open data · Hellinger distance