Three recent years gave no clear gain in monthly fish composition forecasts at Maizuru Bay in 2012–2023

We tested whether three recent years improve forecasts of monthly fish species shares at a Maizuru Bay visual census transect. We used 528 surveys, 113 species columns and 22 full years from 2002 to 2023. At the start of each test year in 2012–2023, we forecast all 12 monthly share vectors from earlier years only. The primary model used the last three years for each calendar month. The baseline used all earlier years for that month. Mean squared Hellinger loss was 0.20839 for the recent model and 0.23081 for the baseline. The error reduction was 9.71%, with a nominal 95% three-year block interval of −2.79% to +20.97%. The model had lower error in seven of 12 years. Skill was −5.41% in 2012–2017 and +21.71% in 2018–2023. Planned five-year and blended forecasts gave larger gains, but they do not change the primary test. The primary interval included zero with all four tested block lengths. A later check with a center matched to Hellinger loss also gave an interval that included zero. Random annual order gave lower skill, which shows order dependence under an exchangeable-year reference. It does not show a cause. The primary gain remains uncertain for this survey area, count-share target and fixed forecast rules. The package has source data, provenance, all forecasts, code, figures, resampling draws and independent numeric checks.

ecological forecasts · fish community · seasonality · Maizuru Bay · forecast benchmark · species composition