Latest counts reduced next-slot squared log-count forecast error beyond seasonal models for 17 Maizuru Bay fish species in 2011–2023

We tested whether the latest fish count improves a later count forecast beyond the usual seasonal cycle at one Maizuru Bay transect in Japan. We selected 17 common species using 2002–2009 and tested all 312 census slots in 2011–2023. Each direct linear forecast used only observations available at its origin. Adding the latest log(1 + count) to annual and semiannual seasonal terms reduced next-slot mean scaled squared log error from 0.76827 to 0.50557, a 34.19% gain. A nominal 95% interval from common three-year blocks was 27.95% to 39.43%. Adding a linear time trend to both models left a 23.54% gain, with an interval of 22.12% to 24.97%. Next-slot point gains were positive for all 17 species and all 13 test years. Gains were smaller at longer leads and depended on the reference model. Copying the latest count gave slightly lower absolute log error than the fitted model, so model ranking depended on the loss function. These results apply to released visual counts at one transect and to ordinal half-month sampling slots. They do not identify a causal process or establish skill at other sites. The package preserves source tables, URLs, hashes, code, all 26,520 primary prediction rows, figures, uncertainty draws and offline checks.

ecological forecasting · fish counts · seasonality · Maizuru Bay · chronological validation · open data