A fixed seasonal branch gave an uncertain 1.1% gain in Maizuru fish-share predictions for 2013–2023

We tested whether a fixed season group improves predictions of fish species count shares beyond measured water temperature. We used 528 visual surveys from Maizuru Bay, Japan. Models used 2002–2012 for training and 2013–2023 for testing. A temperature kernel model was compared with the same model restricted to March–August or September–February. Training validation selected the bandwidths. Mean test Hellinger loss fell from 0.30582 to 0.30233. The relative gain was 1.14%. Its central 95% range from three-year blocks was −0.35% to +2.18%. The primary test did not resolve a positive gain. Both branches had nearby training temperatures for 263 of 264 test surveys. Planned checks changed the gain and whether its range included zero. Equal shares among present species gave a 1.70% gain. Removing the most counted training species gave 2.53%. These checks change the target. Calendar-only and joint calendar–temperature models had losses near the temperature-only model. The primary range also depended on block length. The result applies to one transect, these models and this time split. It does not show that temperature alone is sufficient or identify a biological cause. The package preserves source tables, provenance, code, predictions, all planned results, figures and independent numeric checks.

fish communities · seasonal change · Maizuru Bay · temperature · chronological prediction · Hellinger loss · null result · reproducible research