Catch history reduced mean prediction error more than bottom temperature in tested ridge models for 35 eastern Bering Sea fish taxa, 2007–2025

We tested whether current bottom temperature improves fish catch occurrence predictions after previous local catches are known. We used open NOAA summer trawl data for 35 fish taxa selected in 1982–2006. Ridge logistic models were fixed after 2006 and tested on 5,136 station-year cases in 17 survey years from 2007 to 2025. Catch history reduced mean Brier error by 9.19% relative to a model with station, date, and swept area; the nominal 95% year-block interval was 6.74% to 11.91%. Temperature alone reduced error by 0.33% (−1.10% to 2.19%). In the primary test, temperature beyond catch history reduced error from 0.076665 to 0.076538, a 0.17% gain (−0.80% to 1.25%). Gains differed among taxa. A stronger model penalty gave a 2.25% relative temperature gain but larger absolute error than the primary history-only model. Current temperature was measured during the survey, so these are conditional predictions, not forecasts made before the survey. The result applies to the selected models, taxa, and known stations. It does not establish a causal effect or a general limit on temperature-based models. The package has source data, provenance, code, all primary predictions, figures, and repeat instructions.

Bering Sea · fish distribution · bottom temperature · catch history · temporal validation · occurrence prediction · open data · ridge logistic regression