Our 2026 forecast for the eight Bristol Bay competition rivers is a linear pool (equal-weight per-river mean) of the two published preseason forecasts---ADF&G and UW-FRI---giving a point estimate of 42.8 million. We also built and fully validated an age-structured sibling model of our own, but a three-way accuracy test on the 2025 outcome showed it was materially less accurate than either expert forecast (2025 per-river MAPE: ADF&G 15.7%, UW 18.4%, our model 32.9%), and that adding it to the pool degraded accuracy (two experts 16.8% --> with our model 21.7%). We therefore keep our model out of the point forecast on the evidence, and submit it instead as our entry in the retrospective category, where it has a documented, reproducible backtest.

Two models were developed. (1) The submitted 2026 forecast is a linear opinion pool: the equal-weight, per-river mean of the two published preseason forecasts — ADF&G and UW-FRI (UW Alaska Salmon Program, 14 Nov 2025, Tables 1 & 3). (2) An age-structured sibling model was built and validated as a benchmark and as our retrospective entry: for each river it forecasts the four dominant age classes (1.2, 1.3, 2.2, 2.3) by log–log regression of each age class on its younger sibling's return the previous year, with a geometric-mean fallback, summed to a total run. Data variables: returns by age class per river and year (AgeClass_1.1, 1.2, 1.3, 2.1, 2.2, 2.3) from the competition return table (returns reconstructed by Curry Cunningham, UAF); the sibling predictor (the prior-year younger age class); and the external ADF&G and UW-FRI per-river 2026 forecasts. No environmental covariates are used in the submitted forecast (PDO was tested and not retained). A three-way test on the 2025 outcome showed the sibling model alone (32.9% MAPE) was less accurate than ADF&G (15.7%) and UW-FRI (18.4%) and degraded the pool, so it is excluded from the point forecast.