Please see attached document for full model description with equations. We performed predictions using a Bayesian stock-recruit model. Our model predicts the number of Sockeye Salmon recruits returning to Bonneville Dam from a given brood year (BY) that reared in freshwater for a duration of FA years, migrated to the ocean in year OY, and resided in the marine environment for a duration of MA years. The number of recruits is estimated separately for the Wenatchee (wen) and Okanagan (ok) subbasins. For each unique subbasin, BY, FA, MA combination, the number of recruits is estimated as a function of the number of spawners (S) for that basin and BY, an estimated spawner fecundity parameter (f) where e^f represents the number of eggs produced per spawner, the observed proportion of recruits (p) from a given BY and subbasin combination with a given life-history pattern (ie unique combination of BY, FA, MA), estimated freshwater survival (φ) for each subbasin and BY, estimated marine survival (ϕ) for fish that entered the ocean in a given OY and a residual error term (ε). Freshwater survival is represented on the logit scale as a linear function of estimated average freshwater survival (α_fw), a slope coefficient (β_fw) and a freshwater covariate (F), which is indexed by brood year and represents river conditions in the year of fry emergence (BY + 1). Marine survival is represented on the logit scale as a linear function of estimated average marine survival (α_mar), a slope coefficient (β_mar), freshwater covariate (M), which represents ocean conditions in the year of ocean entry, and an annual random effect (γ) indexed by ocean entry year. Predictions from our model for a given return year are the sum of estimated recruits that return in a given RY = BY + 1 + FA + MA.

Data: Recruits data for our model are the values provided in the Salmon Prize 2026 Columbia River data package. The number of spawners in the Okanagan subbasin are values reported by Ogden et al. (2025) in Table 2 of their report. The number of spawners for the Wenatchee subbasin are estimated based on the number of observed spawners counted at Rock Island Dam, multiplied by the estimated proportion of total spawners bound for the Wenatchee subbasin. Note that because of this approach, our model assumes 100% conversion rates from Rock Island Dam to the Wenatchee River spawning grounds. The proportion of total spawners bound for the Wenatchee River was calculated roughly following the approach of Bailey et al. (2025), by taking the average of proportions estimated from the ratio of spawners at Rocky Reach Dam:Rock Island Dam and from the ratio of spawners at Tumwater Dam:Rock Island Dam. The proportion of recruits from each FA, MA combination in a given BY (p) is the observed ratio of recruits from a given FA,MA combination out of total recruits for that BY. For brood years where subbasin-specific recruits data is not available, p for the Wenatchee and Okanagan subbasins is assumed equal to total p of all recruits returning to Bonneville Dam. For recent brood years with incomplete returns as of 2025, p was assumed to be equal to the mean from the previous 5 brood years with complete returns. The marine covariate we used in our final model is the Bakun Cumulative Upwelling Index for 45°N, and the freshwater covariate was an index of annual mean 1-day maximum streamflow anomalies developed by NOAA NWFSC for the Upper Columbia Spring Chinook ESU (available at https://oceanview.pfeg.noaa.gov/erddap/tabledap/cciea_HB_FLOW.htmlTable?time,freshwater_ indicator®ion_and_type=%22UCS%20ESU%201%20day%20max%20flow%22). Model Estimation: Parameters for our model were estimated using Hamiltonian Monte Carlo sampling in the stan statistics package. For details of our model code and priors, please see the uploaded stan file. While we did not perform an extensive systematic covariate selection process, several different possible freshwater and marine covariate combinations were assessed based on mean absolute prediction error (MAPE) of retrospective predictions for 2021-2025. The LLM Claude Sonnet was used for R and stan coding assistance during implementation of the model.