We used Gaussian Process Empirical Dynamic Models (GP-EDM) to forecast sockeye run sizes. We considered a variety of predictors, including total returns, returns of each age class, and several environmental variables (e.g., NPGO). All models were fit in R using the GPEDM package (https://tanyalrogers.github.io/GPEDM). Data from prior to 2016 were used to train each model, and then one-year-ahead predictions (with sequential model updating) were made on the 2016-2025 data. Of the models that achieved similar median APEs on the ten year test set, the model which did the best in even-numbered years was ultimately selected. For both the whole Columbia and the Okanagan, the best EDM model had a sibling regression component (i.e., used a one year lag of age 1.1 returns) but no environmental lags. The best EDM model for the Wenatchee did use environmental lags, but still had overall poor predictive performance. Ultimately, we chose to simply use the difference between our whole Columbia and Okanagan forecasts to obtain a forecast for Wenatchee.

GP-EDM models for the whole Columbia and the Okanagan both were of the form returns_{t} = f (age1.1_{t-1}, returns_{t-1}, returns_{t-2}, returns_{t-3}, returns_{t-4}) where returns is total returns summed across all age classes (for the whole Columbia or just the Okanagan, respectively).