We used Gaussian Process Empirical Dynamic Models (GPEDM) to forecast sockeye run sizes for each of the fourteen rivers. We considered a variety of individual (single river) and hierarchical (multiple river) GPEDM models, with or without age structure. The models used lags of total (or age-structured) returns to forecast future total (or age-structured) returns. All models were fit in R using the GPEDM package (https://tanyalrogers.github.io/GPEDM). Data from prior to 2015 were used to train each model; for each river, we then selected the model that minimized the MAPE (mean absolute percent error) of leave-future-out predictions on the 2015 - 2024 data. Across the 14 rivers, the average MAPE of the best model on the test set was approximately 31% (min = 15%, max = 69%). For a majority of the rivers, the best model included some form of age structure.

See the attached files for a description of methods/results, the model code, and previous year forecasts.