Retrospective Evaluation of Preseason Forecasting Models for Pink Salmon

Abstract

Models for making preseason forecasts of adult abundance are an important component of the management of many stocks of Pacific salmon Oncorhynchus spp. Reliable forecasts could increase both the profits from fisheries and the probability of achieving conservation and other management targets. However, the predictive performance of salmon forecasting models is generally poor, in part because of the high variability in salmon survival rates. To improve the accuracy of forecasts, we retrospectively evaluated the performance of eight preseason forecasting models for 43 stocks of pink salmon O. gorbuscha over a total of 783 stock-years. The results indicate that no single forecasting model was consistently the most accurate. Nevertheless, across the 43 stocks we found that two naïve time series models (i.e., those without explicitly modeled mechanisms) most frequently performed best based on mean raw error, mean absolute error, mean percent error, and root mean square error for forecasts of total adult recruits. In many cases, though, the best-performing model depended on the stock and performance measure used for ranking. In 21% of the stocks, a new multistock, mixed-effects stock–recruitment model that included early-summer sea surface temperature as an independent variable along with spawner abundance demonstrated the best performance based on root mean square error. The best-performing model for each pink salmon stock explained on average only 20% of the observed variation in recruitment. Owing to the uncertainty in forecasts, a strong precautionary approach should be taken to achieve conservation and management targets for pink salmon on the West Coast of North America.