Expected future performance of salmon abundance forecast models with varying complexity

Abstract

We evaluated the scope for improving abundance forecasts for fishery management using Sacramento River fall Chinook salmon (Oncorhynchus tshawytscha) as a case study. A range of forecast models that related the Sacramento Index (SI; an index of adult ocean abundance) to jack (estimated age 2) spawning escapement the previous year were considered. Alternative models incorporated effects of density dependence, local environmental conditions, the abundance of the previous cohort, and trends or autocorrelation in the jack-to-SI relationship. Forecast performance was assessed in terms of bias, accuracy, ability to track trends in the SI, and management objectives. Several models achieved higher accuracy than the model used for management, but no single model performed best across all criteria, and substantial forecast error remained across all approaches considered. Environmental models generally performed better than the management model, but there were differences in the relative importance of individual environmental variables over time and among model formulations. Accounting for model selection uncertainty in environmental models decreased their forecast performance. Simpler models often had similar or better performance than environmental models. In particular, the model incorporating temporally autocorrelated errors demonstrated potential for modest forecast improvement with relatively little additional model complexity.