Forecasting Pacific salmon recruitment using empirical dynamic modeling

Background/Question/Methods

Attractor reconstruction offers a tantalizing suite of methods to forecast wildlife abundance. Empirical dynamic modeling (EDM) and the closely related multiview embedding (MVE) match recent abundance patterns with similar patterns in the past to forecast abundance in the future. This allows forecasting to proceed without structural assumptions about the relationship between response and explanatory variables. Further, these methods offer a natural way to include environmental covariates among the explanatory variables. Including environmental covariates in single-species assessments and forecasting is a major focus of the ecosystem approach to fisheries management (EAFM), a stepping stone to multi-species ecosystem-based management (EBM). However, empirical dynamic modeling and related methods are often data-intensive, opaque to users, and deployed in unrealistic forecast settings. We ask, do empirical dynamic modeling and multiview embedding outperform conventional model-based methods to forecast sockeye salmon recruitment abundance in the Fraser River? We used 65 years of spawner abundance and four environmental covariates to forecast 20 years of recruitment abundance in 10 sockeye salmon stocks. We developed an R package where the code for these methods is visible to the public. We forecast recruitment each year using only previous data, mimicking real-world forecasting scenarios, and compared forecasts using accepted forecast metrics.

Results/Conclusions

TBD