Forecasting fish recruitment in age-structured population models

Recruitment in age-structured stock assessment models can be forecasted using a variety of algorithms to provide advice on the anticipated consequences of different possible management actions. Selecting one method over another usually involves some subjectivity, yet can be consequential to the provision of advice. Extensive case-specific testing is not always feasible. We evaluated the forecast skill in 3-, 5- and 10-year forecasts of 16 recruitment forecasting methods under various circumstances to provide a broad evaluation and general guidelines on the reliability of forecasts. We used 31 operating models based on existing stock assessment models applied to a diversity of stocks with empirical data, which we show to be generally representative of assessed stocks worldwide. Although no single best-performing method could be identified, we found that time-series methods were most likely to perform poorly. Both forecast skill across all methods and forecast sensitivity to the selected method were linked to the properties of the stock or assessment: age at maturity and recruitment autocorrelation in 3-year forecasts and previous long-term recruitment variability in 10-year forecasts. In some situations, all forecasting methods resulted in systematic over- or underestimation of spawning stock biomass. The simulation approach employed here to assess forecast performance, rooted directly in the predictions of existing stock assessment models, can be a complementary tool to existing simulation approaches which generate alternative sets of population dynamics or observations and we discussed the advantages and limitations.

Automated summary

Various algorithms can be used to forecast recruitment in age-structured stock assessment models, and the choice of method can have a significant impact on the advice provided. Testing the forecast skill of different methods showed that time-series methods were more likely to perform poorly. Factors like age at maturity and recruitment autocorrelation influenced forecast skill across all methods, with some resulting in systematic over- or underestimation of spawning stock biomass.