W-138-10
Combining Data-Limited Fisheries Models to Derive Robust Estimates of Stock Status

Sean Anderson , School of Resource and Environmental Management, Simon Fraser University, Burnaby, BC, Canada
Jamie Afflerbach , National Center for Ecological Analysis and Synthesis
Andrew Cooper , Simon Fraser University
Mark Dickey-Collas , International Council for the Exploration of the Sea
Olaf Jensen , Department of Marine & Coastal Science, Rutgers University, New Brunswick, NJ
Kristin Kleisner , Ecosystem Assessment Program, NOAA/NMFS/NEFSC, Woods Hole, MA
Catherine Longo , National Center for Ecological Analysis and Synthesis
Coilin Minto , Marine and Freshwater Research Centre, Galway-Mayo Institute of Technology, Galway, Ireland
Giacomo Chato Osio , EC JRC, IPSC, MAU
Daniel Ovando , Bren School of Environmental Science and Management, University of California Santa Barbara, Santa Barbara, CA
Andrew Rosenberg , Union of Concerned Scientists
Elizabeth Selig , Conservation International, Arlington, VA
James Thorson , Fisheries Resource Assessment and Monitoring Division, Northwest Fisheries Science Center,, National Marine Fisheries Service, National Oceanic and Atmospheric Administration, Seattle, WA
The majority of the world's fisheries lack sufficient data to apply traditional stock assessment models. In recent years, a number of methods have been developed to estimate stock status for these data-limited stocks. However, some methods may perform better in certain scenarios and there are multiple ways to gauge model performance. How can we reconcile the output from these methods and make the best possible assessment of stock status? A previous working group designed a large-scale fully factorial simulation experiment to evaluate four data-limited assessment models: one model that extrapolates stock status from assessed fisheries based on life history and catch trajectory and three models that use Shaefer-like biomass dynamics combined with assumptions about harvest dynamics and population resiliency. Here, we draw on ensemble methods from the fields of machine learning and climate science to blend estimates of stock status from these models. We train our blended models on our simulated dataset and test them on the RAM Legacy Stock Assessment Database where reference points have been established from data-rich stock assessments. We evaluate the efficacy of different ensemble methods and make recommendations about practical and statistical issues to consider when combining model output for ecological resource decision-making.