Projections are used to explore scenarios for catch advice and rebuilding and are an important tool for sustainably managing fisheries. We tested each projection specification for 12 groundfish stocks in the Northwest Atlantic to identify sources of bias and evaluate techniques for reducing bias. Projections were made from assessments using virtual population analysis (VPA) with 1-7 years of recent data removed from the full time series and were then compared with results from a VPA assessment on the full time series of data. The main source of bias in projections was the assessment model estimates of the numbers at age in the terminal model year + 1 (N a,T+1 ). Recruitment was responsible for more bias in projections beyond 3 years, when population numbers begin to be dominated by cohorts that were statistically generated. Retrospective analysis was performed and several adjustment factors to reduce bias were tested. Even after adjusting for bias, the remaining bias in projections was non-negligible. The direction of bias generally resulted in projected spawning stock biomass (SSB) and catch being overestimated, and the bias in catch was nearly always larger than in SSB. Scientists need to clearly communicate the direction and magnitude of this bias, managers need to consider this additional uncertainty when specifying future catch limits, and both scientists and managers need to develop more robust control rules so that objectives are achieved. Résumé :Les projections sont utilisées pour explorer des scénarios de recommandation de prises et de reconstitution de stock et constituent un important outil de gestion durable des pêches. Nous avons évalué les critères de projection pour 12 stocks de poissons de fond du nord-ouest de l'Atlantique afin de cerner les sources de biais et d'évaluer des techniques de réduction du biais. Les projections ont été faites à partir d'évaluations par analyse virtuelle des populations (AVP) de données récentes pour d'une à sept années, extraites de séries chronologies complètes, et ont été comparées aux résultats d'une évaluation par AVP sur les séries chronologiques complètes. Les estimations tirées du modèle d'évaluation du nombre selon l'âge durant la dernière année + 1 modélisée (N a,T+1 ) constituaient la principale source de biais dans les projections. La part du biais dans les projections attribuable au recrutement augmentait après trois ans, quand les chiffres sur les populations commençaient à être dominés par des cohortes générées statistiquement. Une analyse rétrospective a été réalisée et plusieurs facteurs d'ajustement servant à réduire le biais ont été évalués. Même après ajustement pour le biais, il restait un biais non négligeable dans les projections. La direction du biais se traduisait généralement par une surestimation de la biomasse du stock reproducteur (BSR) et des prises projetées, et le biais des prises était presque toujours plus grand que celui de la BSR. Les scientifiques doivent communiquer clairement la direction et la magnitude de ce biais, les aménage...

No abstract

The World Conference on Stock Assessment Methods (July 2013) included a workshop on testing assessment methods through simulations. The exercise was made up of two steps applied to datasets from 14 representative fish stocks from around the world.Step 1 involved applying stock assessments to datasets with varying degrees of effort dedicated to optimizing fit.Step 2 was applied to a subset of the stocks and involved characteristics of given model fits being used to generate pseudo-data with error. These pseudo-data were then provided to assessment modellers and fits to the pseudo-data provided consistency checks within (self-tests) and among (cross-tests) assessment models. Although trends in biomass were often similar across models, the scaling of absolute biomass was not consistent across models. Similar types of models tended to perform similarly (e.g. age based or production models). Self-testing and cross-testing of models are a useful diagnostic approach, and suggested that estimates in the most recent years of time-series were the least robust. Results from the simulation exercise provide a basis for guidance on future large-scale simulation experiments and demonstrate the need for strategic investments in the evaluation and development of stock assessment methods.