Which fisheries and biological factors affect the misclassification of stock status determined by data-limited methods?

More scientifically based stock assessments of data-limited species are rapidly increasing worldwide. Concur-rently, non-negligible biases of stock status estimated using data-limited methods are becoming an issue. It is well known that the assumed prior distribution on depletion for some data-limited methods strongly affects the estimated stock status. Priors on depletion are best set based on expert knowledge, however, expert subjectivity and experience should be considered. Moreover, for a very data-limited stock, there is no information from experts. Owing to such constraints, the blind application of default priors on depletion often takes place. Here, we examined fishery-related, model assumption-related, biology-related, management-related, and spatial-based characteristics of stocks in which such default priors tend to work well, and vice versa, by applying a machine learning method (XGBoost) to the RAM Legacy data. The results suggest that false healthy misclassification of stock status in Catch Maximum Sustainable Yield (CMSY) and Abundance Maximum Sustainable Yield (AMSY) methods occurs more for stocks that are less managed with short time series length with slow growth and less variation in recruitment. In contrast, false overexploited misclassification of stock status in CMSY and AMSY occurs more for stocks that are well managed, have long time series length, and high variation in recruitment. Filtering out non-suitable stocks based on such characteristics or correcting the bias using machine learning methods will prevent the blind application of default priors and may prevent misclassification of stock status for data-limited species.

Automated summary

Scientifically based stock assessments of data-limited species are increasing worldwide, but the biases associated with data-limited methods are becoming a concern. This study used machine learning to examine characteristics of stocks where default priors on depletion tend to work well, and found that different factors can lead to misclassification of stock status. Filtering out unsuitable stocks or using machine learning methods can prevent blind application of default priors and improve the accuracy of assessments for data-limited species.