MAELSTROM, a machine learning-based approach for stock assessment

Managing the fishing industry is crucial to maintain a balance between the exploitation of marine resources and their natural ability to recover. To achieve such a goal, stock assessment models are built to combine commercial catches and population abundance time series. These approaches can often handle only single-species, losing the information regarding the ecological interactions affecting the dynamics of the species. Moreover, conventional stock assessment models are based on explicit deterministic equations, which can fail to represent the ecological interactions between the species and its environment. In this paper we present Maelstrom, a multispecies predictive model based on neural networks that can interpret fishery-dependent and -independent data to return a forecast of stock abundance considering variations in fishing effort. Although neural networks-based forecasting of ecological and fisheries time series is well established, we present a customizable Shiny tool in which a multi-species, age-structured framing is integrated and devised to be applied in real management frameworks. Namely, we set up five scenarios of increasing complexity in which three commercial species are considered. To assess the model's reliability, we conducted a benchmark test comparing Maelstrom to a routinely used stock assessment tool, the a4a model framework, using RMSE and MAE for validation. The results, besides showing a good degree of accuracy of Maelstrom to classical stock assessment models, endorse the potential of a multi-species approach even when working on shorter time-series. The Shiny application returns statistics, plots, and a report of all the operations conducted during the stock assessment.