Accounting for current challenges in commercial fish stock assessment models: it is worth increasing the complexity?

Marine fisheries lie at the intersection of ecological complexity and human activity, providing food, employment, and cultural value worldwide, while exerting strong pressure on ecosystems. Scientific advice for sustainable management relies on stock assessment models, which transform biological and fisheries data into indicators of stock status and projections of future dynamics. Over recent decades, these models have evolved from simple, data-limited approaches to integrated frameworks capable of assimilating a wide range of inputs. Yet, a persistent challenge remains: finding the right balance between parsimony and complexity, i.e. how much detail is needed to represent essential ecological processes without compromising interpretability, transparency, or robustness. This thesis addresses this question through a general introduction, four analytical chapters, and a general discussion, focused on two species examined through empirical and simulated case studies: European hake (Merluccius merluccius), a demersal species with pronounced sexual size dimorphism in the Atlantic, and yellowfin tuna (Thunnus albacares), a highly migratory pelagic species in the Indian Ocean. The first two chapters refine assessment inputs using the Bayesian Integrated Nested Laplace Approximation (INLA). Fish abundance and distribution are shaped by spatial correlation, environmental variables, and temporal patterns; accounting for these dependencies yield standardized parameters that more accurately reflect ecological processes. For hake, a multivariate model is developed to correct biases in maturity ogives caused by uneven sampling. Similarly, a hierarchical delta model is developed to improve the recruitment index by incorporating spatiotemporal dependence, with depth emerging as the most influential covariate. These applications highlight the potential of INLA as a tool to reduce bias and strengthen fisheries advice, with broad applicability across species and contexts. The following two chapters investigate internal structural complexity configurations in Stock Synthesis 3 (SS3), a modern integrated framework adaptable to multiple settings. For hake, singlesex configurations failed to reproduce the dynamics captured by two-sex models. For tuna, spatially explicit models that incorporated processes such as regional recruitment and movement, along with tagging data, improved the representation of regional dynamics and reduced biases in biomass and depletion estimates. However, certain parameterizations also increased bias and created issues of parameter identifiability. Overall, the thesis concludes that complexity is valuable when hypothesis-driven, introduced incrementally, supported by sufficient data, and accompanied by explicit treatment of uncertainty. Improving inputs with methods such as INLA offers a way to increase realism, while structural extensions should be applied cautiously, supported by diagnostics and best practices. Complexity is therefore not an end in itself, but a strategic tool: it adds value when key processes are central and well understood, while in other contexts more parsimonious approaches may prove more effective.