Contribution of a Pre‐Processing Phase to Create a Biomass Indicator From Big Data With an Application to the Striped Red‐Mullet in the Bay of Biscay

ABSTRACT Biomass indicators derived from fishery‐dependent LPUE data are widely used for data‐limited stocks. But the influence of data pre‐processing on their reliability remains poorly documented. Using Bayonne district fisheries of striped red mullet as a case study, we evaluated how three levels of pre‐processing—none, light, and hard—affect data structure, model behavior, and the resulting biomass signal. Although the raw, lightly filtered, and heavily filtered datasets retained broadly similar temporal patterns, pre‐processing substantially reduced noise and heterogeneity arising from spatial coverage, fleet composition, and declaration practices. Models fitted to pre‐processed datasets showed markedly improved performance, with RMSE reduced by half compared to models using raw LPUE data, and computation time decreasing sharply as filtering increased. Importantly, pre‐processing also altered the set of explanatory variables selected by GAMMs, highlighting that data refinement reshapes the ecological and behavioral information embedded in LPUE. Our results demonstrate that pre‐processing is not a neutral technical step but a determinant of the biological representativeness, interpretability of LPUE‐based indicators, and the frugality of big data models. This study provides a transparent framework for integrating expert knowledge into big‐data workflows and for improving the robustness of populations indicators used in data‐limited stock assessment.