How to best threshold and validate stacked species assemblages? Community optimisation might hold the answer

The popularity of species distribution models (SDMs) and the associated stacked species distribution models (S‐SDMs), as tools for community ecologists, largely increased in recent years. However, while some consensus was reached about the best methods to threshold and evaluate individual SDMs, little agreement exists on how to best assemble individual SDMs into communities, that is, how to build and assess S‐SDM predictions. Here, we used published data of insects and plants collected within the same study region to test (a) if the most established thresholding methods to optimize single species prediction are also the best choice for predicting species assemblage composition, or if community‐based thresholding can be a better alternative, and (b) whether the optimal thresholding method depends on taxa, prevalence distribution and/or species richness. Based on a comparison of different evaluation approaches, we provide guidelines for a robust community cross‐validation framework, to use if spatial or temporal independent data are unavailable. Our results showed that the selection of the “optimal” assembly strategy mostly depends on the evaluation approach rather than taxa, prevalence distribution, regional species pool or species richness. If evaluated with independent data or reliable cross‐validation, community‐based thresholding seems superior compared to single species optimisation. However, many published studies did not evaluate community projections with independent data, often leading to overoptimistic community evaluation metrics based on single species optimisation. The fact that most of the reviewed S‐SDM studies reported over‐fitted community evaluation metrics highlights the importance of developing clear evaluation guidelines for community models. Here, we move a first step in this direction, providing a framework for cross‐validation at the community level. Foreign Language La popularité des modèles de distribution d'espèces (SDMs) et leur forme cumulée multi‐espèces (S‐SDMs) comme outils d’étude dans le domaine de l’écologie des communautés a fortement augmenté ces dernières années. Cependant, alors qu'un certain consensus a été atteint concernant les meilleures méthodes pour sélectionner un seuil et évaluer les predictions binaires (presence‐absence) d'espèces individuelles, cette question du seuil n'a pas été abordée du point de vue des prédictions de communautés. Dans cet article, nous utilisons des données publiées de distribution de p