75 years since Monod: It is time to increase the complexity of our predictive ecosystem models (opinion)

The complexity of our mechanistic ecosystem (biogeochemical, eutrophication, phytoplankton, water quality) models has not changed substantially since they were developed in the 1970s. Consequently, there is now a large disconnect with contemporary problems (e.g., toxin production), knowledge of biological and ecological processes (e.g., intracellular mechanisms, sediment bed overwintering) and environmental observational technologies (e.g., metatransciptomics). This limits the utility of models for making predictions and supporting management. There are several reasons against increasing complexity, including (a) number of required assumptions, (b) risk of overfitting, (c) higher uncertainty, (d) missing knowledge, (e) lack of observations for calibration and validation and (f) difficulty of developing, running, analyzing and communicating the model. Here I review those arguments and conclude that, for mechanistic, predictive ecosystem modeling, they either do not apply, are not a significant problem in practice or can readily be solved by providing more resources to modelers. Further, a review of these issues leads to the conclusion that more complexity generally increases the predictive skill of a model, because more information is used to constrain it. This can be formulated as a new rule: more in, more out (MIMO). MIMO suggests that more complex models make better predictions, but this should not be adopted as a universal modeling strategy, because in practice, the difficulty associated with developing, understanding and communicating complex models has to be considered. However, those are problems readily solved by more resources and I argue that more funding needs to be made available to develop complex ecosystem models.