Closing the gap between statistical and scientific workflows for improved forecasts in ecology

Concerns about increasing biodiversity loss and climate change have led to greater demands for useful ecological models. Datasets relevant for developing these models have also increased in size and complexity, including in their geographical, temporal and phylogenetic dimensions. New research often suggests that models accounting for these complexities can yield more accurate trends and predictions. We argue, however, that the usual workflows for model fitting in ecology make it difficult to evaluate and compare current models for several reasons. First, the research community is split between two disconnected spheres that prevent uniting the natural connection between trend estimation and forecasting. One research sphere focuses on using data to fit simple, trend-like models with few parameters, which is separated from research developing forecasting models, which often include complex, mechanistic submodels only indirectly informed by data. Second, in both cases, models tend to be developed without a coherent framework for linking scientific questions and understanding to statistical and modelling decisions throughout the modelling process. To address these challenges, we propose a workflow that integrates statistical and scientific practices through clear steps, many of which rely on data simulation to inform decisions in the process and where forecasting a natural output. We show how this approach, coupled with a shift towards universal training, more open model sharing and alignment on common datasets, could harmonize currently divided efforts at trend estimation and forecasting to better inform sustainable policies. This article is part of the theme issue 'Statistical workflow'.