Nicholas Joshua Clark, SKM Ernest, Henry Senyondo, Juniper Simonis, Ethan P White, Glenda M Yenni, KANK Karunarathna

Forecasting the responses of natural populations to environmental change is a key priority in the management of ecological systems. This is challenging because the dynamics of multispecies ecological communities are influenced by many factors. Populations can exhibit complex, nonlinear responses to environmental change, often over multiple temporal lags. In addition, biotic interactions, and other sources of multi-species dependence, are major contributors to patterns of population variation. Theory suggests that near-term ecological forecasts of population abundances can be improved by modelling these dependencies, but empirical support for this idea is lacking. We test whether models that learn from multiple species, both to estimate nonlinear environmental effects and temporal interactions, improve ecological forecasts for a semi-arid rodent community. Using Dynamic Generalized Additive Models, we analyze monthly captures for nine rodents over 25 years. Model comparisons provide strong evidence that multi-species dependencies improve both hindcast and forecast performance, as models that captured these effects gave superior predictions than models that ignored them. We show changes in abundance for some species can have delayed, nonlinear effects on others, and that lagged effects of temperature and vegetation greenness are key drivers of change. Our findings highlight that multivariate models are useful not only to improve near-term ecological forecasts but also to ask targeted questions about community dynamics.