Forecasting the future state of biodiversity is increasingly important for planning, management, and assessing our understanding of ecological systems. We evaluate our ability to forecast the diversity of breeding birds at hundreds of locations throughout North America. To do this we combine 30 years of time-series data on bird distribution and abundance with monthly time-series of climate data and satellite-based remote-sensing. We compare an array of different spatial and dynamic modeling approaches, explicitly model uncertainty and observation, and evaluate the performance of forecasts across different time lags using existing time-series.
Most approaches to forecasting species richness yield reasonable cross-site predictions, with good matches between locations predicted to have large or small numbers of species in the future and the actual diversity values. However most approaches also perform poorly at predicting the dynamics at individual locations, with most approaches failing to meaningfully outperform models that assume diversity is static. Simple models of individual time-series perform best for predicting richness dynamics, but the accuracy of these models decays more rapidly with the lead time of the forecast than models that use information on how richness varies across space. Increasing the spatial and temporal scales predicted generally improves the accuracy of forecasts. Results of forecasts of the abundance of individual species will also be presented.