Statistical models that combine information on species occurrences with abiotic predictor variables are commonly used to make projections of the distribution of biodiversity under future climate. Given the expected novelty of future climate and biotic assemblages, the transferability of such correlative models through time is frequently questioned. The expectation is that model transferability will decline as climate becomes increasingly novel relative to the period in which the model was fit and as fitted species- and assemblage-climate relationships break down. Thus, key questions regarding the use of statistical models for projecting future biodiversity patterns are: (i) when in the future will climate become “too novel” and (ii) might some modeling approaches be more transferable through time than others? Using the fossil pollen record and a comprehensive model intercomparison, we determined how model performance declines with climatic and biotic novelty over the last 22,000 years before present. We then used this information to make inferences regarding model transferability to future climate. We also explored whether the drivers of climate novelty differ between past and future climate.
Results/Conclusions
Using the paleo-record, we found that model transferability (as measured using standard metrics of model performance) declined precipitously with increasing climatic and biotic novelty until a threshold value was reached. At this value, the decline in model performance slowed and remained only marginally better than random. When considered in the context of future climate, similar declines in model performance would be expected starting less than two decades from present, with the threshold value at which models become largely uninformative being reached before 2060, depending on climate scenario. We also found that the drivers of climate novelty differed between past and future. Throughout the late Pleistocene, change in potential evapotranspiration was the primary driver of climate novelty, where as changes in precipitation of the wettest quarter and mean yearly water deficit index emerged as important drivers of climate novelty in the Younger Dryas and Bølling-Allerød respectively. However, in the Holocene no single variable emerged as being an important driver of climate novelty. In contrast, change in actual evapotranspiration was the primary driver of future climate novelty and most prominently so after 2040. Our results suggest that correlative biodiversity models may have little utility beyond mid-century.