Action plans for the conservation of biodiversity rely on forecasts of how species and natural systems will change in the future, and how they will respond to alternative management strategies. In an ideal world we – scientists and ecologists – would be able to provide those forecasts with great accuracy. The problem about making such predictions is that they cannot be validated without the ability to travel in time. Lacking this possibility, decision makers have to rely on current model predictions for making decisions that will have consequences on the species in the long term. But how do we know if our current predictions about the future distributions of species are accurate?

In a new study published in Global Ecology and Biogeography, we evaluated what aspects of the data used to fit models, the model settings and the species traits determine robust predictions of species distributions into the future.

Predictive accuracy transferability
This figure shows the changes in predictive performance over time (AUC values) of models fit across different time periods (t1 to t4) and all species groups. Years covered by the different time periods: t1=1950-1980, t2=1980-1990, t3=1990-2000 and t4=2000-2013. Models based on the same training dataset are shown under the same grey-scale shade. For example, AUC values for t1.t4 refer to models trained with t1 data and evaluated with t4 data (the longest time lag evaluated in the study). The figure suggests that models transfer well into the future (there were not large drops in AUC values over time). Adapted from Figure 2 in the article.


In the absence of data from the future, we looked at a set of 318 Australian terrestrial mammals, birds, amphibians and reptiles using a historical species dataset covering the period 1950–2013. For example, we put ourselves in the shoes of a modeler in the 1980s, made predictions about the 1990s and 2000s, and then checked how good those predictions were. We found that in general models that predict well for the current time period are also likely to provide good predictions into the future time periods, that is, they are likely to transfer well into the future. However, this transferability depends on many factors.


The most influential are the breadth of a species’ range (models for species with broad geographical ranges tended to perform worse over time than models for locally restricted species) and the (how well all environmental gradients suitable for the species have been sampled). Factors related to the species traits (e.g. taxonomic group, preference for a given habitat type or species body size) were not highly influential in describing the variation in predictive performance over time.

Relationship between model accuracy (area under the receiver operator characteristic curve, AUC values) measured across all species groups evaluated in the study (amphibians, reptiles, mammals and birds) and factors describing (1) intrinsic aspects of distribution data (e.g. sample size, the geographical spread of presence data); (2) time lag (years) between the collection of model fitting data and observation data used to evaluate models; and (3) ecological traits of the species (e.g. type of preferred habitat: forest, woodland, shrublands, grasslands or waterbodies). Figure adapted from Figure 3 in the manuscript.

The key message of our study is that we can make more robust predictions by ensuring we use data that sample well the environmental and geographical space in which the species is known to exist. We should also strive to identify and map drivers of widespread and generalist species.

Ref.  Morán-Ordóñez, A., Lahoz-Monfort, J. J., Elith, J. and Wintle, B. (In press) Evaluating 318 continental-scale species distribution models over a 60-year prediction horizon: what factors influence the reliability of predictions? Global Ecology and Biogeography. DOI: 10.1111/geb.12545.