Statistical methods can be used to detect the accuracy of climate model projections, by assessing how well the models capture the feedback and interaction between all of the components of the climate system: the system dynamics. A study published online this week in Nature Climate Change may help to improve climate model predictions and help increase confidence in making decisions informed by those predictions.

Michael Runge and colleagues develop a method to detect when predictions from a single model - or set of models - are failing to match actual observational data, and apply the method in two examples: the change in range of the northern pintail duck, and Arctic sea-ice projections. For the northern pintail, they compared the observed latitude of the North American breeding population and the predicted latitude from two models. Their analysis shows that these methods would have detected a shift in the breeding range in 1985, 20 years earlier than it was observed.

The second example examines the ability of models to accurately predict the level of Arctic sea-ice in September. Their analysis of 11 climate models under a high emissions scenario suggests that the current set of models is accurately representing the observed system dynamics and, therefore, capable of accurate predictions. However, the authors note that some individual models are showing rapid changes in their fit to observations, suggesting the model ensemble may be at risk of failure in the future. The authors conclude that more weight should be given to those climate models that forecast an ice-free Arctic by September 2055.