Model projections and weighting

 

Climate change projections are generally based on an ensemble of climate models representing the state-ofthe-art in understanding and modelling climate. The reason for using an ensemble of models is that no single model can be considered the best, since different models exhibit varying levels of realism in simulating climate, depending on the region and variable of interest. Even if a single best model could be determined, there is no guarantee that its present-day performance would cause it to give more reliable projections of future climate.

Climate change projections differ from weather forecasts in several crucial respects. One important difference is that, while we learn the accuracy of weather forecasts in the next few days, the true performance of future climate projections will remain unknown until many decades from now (Weigel et al., 2010). In the absence of a consensus on which models are the best, common practice has been to rely on “model democracy,” whereby each model in a multi-model ensemble is treated equally. This equal-weighting method assumes that each model is different and yet equally plausible.

In recent years, however, there is increasing evidence in the scientific literature that model democracy has some drawbacks. Accurate present-day model performance may not guarantee future performance, but poor performance clearly does not inspire confidence (for example, models that severely underestimate current Arctic sea ice coverage may not be reliable in projecting future changes in sea ice coverage). As a result, there is a growing appreciation that some performance-based weighting of model projections may be appropriate. Indeed, the IPCC Fifth Assessment illustrated this for the case of Arctic sea ice (Collins et al., 2013). However, a clear consensus on to how to weight models has not yet emerged.

A further drawback of model democracy is that it assumes each model is independent. However, climate models often share common features because one model may use computer code adopted from another model with minor adjustments, or two models may have been developed from a common earlier model. Although schemes to account for model performance and independence are being developed and tested (e.g., Sanderson et al., 2017; Knutti et al., 2017), this is still an emerging area of research. Initial exploration of weighting approaches suggests that differences between weighted and unweighted projections for Canada are small, and so traditional, unweighted multi-model projections are presented in this report.