Quantifying the uncertainty associated with extreme weather likelihood estimates derived from large model ensembles

In the aftermath of extreme weather, policy makers, contingency planners and insurers often seek to understand the likelihood of experiencing such events. The most common tool for this is extreme value analysis (EVA), but likelihood estimates based on observed or reanalysis data can be highly uncertain due to the relatively short observational record. Substantially larger samples of plausible extreme weather events can be obtained using the UNprecedented Simulated Extremes using ENsembles (UNSEEN) approach, which involves applying EVA to large forecast/hindcast ensembles. While larger sample sizes generally reduce the uncertainty associated with EVA, using seasonal or decadal forecast data introduces additional uncertainties related to model bias and model diversity. In this study, a multi-model ensemble of hindcast data from the CMIP6 Decadal Climate Prediction Project was analysed to quantify these additional uncertainties in the context of extreme temperature and rainfall across Australia. Factoring in model bias and diversity dramatically increased the uncertainty associated with estimated event likelihoods from the UNSEEN approach, to the point that it equaled or exceeded the uncertainty from an observation-based approach at most locations. Model diversity tended to be the largest source of uncertainty (60-70% of the total). Bias correction was also a significant source of uncertainty (30-40%), while the uncertainty associated with EVA was trivial. Our results suggest that an UNSEEN-based approach to estimating the likelihood of climate extremes should be understood as an approach that has different uncertainty characteristics to an observation-based approach, as opposed to less uncertainty.