Understanding extreme weather is imperative to society, particularly given the growing body of evidence suggesting that the characteristics of weather extremes are changing. Given this imperative, it is crucial that climate models accurately simulate such extremes. Models are routinely validated with respect to appropriate datasets, but often the extremes are a secondary focus. Furthermore, uncertainty in the extremes is rarely quantified, or even characterised, and it remains unclear how well they are represented in the observational record and, by extension, simulation output.

Therefore, towards quantifying uncertainty in the extremes of precipitation, we examine the correspondence between a variety of data sources. Specifically, we consider widely used observational precipitation datasets from both land based gauge and satellite measurements, reanalysis products and simulation output over the contiguous United States. The simulations were performed using the Community Earth System Model under the ILIAD framework, which runs a large, multiresolution ensemble of five-day reforecasts to achieve a long-term dataset in which simulated weather events correspond to observed weather events.

We apply bivariate methods of extreme value statistics to combinations of these datasets in a pair-wise approach, which yields quantitative measures of correspondence (or equivalently dependence) into the tails of the respective precipitation distributions. While a multivariate extension to this approach has not yet been developed, we describe how results from the bivariate analysis can be used to quantify uncertainty in the extremes, and thus how model fidelity can be assessed.