Quantifying the value of quantile mapping methods using a perfect-model approach

Here we compare three quantile mapping statistical downscaling methods (i.e. CDFt, Change Factor and Equidistant) to determine the methods' performance simulating means, medians, standard deviations, inter-quantile ranges and tail behavior. The methods were analyzed using synthetic data from different types of distributions; and additionally to test the statistical downscaling time-invariance assumption, we used historical and future daily maximum and minimum temperature outputs from a high resolution global atmospheric model as pseudo-observations, and a coarsened version of the same high resolution outputs as predictors. The study region includes 16 gridpoints from different climate regions across USA and Canada.

Considering that all the methods analyzed are quantile-based approaches, the downscaled results were notably different. Moreover, we caution the users on using the methods in data scarce situations as some methods show instabilities in the tails of the distributions, which can cause over/under prediction of the return periods associated to specific events. These instabilities were not found when using the data-rich synthetically generated datasets.