Up to know, almost everybody predicts mean values over a given time span. While it is convenient to do, it is only part of the full story. Because distributions of a variable are not necessarily reflected by an average value. 2/7
In this study we propose a strategy to evaluate full distributions relatively against each other. For this we employ the IQD to compare hindcasts, historicals and climatology vs. a reference (assimilation simulation). 3/7
But as we cannot evaluate them by a single value like correlation, we do it by counting how often over a given time span one simulation wins against another. It is a different form of looking at verification and a hopefully much more accessible for communication purposes. 4/7
Main physical result: different seasons show different skills. Especially with involved ice processes the skill between hindcasts and historicals vary considerately. We use a quite normal distributed variable, so the results are in most cases close to a correlation analysis. 5/7
This study demonstrate that predictions of distributions are possible, but require creative approaches for the verification. It opens a new dimension and increases the temporal resolution to look at these predictions, 6/7
The research was done together with Sebastian at Uni Hamburg supported by the project Coming decade funded by BMBF. Myself funded for this by #nckf at DMI and A4 project funded by Marine Institute 7/7
Comments