Our intern Ivan Vorobevskii did a great job in crunching Early Warning System output and putting it together with regionally observed flood/warning levels. Typical scores like POD, FAR, AUC are also calculated. The plot below gives an idea on the verification problem, which is kind of fuzzy, rather than really being crisp/dichotomous. We have to put some more effort in the specific verification approach (preferably, thinking the problem a bit from the perspective of warning product users, which typically do some sort of fuzzy decision making, too!)
Besides the forecasts, the Early Warning System operationally also runs in hindcast mode (driven only with QPE data), delivering some possibility to separate input parameter (i.e., QPF data) uncertainty from model uncertainty, by using hindcast data for verification.