Ensemble Verification

How to read this. A deterministic score cannot say whether an ensemble is any good. The question here is calibration: does the spread of the members actually describe the uncertainty? The rank histogram asks whether the gauge behaves like one more member; spread–skill asks whether the ensemble knows when it is uncertain; CRPS scores the whole distribution in mm and collapses to MAE for a single forecast, so the ensemble and the control are directly comparable; reliability asks whether “70 % chance” happens 70 % of the time; ROC asks whether the ensemble can separate events from non-events at all, independent of calibration.