The first issue when applying the decaying averaging method is the choice of the decaying weight w.
Different w’s had been chosen and tested. Figure 2 shows some of the decaying weight sensitivity test results.
Among the six curves in Fig. 2, the curve OPR_OPT is for the calibrated NCEP/GEFS result using dependent data and the curve OPR_RAW is for the raw NCEP ensemble forecast.
The other four curves (OPR_RUN_DAV2%, OPR_RUN_DAV1%, OPR_RUN_DAV0.5%, and OPR_RUN_DAV0.25%) are for the RPSSs of 500-hPa geopotential height that are averaged from 1 March
2004 to 28 February 2005 for the Northern Hemisphere with decaying weights of 2%, 1%, 0.5%, and 0.25%, respectively.
All four calibrated ensembles show improvements compared with the raw ensemble OPR_RAW
for all lead times. There is little room for further improvement
compared with the OPR_OPT test for short
lead times until day 4. Though the four curves are
close together, for short lead times OPR_DAV2% is
better than the other decaying weights. On the other
hand, OPR_DAV0.25% is the best for week-2 forecasts.
Other statistics are calculated that show the 2%
ensemble produces large improvements in ROC and
BSS scores over the Northern and Southern Hemispheres.
The improvement of these scores in summer
is more significant than in spring. A higher decaying
weight (10%) is also investigated and compared with
2%. The choice of a 10% weight works better for the
tropics compared to 1% or 2%. In general, the 2% weight
works better for most regions and seasons (not shown).
For an optimal result, the decaying accumulated bias with
a 2% weight is used in operations and is updated every
day for all 35 selected variables.