Figure 4. Six-month running-mean time series of global 500-mb
geopotential height ACC (top) and RMSE (bottom, m) at day 5 for each
version of the 13-km SHiELD and the contemporary operational GFS. Note
that the operational GFS upgraded to v13 on 11 May 2016 and v14 on 19
July 2017.
Precipitation RMSE and biases have also improved during SHiELD
development. The 2018 version significantly reduced both RMSE (Figure 5)
and Bias (arithmetic difference between time-mean model and observed
precipitation; Figure 6) at all lead times compared to earlier versions.
Prediction of CONUS precipitation is more challenging given the smaller
area and larger seasonal cycle but RMSE still improves every year and
there is nearly no bias, especially in the 2019 version. Zhou et al
(2019) give a more thorough description of precipitation forecast skill,
including other metrics. Probability distribution functions (PDFs) of
precipitation (Figure 7) show that all of the versions depicted here
have a low bias in the frequency of moderate precipitation and a high
bias of both light and heavy precipitation rates compared to TRMM,
although versions of SHiELD using the GFDL microphysics (2017 and later)
modestly alleviate these biases. Both the GFS and all versions of 13-km
SHiELD capture the observed CONUS PDF very well.