The predictability of atmospheric responses to global sea surface temp
erature (SST) anomalies is evaluated using ensemble simulations of two
general circulation models (GCMs): the GENESIS version 1.5 (GEN) and
the ECMWF cycle 36 (ECM). The integrations incorporate observed SST va
riations but start from different initial land and atmospheric states.
Five GEN 1980-1992 and six ECM 1980-1988 realizations are compared wi
th observations to distinguish predictable SST forced climate signals
from internal variability. To facilitate the study, correlation analys
is and significance evaluation techniques are developed on the basis o
f time series permutations. It is found that the annual mean global ar
ea with realistic signals is variable dependent and ranges from 3 to 2
0% in GEN and 6 to 28% in ECM. More than 95% of these signal areas occ
ur between 35 degrees S-35 degrees N. Due to the existence of model bi
ases, robust responses, which are independent of initial condition, ar
e identified over broader areas. Both GCMs demonstrate that the sensit
ivity to initial conditions decreases and the predictability of SST fo
rced responses increases, in order, from 850 hPa zonal wind, outgoing
longwave radiation, 200 hPa zonal wind, sea-level pressure to 500 hPa
height. The predictable signals are concentrated in the tropical and s
ubtropical Pacific Ocean and are identified with typical El Nino/ Sout
hern Oscillation phenomena that occur in response to SST and diabatic
heating anomalies over the equatorial central Pacific. ECM is less sen
sitive to initial conditions and better predicts SST forced climate ch
anges. This results from (1) a more realistic basic climatology, espec
ially of the upper-level wind circulation, that produces more realisti
c interactions between the mean flow, stationary waves and tropical fo
rcing; (2) a more vigorous hydrologic cycle that amplifies the tropica
l forcing signals, which can exceed internal variability and be more e
fficiently transported from the forcing region. Differences between th
e models and observations are identified. For GEN during El Nino, the
convection does not carry energy to a sufficiently high altitude, whil
e the spread of the tropospheric warming along the equator is slower a
nd the anomaly magnitude smaller than observed. This impacts model abi
lity to simulate realistic responses over Eurasia and the Indian Ocean
. Similar biases exist in the ECM responses. In addition, the relation
ships between upper and lower tropospheric wind responses to SST forci
ng are not well reproduced by either model. The identification of thes
e model biases leads to the conclusion that improvements in convective
heat and momentum transport parametrizations and basic climate simula
tions could substantially increase predictive skill.