This paper provides a detailed illustration that it is beneficial for ENSO
forecasting to improve in priority the model parameterizations, instead of
increasing only the consistency of the initial conditions with the coupled
model, Moreover it is shown that the latter can lead to misleading results.
Using sea level data in addition to wind to initialize the Cane and Zebiak
model does not improve El Nino forecasts. Nudging the observed wind to the
model one to initialize the forecasts as proposed by Chen et al. also fail
s to correct the model deficiencies and degrades the initial conditions of
the ocean and atmosphere. These failures are explained by large model error
s in the off-equatorial sea level and wind anomalies that play a key role i
n the coupled behavior. The use of data to estimate new model parameterizat
ions allows for significantly improving both the initial conditions and the
forecasts up to 6-month lead time. This success holds for all the various
initialization procedures used in this study. Because of erroneous winds si
mulated by the atmospheric component in the eastern Pacific, errors grow fa
st though. Replacing the atmospheric model by a statistical one results in
more reliable predictions over 1980-98. Por lead times up to 1 yr, the mode
l predicts well the observed anomalies between 1984 and 1993, including the
sea level rises along the ITCZ after warm events and their subsequent equa
torward migration. This success is attributed to the consistency between th
e observed anomalies over this period and the mechanisms involved in mainta
ining the oscillatory behavior of the model, including the off-equatorial m
eridional wind anomalies.