ENSO Predictions with Coupled Ocean Atmosphere Models
319
16.3 ENSO prediction models
In the past a large variety of successful ENSO prediction models has been developed and applied. The models can be divided into purely statistical models and
physical models. In the following I describe very briefly statistical models, how
they are initialized and what potential forecast skill they have. A more detailed and
complete explanation of statistical models and methods would be beyond the scope
of this paper. I therefore refer to the literature (e.g. (Latif et al. 2001 and references
therein), (Bamett and Hasselmann 1979), (Navarra 2001)). Physical models are
described in more detail in later sections.
16.3.1 Statistic al ENSO prediction models
Statistical methods of different complexity have been used over the last decades
to identify the principal modes of climate variability. The same statistical models
may be also applied for climate predictions. Common to all statistical methods is
that they need a so called training data set to fit prediction parameters. Most ofthe
data records used in ENSO predictions do not go back earlier than 1950, and thus
capture only a few realizations ofENSO extremes. Therefore the prediction parameters might be difficult to estimate.
The most simple statistical model that can be applied to ENSO predictions is persistence. From the heat content time series in Fig. 16.8 it may be inferred that usefui prediction skill is obtainable up to lead times of about 3 months. This is far
below the 12 to 18 months ofpredictability that were obtained from predictability
studies.
The next step in a hierarchy of statistical schemes is a correlation model. In section 16.2.2 it was shown that the heat content anomalies in NINO-4 lead NINIO-3
temperature anomalies by about Il months, and are thus useful as a precursor for
El Nifio events. However, the extremely high lag correlation of 0.82 was only
found in the eighties which was a period of very regular and strong ENSO events.
During the nineties the lag correlations are much lower and useful prediction skill
can be expected for only a few months.
A more sophisticated scheme would be the POP analysis presented in section
16.2.1. By projecting a combined data set of observed heat content, SST and
pseudo wind stress anomalies onto the leading POP pattems the actual position in
the two dimensional phase space can be identified, and thus the further evolution of
the system can be predicted. Altematively other variables may be used as predictors. Xu and Storch (Xu and Storch 1990) used in their study sea level pressure
anomalies in the latitude band 15°S to 40
0
N, and achieved a prediction skill (anomaly correlation between prediction and observation) of about 0.5 after a lead time of
about 10 months.
A large variety of other sophisticated statistical methods with different predictors
have been applied to forecast the state of the ENSO cycle. A good and complete
overview is given in the review paper of (Latif et al. 1997 and references therein).
319
16.3 ENSO prediction models
In the past a large variety of successful ENSO prediction models has been developed and applied. The models can be divided into purely statistical models and
physical models. In the following I describe very briefly statistical models, how
they are initialized and what potential forecast skill they have. A more detailed and
complete explanation of statistical models and methods would be beyond the scope
of this paper. I therefore refer to the literature (e.g. (Latif et al. 2001 and references
therein), (Bamett and Hasselmann 1979), (Navarra 2001)). Physical models are
described in more detail in later sections.
16.3.1 Statistic al ENSO prediction models
Statistical methods of different complexity have been used over the last decades
to identify the principal modes of climate variability. The same statistical models
may be also applied for climate predictions. Common to all statistical methods is
that they need a so called training data set to fit prediction parameters. Most ofthe
data records used in ENSO predictions do not go back earlier than 1950, and thus
capture only a few realizations ofENSO extremes. Therefore the prediction parameters might be difficult to estimate.
The most simple statistical model that can be applied to ENSO predictions is persistence. From the heat content time series in Fig. 16.8 it may be inferred that usefui prediction skill is obtainable up to lead times of about 3 months. This is far
below the 12 to 18 months ofpredictability that were obtained from predictability
studies.
The next step in a hierarchy of statistical schemes is a correlation model. In section 16.2.2 it was shown that the heat content anomalies in NINO-4 lead NINIO-3
temperature anomalies by about Il months, and are thus useful as a precursor for
El Nifio events. However, the extremely high lag correlation of 0.82 was only
found in the eighties which was a period of very regular and strong ENSO events.
During the nineties the lag correlations are much lower and useful prediction skill
can be expected for only a few months.
A more sophisticated scheme would be the POP analysis presented in section
16.2.1. By projecting a combined data set of observed heat content, SST and
pseudo wind stress anomalies onto the leading POP pattems the actual position in
the two dimensional phase space can be identified, and thus the further evolution of
the system can be predicted. Altematively other variables may be used as predictors. Xu and Storch (Xu and Storch 1990) used in their study sea level pressure
anomalies in the latitude band 15°S to 40
0
N, and achieved a prediction skill (anomaly correlation between prediction and observation) of about 0.5 after a lead time of
about 10 months.
A large variety of other sophisticated statistical methods with different predictors
have been applied to forecast the state of the ENSO cycle. A good and complete
overview is given in the review paper of (Latif et al. 1997 and references therein).