[Previous Article] [Next Article]
Consolidated Forecasts of Tropical Pacific SST in Niño 3.4
Using Two Dynamical Models and Two Statistical Models
contributed by David Unger, Anthony Barnston, Huug van den Dool and Vern Kousky
Climate Prediction Center, NOAA, Camp Springs, Maryland
In this Bulletin we find a fairly large number of forecasts for the east-central tropical Pacific SST
for the coming year. Our objective here is to synthesize information from some of the predictive
sources into a single objective estimate of the likely evolution of the SST's in the tropical Pacific.
One approach to the problem is to combine, or consolidate, the forecasts of several models into a
single forecast based on the past behavior of each contributing model. Multiple linear regression
is used here to extract information from several models to produce a consolidation forecast. In
this case we use three input models. One is dynamical: the CMP12 NCEP coupled model (Ji et al.
1994); and the other two models are statistical: the NCEP constructed analogue (CA) model (Van
den Dool 1994; Van den Dool and Barnston 1995), and the NCEP canonical correlation analysis
(CCA) model (Barnston 1994). The individual forecasts of each model are shown elsewhere in
this Bulletin issue.
To derive the multiple regression equations for each target season for each lead time, histories of
the forecasts of each model were obtained. The CCA and CA models have histories extending
back to 1956 and the NCEP coupled model 1981-96. To circumvent the problem of the differing
units and climatologies used, all forecasts were converted to actual C prior to the equation
derivation. The observations were expressed likewise. The regressions are based on forecasts for
the Niño 3.4 region (5N-5S, 120-170W).
The desired lead times of the consolidated forecasts range from 0.5 months to 12.5 months by 1
month increments, where lead time is defined as the time skipped between the time of the forecast
and the beginning of the forecasted (target) period. For example, the forecasts shown here, which
are issued in the middle of December, 1997, have target periods including Jan-Feb-Mar
1998,...Jan-Feb-Mar 1999. Two of the three individual models have forecast histories whose
leads range to 12.5 months or greater, while one (the NCEP coupled model) has a maximum lead
of only 8.5 months. Consolidated forecasts for lead times higher than 8.5 months, therefore, are
based only on the other two models.
Because the NCEP coupled model forecast only has a 1981-96 history, the training period for the
regression is limited to that period and thus results in greater uncertainty in the coefficients than
would be the case if a longer history could be used. When that model is not included in the
consolidation process for the longer lead times, the 1956-96 period is used to derive the
regression equations, making for a more favorable training sample. Data from three lead times
were pooled together to help equation stability and help smooth forecasts from projection to
projection. Predictor and predictand data from the season preceding and following the target
season were combined to form the regression equation. The first (last) target season shares the
equation with the adjacent season. As a final step, the consolidation forecast is adjusted, where
necessary, to prevent a prediction outside of the envelope formed by the highest or lowest of the
input model prediction. Because of the limited sample size, regression may not be trusted to
inflate a forecast, but, rather, should be regarded as a skill-weighted mean of the input forecasts.
Limiting its value to either the highest or lowest model essentially allows the consolidation to give
total weight to that model's forecast.
The consolidated forecast for Niño 3.4 made in mid-December 1997 is shown in Fig. 1. Forecasts
are expressed as standardized anomalies relative to the 1961-1990 climatology. The box and
whisker intervals for the forecasts indicate the one and two error standard deviations, based on
estimated skill following shrinkage of the dependent sample skill results in accordance with the
sample size and number of predictors. The component forecasts are displayed on the same chart
for comparison. The observed SST anomaly for the most recent 3-mo period is also shown. For
this prediction, the first 4 lead times have been adjusted downward to prevent the consolidation
forecast from forecasting values higher than those from the coupled model.
The consolidation forecast indicates that the strong warm event in progress will continue into
early 1998 and will rapidly return to normal, starting in late spring. Standardized anomalies peak
at 3.3 standard deviations above normal in FMA 1998. The absolute temperature anomaly
declines steadily throughout 1998. Temperatures decrease from about 2.2 degrees C above
normal in JFM, to an anomaly of near zero by ASO 1998, and further declines to about 1 degree
C below normal by the end of the year.
Acknowledgments: We are grateful to Ming Ji and Ants Leetmaa from the National Centers for
Environmental Prediction, for providing the forecast histories from their respective dynamical
models, as well as their current real-time forecasts.
REFERENCES
Barnston, A.G., 1994: Linear statistical short-term climate predictive skill in the Northern
Hemisphere. J. Climate, 5, 1514-1564.
Ji, M., A. Kumar and A. Leetmaa, 1994: An experimental coupled forecast system at the
National Meteorological Center: Some early results. Tellus, 46A, 398-418.
van den Dool, H.M., 1994: Searching for analogues, how long must we wait? Tellus, 46A,
314-324.
van den Dool, H.M. and A.G. Barnston, 1995: Forecasts of global sea surface temperature out to
a year using the constructed analogue method. Proceedings of the 19th Annual Climate
Diagnostics Workshop, November 14-18, 1994, College Park, Maryland, 416-419.
Fig. 1. Consolidated forecast (thick line) for the standardized anomaly of the SST in the Niño
3.4 region (5N-5S, 120-170W) for the next 13-running 3-month periods. Month labels on the
abscissa denote the middle months of the 3-month predictand period. Box and whiskers for each
point indicate the one and two error standard deviation intervals. The latest observation (Sep -
Oct - Nov 1997) is also shown by the filled ellipse. The prediction from each component model
is shown for comparison.
