Box 1. Models for prediction

From the following article:

The past and the future of El Niño

Peter J. Webster and Timothy N. Palmer

Nature 390, 562-564(11 December 1997)



A great deal of progress has been made in forecasting El Niños. For example, the event of 1982-83 was only evident once it had started. At that time, seasonal predictions were made using empirical models whose equations were based on statistical relationships derived from historical time-series data15. In 1986, however, Cane and Zebiak16 demonstrated the possibility of making very useful forecasts several seasons ahead, by taking the basic equations which describe Newton's laws of motion, together with the laws of thermodynamics, and applying them to describe the coupled dynamics of the ocean and atmosphere of the tropical Pacific. By the standards of comprehensive numerical weather-prediction models, the Cane-Zebiak model is relatively simple17. But, impressively, during the late 1980s and early 1990s it was used to predict the timing and magnitude of the variations of sea-surface temperatures in the tropical Pacific (and hence El Niño).

Those successes led to the development of comprehensive ocean-atmosphere models, where the atmospheric component was either a numerical weather-prediction model, or a global climate model, and where the ocean component covered all of the ocean basins6. For a decade, it has been hard to prove that the comprehensive coupled models were better than the Cane-Zebiak model. That may have changed with the 1997-98 event, which comprehensive coupled models forecast whereas the Cane-Zebiak model did not.

Part of the answer may lie in the relatively simpler way that models of Cane-Zebiak complexity handle the data used to create the initial conditions for the ocean-atmosphere forecasts. Essentially, the comparatively large systematic errors of the simple models makes them harder to use when assimilating these initial data. (Indeed, this was probably the case with the 1997-98 warming — later experiments have indicated that if a more comprehensive initial data set had been run on the Cane-Zebiak model, it too would have forecast a major El Niño with a similar lead time.) The comprehensive models also have such errors, but these have been reduced over the years and, in the long run, there is more scope to reduce them further than in the simpler models.

Empirical models can still be useful in prediction. Unlike the dynamically based models, however, their skill is obtained from statistical relationships derived from earlier, training data. Therein lies the rub, because on timescales of decades and centuries the world's climate is not stationary (which may or may not be due to anthropogenic effects)18, and that will always limit the predictive power of empirical models.