consists of 70 stationary platforms moored to the ocean bottom by 5 kilometers of nylon and kevlar line. Each surface platform measures winds, humidity, atmospheric pressure, and the sea surface temperature. Attached to the line are a series of instruments that measure the temperature and pressure at intervals down to 500m below the surface. The entire observing system for the tropical Pacific can be viewed at http://www.pmel.noaa.gov/toga-tao/noaa/elnino.html and is described in National Research Council (1994). The ocean data are quality controlled by a variety of checks.
Second, the ocean data are combined with the atmospheric data to provide an estimate of the initial state of the coupled system. In practice, the time scales of the atmosphere are short compared with those of the ocean so that the surface winds and subsurface temperatures (at various depths) are assimilated into an ocean model to gain an estimate of the initial state of the ocean alone. The atmospheric state is estimated from the analysis performed for weather prediction. This gives an initial state for the coupled atmosphere-ocean system.
Third, starting from the initial state, a forecast is made. The coupled system is allowed to evolve freely for a given lead time and the forecast is the state of the coupled model after that lead time. Sometimes the initial atmospheric state is not known even though the ocean initial state is known, so that an ensemble of forecasts is made starting from various possible atmospheric initial states. This approach provides an envelope of possible forecasts, and, from the distribution of the final ensemble members, an estimate of the uncertainty of the forecasts. Finally, the forecast is evaluated by statistically comparing the forecast state with the analysis of the current state at the time for which the forecast was made, due regard being paid to the uncertainty of the current analysis.
The forecasts are made at ranges of months to years, so that, for each forecast made, it would normally take months to years to find out to what extent it proved accurate. It is therefore very cumbersome (basically impractical) to develop forecast systems by waiting for the many forecast-analysis cycles needed to evaluate a system. For example, since the first successful ENSO forecast by a coupled atmosphere-ocean model (Cane et al., 1986), a forecast every month would yield a total of only about 120 forecasts. By contrast, since numerical weather prediction was developed in 1948, over 20,000 forecasts have been made. To develop prediction systems more efficiently, past data are used to initialize the state of a model and "forecasts" are made of events that have already occurred and scored by data already in hand. These retrospective forecasts are called "hindcasts."
We do not want to leave the impression that the only way to forecast is with numerical coupled models. Statistical methods are routinely used