While there is valuable prediction information in forecasts of tropical Pacific sea-surface temperature anomalies, the prediction problem is again far from being deterministic. That is, the forecast problem is again analogous to the ball drop problem described in Fig. 2.4. Thus, an ensemble of forecasts are run (analogous to releasing the ball several times down the slope to find the tendency of the ball). The ensemble is made by starting the forecasts from slightly different initial conditions, so that the ensemble of forecasts have the opportunity to sample the range of possible evolutions that nature may follow in the particular forecast period.

Each ensemble member produces a forecast of all model parameters for the forecast period, at each grid-point in the model. To look at the model predictions of tropical Pacific SST, the model's predicted sea-surface temperature can be averaged over the forecast season and across a region that is central to the El Niño pattern (Fig. 1.3c). Figs. 3.1a,b show such examples from two different models.

How accurate are the forecasts of the Nino 3 SST? Experiments suggest that when models are given the state of the tropical Pacific Ocean in boreal summer (say in July), then they can project forward the expected Nino3 SST anomaly (and associated likelihood of El Niño or La Niña) with good skill for about 6 months (correlation skill above 0.7) and with reasonable skill for about 9 months (correlation skill above 0.5). When models are provided the current state of the tropical Pacific in boreal winter (say in January), they have more difficulty in predicting the Nino3 SST anomaly, though correlation skills are still typically above 0.5 for the coming boreal summer. This seasonal dependence of skill has been widely discussed in the literature. It is partly due to the increased likelihood of rapid developments of El Niño or La Niña during the approximate March-May period. Indeed, a distinction may be made between predicting an event not yet started (skill only fair) compared to predicting the evolution of an event already begun, for which there is good skill, often even including the dissipation of the event. Recognition of a period when El Niño or La Niña is becoming well established and can be projected forward for 6-9 months or more, can represent an increase in lead time and intensification of skill levels for predictions of seasonal to interannual climate variations.

An improved understanding and basis for forecasts for the tropical Pacific were helped by the development of an observing system for the tropical Pacific, that was implemented around 1982 (Fig 3.2). Skill in Atlantic and Indian Oceans is under investigation, in part hindered by lack of sub-surface observations. This situation has encouraged some coupled models to be run only for the tropical Pacific, rather than for a global domain, and to focus on the prediction SST anomalies associated with the El Niño phenomenon. However, the seasonal forecast for many parts of the tropics is substantially dependent on accurate characterization of Atlantic and Indian Ocean SST, and this represents a major challenge to CGCMs. Making use of the limited data to initialize a large set of past forecasts, estimates of the skill of a number of CGCMs are now available, such as through the DEMETER project.