IRI uses a 2-tiered climate prediction system, in which SST predictions are first developed (tier one), and then atmospheric GCMs are run using the SST predictions as the lower boundary condition over the oceans. As detailed in the section called "SST Forecasts", several different predicted SST scenarios are used. In the late 1990s and very early 2000s, three atmospheric GCMs were used for the multi-model ensemble climate forecasts. Since 2005, seven such models are used, and among all SST scenarios and models, more than 100 ensemble predictions are generated per forecast. Access the predictions here.Initial conditions for the integrations are not observed, but rather "restart files" from continuing AMIP-type historical simulations. Using observed initial conditions is believed to be of little value when the first forecast target period begins about two weeks later than when the forecasts are run, and the benefit of weather prediction skill would be small. The predictions are influenced mainly by the SST anomaly patterns. The mixture of initial atmospheric conditions, all appropriate for the season, give rise to the spread of seasonal mean climate predictions-this spread representing the probability distribution of the climate forecasts across the globe. The forecasts are run out to 7 months into the future.

The distribution of the forecasts, originally expressed as a set of predicted values having a mean and a pattern of spread, can be represented by a forecast probability density curve. This curve can be compared to a similar curve, but one that represents the historical distribution of the observations for the season and location in question. The proportion of the forecast curve falling into the below-normal, near-normal, and above-normal thirds of the historical distribution curve then become the forecast probabilities for those three portions of the historical distribution. In cases when the two curves are nearly the same, the forecast would call for equal chances for the below-normal, near-normal, and above-normal categories. Otherwise, the forecast may indicate an enhanced probability for one of the categories. For example, the probabilities for below, near, and above normal indicated in the figure above are 15%, 32% and 53% respectively.