Reliable ensemble-based probability forecasts of weather and climate require a representation of the inevitable uncertainties that arise from the computational representation of the partial differential equations of climate. A number of different representations are under development: the multi-model ensemble, the perturbed parameter ensemble, stochastic physics, forcing singular vectors and stochastic optimals. Each of these techniques will be reviewed, and results shown.

In the case of the multi-model ensemble, results will be drawn from the European Union project DEMETER on seasonal forecasting. The DEMETER ensemble comprises 7 global coupled ocean-atmosphere models run in hindcast mode over the ERA-40 period. It will be shown that seasonal ensemble forecasts made using any of the individual DEMETER models are underdispersive, and the corresponding forecast probabilities overconfident. By contrast, probabilities from the multi-model ensemble forecast are generally much more reliable, and consequently more skilful. The implications of such enhanced reliability for specific quantitative applications, such as malaria prediction, will be discussed.

It will be concluded that an ability to inter-compare the probabilistic reliability of seasonal and decadal forecasts made using different representations of model uncertainty, will be important in determining the best method for representing model uncertainty in forecasts of climate change.