A method for recalibrating and combining ensemble forecasts

Because of the inherent uncertainty in the evolution of the atmosphere at time- scales longer than a few days, seasonal climate forecasts generally are presented in probabilistic terms. The forecast probabilities provide an indication not only of the range of possible outcomes, but also the level of confidence the forecaster has in the forecast. Although it can be demonstrated that probabilistic forecasts are of potentially greater value to decision-makers than deterministic forecasts even when the probabilities are unreliable, it is clearly preferable for the forecast probabilities to correspond with the observed relative frequency of the forecast event as closely as possible.

Ensemble methods are used to obtain forecast probabilities from numerical forecasting methods, but little attempt has been made to correct these probabilities for model errors and biases, or to combine ensembles from different models. In this paper, the reliability of seasonal forecast probabilities provided by individual general circulation models is tested, and compared with the reliability of pooled ensembles from different models. A method is described for recalibrating the ensemble model output and for combining the forecasts from different models more effectively than by simply pooling ensembles from models with different skill levels. The method is based on estimates of the dependence of the capture rates of the ensembles on the values of the individual ensemble member predictions. Logistic regression is used to estimate the capture rates.