Forecast calibration and combination: Bayesian assimilation of seasonal climate predictions
Caio A. S. Coelho, University of Reading, Reading, Berkshire, United Kingdom; and D. B. Stephenson, F. J. Doblas-Reyes, and M. A. Balmaseda
The ultimate aim of this study is to produce improved probability forecasts of seasonal rainfall for South America. Such forecasts allow local governments plan their actions prior to the occurrence of climate anomalies such as those observed during El Nino-Southern Oscillation (ENSO) events. ENSO is one of the most important modes of climate variability affecting precipitation over South America. Therefore, the improvement of ENSO seasonal forecasts can improve the quality of rainfall forecasts for South America.
This study establishes a unified framework for the production of calibrated probability forecasts of observable variables based on information from ensembles of climate model predictions. In the same way that data assimilation is needed to get observed information into climate models, an analogous assimilation is required to convert multi-model climate predictions into well-calibrated forecasts of real-world observable variables. This Bayesian combination and calibration procedure is referred to as forecast assimilation.
The methodology, which allows the combination of coupled model with empirical predictions, is developed and tested in three stages. First, the Bayesian procedure is developed for the calibration of forecasts of an ENSO index (Nino-3.4) obtained from an individual coupled model. Next, the method is extended for the calibration and combination of equatorial Pacific sea surface temperature (SST) anomaly forecasts from seven DEMETER coupled models. Hence, in the second stage the method deals with multi-model forecasts and acquires the first spatial (longitudinal) component. Finally, in the third stage the Bayesian multi-model method is applied to the calibration and combination of spatial field forecasts of rainfall over South America. Results show that Bayesian combined forecasts are better calibrated and more reliable than both raw and bias-corrected coupled model forecasts.
Session 3, Bayesian Probability Forecasting
Monday, 30 January 2006, 4:00 PM-5:00 PM, A304
Previous paper Next paper
Browse or search entire meeting
AMS Home Page