We assessed current status of multi-model ensemble (MME) deterministic and probabilistic seasonal prediction based on 25-year (1980-2004) retrospective forecasts performed by 14 climate model systems (7 onetier and 7 two-tier systems) that participate in the Climate Prediction and its Application to Society (CliPAS) project sponsored by the Asian-Pacific Economic Cooperation Climate Center (APCC). We also evaluated seven DEMETER models' MME for the period of 1981-2001 for comparison. Based on the assessment, future direction for improvement of seasonal prediction is discussed. We found that two measures of probabilistic forecast skill, the Brier Skill Score (BSS) and Area under the Relative Operating Characteristic curve (AROC), display similar spatial patterns as those represented by temporal correlation coefficient (TCC) score of deterministic MME forecast. A TCC score of 0.6 corresponds approximately to a BSS of 0.1 and an AROC of 0.7 and beyond these critical threshold values, they are almost linearly correlated. The MME method is demonstrated to be a valuable approach for reducing errors and quantifying forecast uncertainty due to model formulation. The MME prediction skill is substantially better than the averaged skill of all individual models. For instance, the TCC score of CliPAS one-tier MME forecast of Niño 3.4 index at a 6-month lead initiated from 1 May is 0.77, which is significantly higher than the

No abstract

In this paper, a multi-model ensemble approach with statistical correction for seasonal precipitation forecasts using a coupled DEMETER model data set is presented. Despite the continuous improvement of coupled models, they have serious systematic errors in terms of the mean, the annual cycle and the interannual variability; consequently, the predictive skill of extended forecasts remains quite low. One of the approaches to the improvement of seasonal prediction is the empirical weighted multi-model ensemble, or superensemble, combination. In the superensemble approach, the different model forecasts are statistically combined during the training phase using multiple linear regression, with the skill of each ensemble member implicitly factored into the superensemble forecast. The skill of a superensemble relies strongly on the past performance of the individual member models used in its construction. The algorithm proposed here involves empirical orthogonal function (EOF) filtering of the actual data set prior to the construction of a multimodel ensemble or superensemble as an alternative solution for seasonal prediction. This algorithm generates a new data set from the input multi-model data set by finding a consistent spatial pattern between the observed analysis and the individual model forecast. This procedure is a multiple linear regression problem in the EOF space. The newly generated EOF-filtered data set is then used as an input data set for the construction of a multi-model ensemble and superensemble. The skill of forecast anomalies is assessed using statistics of categorical forecast, spatial anomaly correlation and root mean square (RMS) errors. The various verifications show that the unbiased multi-model ensemble of DEMETER forecasts improves the prediction of spatial patterns (i.e. the anomaly correlation), but it shows poor skill in categorical forecast. Due to the removal of seasonal mean biases of the different models, the forecast errors of the bias-corrected multi-model ensemble and superensemble are already quite small. Based on the anomaly correlation and RMS measures, the forecasts produced by the proposed method slightly outperform the other conventional forecasts.

A B S T R A C TIn this paper, a multi-model ensemble approach with statistical correction for seasonal precipitation forecasts using a coupled DEMETER model data set is presented. Despite the continuous improvement of coupled models, they have serious systematic errors in terms of the mean, the annual cycle and the interannual variability; consequently, the predictive skill of extended forecasts remains quite low. One of the approaches to the improvement of seasonal prediction is the empirical weighted multi-model ensemble, or superensemble, combination. In the superensemble approach, the different model forecasts are statistically combined during the training phase using multiple linear regression, with the skill of each ensemble member implicitly factored into the superensemble forecast. The skill of a superensemble relies strongly on the past performance of the individual member models used in its construction. The algorithm proposed here involves empirical orthogonal function (EOF) filtering of the actual data set prior to the construction of a multimodel ensemble or superensemble as an alternative solution for seasonal prediction. This algorithm generates a new data set from the input multi-model data set by finding a consistent spatial pattern between the observed analysis and the individual model forecast. This procedure is a multiple linear regression problem in the EOF space. The newly generated EOF-filtered data set is then used as an input data set for the construction of a multi-model ensemble and superensemble. The skill of forecast anomalies is assessed using statistics of categorical forecast, spatial anomaly correlation and root mean square (RMS) errors. The various verifications show that the unbiased multi-model ensemble of DEMETER forecasts improves the prediction of spatial patterns (i.e. the anomaly correlation), but it shows poor skill in categorical forecast. Due to the removal of seasonal mean biases of the different models, the forecast errors of the bias-corrected multi-model ensemble and superensemble are already quite small. Based on the anomaly correlation and RMS measures, the forecasts produced by the proposed method slightly outperform the other conventional forecasts.