Efforts have been devoted to improve the probabilistic multi-model prediction (PMMP) system that is operationally employed in the Asia-Pacific Economic Cooperation (APEC) Climate Center (APCC). The novelty of the proposed system lies in (i) the use of an upgraded multi-variable version of a stepwise pattern projection method (SPM) to calibrate single model predictions and obtain a more reliable forecast probability, and (ii) the combination of skillful models based on the selection of such models from among calibrated single-model predictions to formulate multi-model probabilistic prediction. The former first corrects errors in the predicted anomalies and then inflates the variance of the corrected prediction to match that of the corresponding observed variance in the individual models. The latter produces a tercile-based categorical PMMP based on the combination of skillful models from among all possible candidates after the calibration.

A comprehensive assessment of the benefits of the calibration (using the SPM and variance inflation) and combination (based on skill-based model selection) in the new PMMP system was first carried out for 23-year retrospective forecasts (1981-2003) of temperature and precipitation in a double cross-validation mode. The results indicate that both the calibration and combination significantly contribute to improving the PMMP skill for most regions of the globe. The resolution of the forecasts is improved by model correction and combination, while reliability is mainly increased by inflation. As a result, the calibrated PMMP based on model combination has significantly higher skill than the current version of (uncalibrated) the operational PMMP system for both variables over the globe. It is further shown that the new PMMP system also improves forecast skill during the real-time forecast period of 2008-2010 relative to the performance of the current operational system.

1. INTRODUCTION

2. DATA AND EVALUATION OF FORECAST SKILLS

2.1 Model and observed data

2.2 Forecast skill measures

2.3 Experimental design

3. METHODS FOR THE PMMP IMPROVEMENT

3.1 Model correction and combination

3.2 Variance inflation and probabilistic approach

4. SENSITIVITY TEST FOR THE PROPOSED METHOD

4.1 Model correction and combination

4.2 Variance inflation

5. IMPROVEMENT OF THE PMMP

5.1 Application to retrospective forecast

5.2 Application to real-time forecast

6. SUMMARY AND DISCUSSION