Since its inception in 2005, the Asia-Pacific Economic Cooperation (APEC) Climate Center (APCC) has devoted considerable effort to developing a multi-model ensemble (MME) prediction system for producing improved and well-validated seasonal global forecasts. Since 2007, the APCC has issued monthly MME seasonal climate predictions for the upcoming three-month period with a one-month lead time, and has disseminated it to APEC member economies. This paper gives a comprehensive documentation of the current status of the APCC operational multi-model performance, with a large set of predictions of global temperature and precipitation in different seasons for the retrospective (1981–2003) and real-time (2008–2011) forecast periods. It focuses on a comparison between the different multi-model combination schemes, in terms of the skill of temperature and precipitation forecasts. In particular, simple averaged MME in which model forecasts are combined with equal weights, two empirically-weighted MMEs based on multiple regression and a calibrated MME based on a step-wise pattern projection method are considered. Results indicate that when considering all aspects of the predictions, (i.e. with different variables, regions and seasons), the methods of the multiple regression-based weighted MMEs consistently show a lower skill than the simple averaged MME. The calibrated MME predictions by model correction provide the most effective way of reducing errors and improving forecasts for both variables than other MME method during the cross-validated 21-year retrospective, as well as during independent 4-year real-time forecast periods. Basics for the rationality of the MME and the possible causes of the failure and success of the different MME methods implemented in the APCC operations, are also discussed.

1. INTRODUCTION

2. DATA AND FORECAST SKILL MEASURES

3. METHODOLOGY OF THE APCC MME PREDICTION SYSTEM

3.1 Simple composite method

3.2 Point-wise multiple regression method

3.3 Synthetic super-ensemble method

3.4 Step-wise pattern projection method

4. PERFORMANCE OF APCC SINGLE-MODEL AND MULTI-MODEL PREDICTIONS

4.1 Seasonality and spatial distribution of forecast skill

4.2 Interannual variation of forecast skill

5. COMPARISON OF APCC MULTI-MODEL PREDICTION SYSTEMS

5.1 Retrospective forecast

5.2 Real-time forecast

6. A BASIC FOR RATIONALITY OF MULTI-MODEL ENSEMBLE

6.1 Simple averaged method

6.2 Point-wise multiple regression method

6.3 Spatial shift correcting method

7. SUMMARY