Abstract: In order to solve the practical problems in short range climate prediction, an operational system has been developed for monthly scale climate prediction based on Dynamical Extended Range Forecast (DERF) model output, statistical prediction methods and downscaling techniques. The system has the following features. It provides two subjunctive methods including Perfect Prediction (PP) and Model Output Statistics (MOS) methods. The former supposes that the prediction of model is perfect enough and needn't to be modified. The downscaling model can be built on the historical observed data. The latter supposes that the prediction of model has certain bias and the downscaling model is developed using the hindcast data of model output. Predictants can be determined in two ways. One is called the single station method and predictants are determined at each station within the studied area based on the reasonable physical mechanism. The other is called the regional average method and predictants are determined based on the relationship of regional average features and predictants. Three types of high correlation centers, i.e., positive correlation centers, negative correlation centers and local correlation centers are used to determine key circulation regions which could be taken as predictants. Six downscaling methods are used to obtain predictants from key circulation regions, and seven combinations of correlation coefficients within key circulation regions are used to find optimal prediction result. The stepwise regression, optimal sub tree regression, analogous regression and minimum distance resemblance are used to develop statistic prediction models. Predicted results can be assessed after the data is updated. The output of the prediction methods provided by the system is compared with observed precipitation data at 88 stations of Guangxi in June, 2005—2008. The results of the independent samples show that the skill of the MOS method is much better than the PP method in the downscaling techniques. The best forecast method is based on the predictors which are selected from the key circulation region near the station. The Empirical Orthogonal Functions (EOF) and combined dynamical statistical prediction method are more accurate and stable than the other downscaling methods. In determining key areas which affected predictants, the regions where model output and predictants, reanalysis data and predictants are well correlated are selected. The prediction skill of the downscaling techniques is generally above 70%, which is higher than that of the conventional physical-statistical prediction.