Multi-model data from DEMETER (Development of a European Multimodel Ensemble System for Seasonal to Interannual Prediction) are used to investigate the performance of multi-model ensemble of seasonal precipitation in China during 1959－2001. Moreover, combined with multi-model data, an optimal subset regression (OSR) approach is used to perform a statistical downscaling forecast for the seasonal precipitation in China. Its forecast skill is compared with those of different multi-model ensemble methods. Results show that similarly poor simulating abilities to seasonal precipitation in China can be found in several models, and the multiple linear regression (MLR) ensemble forecast performs worse than ensemble mean (EM). The forecast skill of seasonal precipitation can be significantly improved by the OSR approach in China. In summer, the temporal anomaly correlation coefficient (ACC) advances obviously in the south of Yangtze River, Tibet, and the central area of Inner Mongolia. The area-averaged ACC is up to 0.29 in China, which is clearly better than those of EM and MLR ensembles. In winter, the OSR approach is helpful to improve the low level which occurs in the multi-model ensemble forecasts in the north of China. Moreover, probabilistic Brier skill score (BSS) also indicates the advantage of OSR approach over multi-model ensembles for the seasonal precipitation forecast. It is important to note that the physical mechanism between the predictor and seasonal precipitation in China should be further investigated, although a significant improvement in the seasonal precipitation forecast can be achieved by the OSR method.