Abstract: Stepwise regression MOS statistical method is applied to predict the pentad mean temperature of future 40 days at 24 weather stations in Ningxia, using the pentad mean temperature and dynamical extended range forecasting products from January 1982 to March 2010. In order to evaluate the predicting capability of the direct numerical model output products (DMO) and MOS which use the seasonal and monthly data, the prediction results of DMO and MOS from January to March in 2009 and 2010 are compared.The prediction accuracy and stability of MOS improved remarkably comparing with DMO, MOS method can predict the trend and extent of violent weather changes in temperature. With the drawing near of predicting time and successive correction, its prediction error values decreases gradually, and predicting results can be for the reference of medium term prediction operation.MOS prediction capability is different when using data of different lengths. The prediction capability of MOS using monthly data is better, because the chosen prediction factors using monthly data can better indicate the correlation of prediction objects in this period, and its physical meaning is much distinct. So more monthly data samples lead to better temperature prediction result and stability.MOS method merely applies the output products of dynamical extended range prediction model, so its prediction results thoroughly rely on the accuracy and stability of numerical prediction model. The prediction results may be more accurate if some observation factor, local experimental factor and climatic factor are added. There are 4 aspects which need attentions when establishing MOS equations: The equations should be established based on experimental calculating error values; the establishment of F-test value at each weather station should also consider experimental calculating error values; using data samples of longer temporal scales to verify if MOS prediction results will be better using more monthly data samples; finally, more new data are recommended to improve MOS prediction capability in the future.