Abstract:A statistical downscaling model for forecasting autumn rainfall at stations over Southwest China was established in this study based on real-time prediction of numerical products from the Climate Forecast System (CFS) and observation data. The autumn 500 hPa geopotential height from CFS and summer sea surface temperature from reanalysis data, which are relatively clear in physics for autumn rainfall over Southwest China, were selected as the two predictors. The key regions of the two predictors are 10°S-50°N/70°-180°E and 30°S-30°N/30°-120°E. There exist high correlations of time coefficient of first leading SVD modes between predictors and observation. The correlation coefficients have passed the 0.01 significance level test on the 500 hPa geopotential height field and the 0.001 significance level test on the sea surface temperature field. The statistical downscaling hindcast for the 1982-2017 result shows that the spatial correlation coefficient can improve the performance of prediction compared with that of the original CFS, and multi-year mean is increased from -0.06 to 0.38 with the maximum getting up to 0.7. The root mean square error decreases in comparison with the output of original CFS at most stations with the maximum being 40%. At the same time, the statistical downscaling hindcasts on spatial pattern of extreme minimum and maximum are fine.