Abstract: Based on the re-forecast data from the second-generation seasonal prediction model of National Climate Center, the model's capability to predict summer rainfall over East China and possible reasons for the forecast errors are investigated. Furthermore, the rainfall forecast skill is improved by the application of downscaling approaches. Results indicate that the model is able to capture the two major modes of spatiotemporal variability of summer rainfall over East China to some extent (i.e. the dipole mode and the uniform-distribution mode). However, forecasts at various lead times show obvious errors in variance contributions of these modes and spatial distributions of anomalies and interannual variations of time coefficients, etc. In addition, although the model can reasonably reproduce variations of large-scale circulation and sea surface temperature (SST), it shows limited skills in forecasting summer rainfall over East China. This is partially due to the model's inability to realistically depict the impacts of circulation systems such as the West Pacific subtropical high, the continental high and the middle-high-latitude blocking high. Influences of SST in the tropical Pacific and Indian Ocean on major rainfall modes over East China are also not well described in the model. Furthermore, in terms of the 500-hPa geopotential height, 850-hPa zonal and meridional winds, and SST in reforecasts for 1991-2003, predictors with the closest relationship with East China rainfall are identified on global scale and used to establish the single-factor linear regression, multi-factor stepwise regression, and multiple regression downscaling models for rainfall prediction. These downscaling rainfall prediction models are tested independently using reforecasts for 2004-2013, and significant improvements in the forecast of East China summer rainfall are obtained. For the forecast initialized on June 1, for example, the spatial correlation coefficient between predicted and observed EOF1 (EOF2) modes increases from 0.12 (0.48) for the original prediction to 0.58 (0.80) for the downscale prediction, and the corresponding temporal correlation coefficient rises from 0.47 (0.15) to 0.80 (0.67). Compared to the original forecasts by the model at other lead times, the downscaling forecast models also significantly enhance the prediction skill of rainfall.