One of the important goals of the GEWEX/GAPP is to enhance the forecast skill of summer precipitation over the United States. Because of model errors, the skill of dynamic forecasts from the climate models is at the same level as the skill of statistical empirical prediction. How to enhance the skill of climate model outputs is an important task. In this paper, ensemble canonical correlation prediction method developed by Lau et al. (2002) was applied to predict summer (JAS) rainfall over the United States. Variables used as predictors are the observational sea surface temperature (SST), sea level pressure for the Northern Hemisphere, soil moisture, rainfall for the previous winter in addition to the climate model forecast fields. The canonical correlation analysis (CCA) is performed for each single predictor separately. These predicted precipitation fields from different variables form an ensemble. Ensemble means can be constructed based on previous performance of each variable. If only observational data used, the ensemble forecasts with members consist of predicted rainfall based on SST, sea level pressure and soil moisture have higher skill than the CCA forecasts based on SST or sea level pressure alone. Different rainfall patterns are associated with the variations of SSTs and soil moisture. Therefore the ensemble mean of predicted rainfall improves the forecast skill. Because of no hindcasts available, the AMIP simulations were used to study the ensemble canonical correlation prediction method by adding model simulation from each model as a variable. There are AMIP simulations from 5 models. Overall, the CCA corrected model precipitation has higher skill than the uncorrected model precipitation, because the CCA corrects more than the mean field, it also corrects the leading modes of precipitation variations. The best forecasts are the ensemble means of members from CCA corrected model precipitation and precipitation predicted based on the observed SST, SLP and soil moisture. Each member has skill over the different parts of the United States. Therefore, the ensemble mean based on both model forecasts and observations together has higher skill than the CCA prediction made based on the model or observations alone.