Combination forecasting pioneered by Granger is an efficient way to diversify forecasting risk and thus to deal with model uncertainty. However, the empirical evidences have shown that the simple combining forecasts usually outperform the complex models due to the uncertainty and instability of the relationship between single forecasts. In this paper, the combination forecasting methods and model specification based on Stein rule shrinkage estimation and error correction mechanism are studied. The empirical results show that the perform ance of combining forecasts can be improved by adopting Stein rule estimators to combine non—sample information and sample information．Furthermore，the error correction models of combining forecasts are more accurate than other model forms． The error correction models proposed in this paper are more efficiency than previous models in"utilizing the co—integration relationship between non—stationary series and their forecasts，and therefore even better and more applicable．