The maximum signal-to-noise ratio empirical orthogonal function (MSN-EOF) is used to evaluate the midsummer 2 m temperature over Eastern China using subseasonal forecast data in the ECMWF model. The predictable sources of the most predictable components in the ECMWF model are analyzed, and the improvement of reconstructed predictions with leading predictable components is also investigated. ECMWF has the highest forecast ability among all S2S models but decays significantly after 10 d. The leading predictable mode mainly presents a dipole mode over Eastern China. Both the El Niño-Southern Oscillation and tropical Indian Ocean sea surface temperature anomalies are the main predictable sources of MSN-EOF1. A positive MSN-EOF1 is accompanied by El Niño decay, which can cause the strong and westward West Pacific Subtropical High, accompanied by anomalous southwesterlies over the north of Eastern China. The second predictable mode is the warmer characteristic of the Yangtze-Yellow River Basin, and the sea ice anomalies over the Barents Sea in the previous winter are the main predictable sources. This is accompanied by the wave train propagating from northwestern Russia to northeast Asia. The third predictable mode is mainly the temperature trend item extracted from the ECMWF model. The reconstructed predictions with the leading four MSN-EOF components show higher ability than the model raw predictions, therefore, this method can use the predictable components to filter the model noise and reduce the model error.