At NCAR, scientists are seeking to build predictive capability of the CCSM4 on decadal time scales by initializing the ocean and sea ice components with observed information, as a means to reduce errors associated with natural variability. A large set of hindcast experiments starting from different initial years in the past five decades have been carried out. A previous study (Yeager et al. 2012) shows that the hindcast experiments have tremendous decadal predictive skills in the North Atlantic subpolar gyre heat content, while the uninitialized simulations can’t simulate the correct tendencies due to errors in the initial states of ocean circulation variables.  In this study a more sophisticated statistical approach was developed in order to further evaluate the predictive skills in this prediction system.