In this study, we use the three versions of AOGCM MIROC: MIROC3m, MIROC4h, MIROC5. In MIROC3m, the atmospheric and oceanic resolutions are a T42 spectral model with 20 levels on vertical σ-coordinates and an approximately 1° × 1° longitude-latitude grid with 44 vertical levels, respectively. MIROC4h includes the same model physics as MIROC3m but at an eddy-permitting resolution for the ocean: a T213 spectral model with 56 levels in the atmosphere and an approximately 1/4° × 1/6° longitude-latitude grid with 47 vertical levels in the ocean. In MIROC5, most parts of the model physics are updated or replaced with new parameterization schemes derived from MIROC3m and MIROC4h. The resolution of MIROC5 is T85 spectral with 40 sigma-pressure hybrid vertical levels in the atmosphere and approximately 1° horizontal grid of curvilinear coordinates with 49 vertical levels in the ocean. Details of the performance and settings of MIROC3m, MIROC4h, and MIROC5 are described inNozawa et al. , Sakamoto et al. , and Watanabe et al. , respectively.
 To evaluate the decadal climate predictability, we conducted the following three experiments using the three versions of MIROC: the 20th century climate simulation (the uninitialized simulation), the data assimilation, and the hindcast experiments (the initialized simulation). In the uninitialized simulation, the model was prescribed by historical natural and anthropogenic forcings, such as greenhouse gases and aerosol concentrations, solar cycle variations, major volcanic eruptions and future emission scenarios based on the CMIP3 and CMIP5 protocols. Using the model climatology defined by the uninitialized simulation, the observed temperature and salinity anomalies in the ocean [Ishii and Kimoto, 2009] were incorporated into model anomalies by an incremental analysis update scheme [Bloom et al., 1996] in the assimilation experiment. The observed temperature and salinity anomalies, obtained from the gridded monthly objective analysis covering the whole ocean produced by Ishii and Kimoto , were linearly interpolated for each day and the ocean model grid. From the assimilation experiment, we can first obtain a pair of atmospheric and oceanic initial conditions and then conduct an ensemble hindcast experiment with a total of 9 hindcasts initialized in 1961, 1966, 1971, 1976, 1981, 1986, 1991, 1996, and 2001 (hereafter, we mainly used the hindcast experiment initialized in 1996). The hindcast experiment has 10, 3, and 6 ensemble members for each of the starting dates in MIROC3m, MIROC4h, and MIROC5, respectively. Details of the model experiment and the assimilation procedure are described in the previous manuscripts [Tatebe et al., 2012; Mochizuki et al., 2012; Chikamoto et al., 2012a, 2012b]. To evaluate the robustness of our results, we also analyze near-term climate prediction experiments in the CMIP5 models. After removing the model drift that arose during prediction from each model, a multi-model ensemble in the initialized simulation is obtained by averaging over three ensemble means in each model in a similar manner obtained byChikamoto et al. [2012b]. The model drifts in SST and precipitation are estimated on the basis of the assimilation experiment in the MIROC multi-model but of the observation in the CMIP5 models including MIROC5 and MIROC4h. Therefore, precipitation anomalies are obtained in the MIROC models but not in the CMIP5 models because of few observed precipitation datasets covering the entire globe before 1979 owing to the lack of satellite data.