This thesis analyses to what extent the dynamic of the atmosphere is represented by decadal climate models. To capture this, two different methodological approaches are used. On the one hand, different teleconnection indices like the North Atlantic Oscillation (NAO) or the Southern Oscillation Index (SOI),in which the internal variability of the atmosphere is expressed, serve as an object of investigation.On the other hand, different classification procedures are used (based on correlation coefficients, cluster or principal component analyses) for the generation of circulation types. For this purpose, aspatial restriction on the North Atlantic and European area (90°Wto 40°E, 20°Nto 80°N) is carried out.To examine the decadal prediction skill, the global climate model of the Max Planck Institute (MPI-ESM) is used indifferent model generations (historical, baseline0, base-line1, prototype). These differ concerning their initialization technique as well as the number of available ensemble members. This work examines, based on different model generations, to what extent an initialization of the model as well as an enlargement of the ensemble have an impact on the prediction skill. The methodological approach is oriented towards the hindcast analyses, which are common in the decadal field of research. This implies,that with the aid of the decadal climate model retrospective decadal predictions for the investigation period (1979-2011) are provided and compared with observed values of the different reanalyses. The evaluation is not carried out for single predictions of a certain decadal period within the whole investigation period, but by so-called “lead years”. For this purpose, one year is picked out from all annually started predictions within the investigation period after the initialization of the model (e. g.,for "lead year 1” the first year of the prediction). To measure the prediction skill, different skill scores are used: Mean Squared Error Skill score (MSSS) and Ranked Probability Skill score (RPSS), a deterministic one as well as a probabilistic skill score. Moreover, the linear correlation coefficient (CORR) is used for the evaluation of the time series.The results show, that for the teleconnection indices as well as for the different classification procedures improvements can be observed in the prediction skill. They appear basically only in the first year after the initialization of the model, i.e. in the first lead year. The improvements can especially be observed for the initialized second (baseline1) and third model generation (prototype) of the decadal prediction model MPI-ESM, but not for the uninitialized model runs (historical) and the first model generation (baseline0). How-ever, the different teleconnection indices and classification procedures show big differences. On the one hand, in the spring (MAM) of the first lead year, i.e. for the third, fourth and fifth month after the initialization of the model, correlation coefficients near 0 occur for single teleconnection indices (e.g. NAO). On the other hand, values greater than r=0,7 can be observed for certain indices (e.g. SOI). Moreover, this thesis indicates that not only the implementing of an initialization on to the climate model positively affects the prediction skill. Rather, it can be proved that the rise of the ensemble number has a decisive influence on the improvement of the prediction skill. Nevertheless, it must be ascertained that the improvements can be shown primarily for the first lead year but cannot be observed for the whole decadal period .