We present an analysis of the large-scale atmospheric circulation variability since 1900 based on various circulation indices. They represent the main features of the zonal mean circulation in the northern hemisphere in boreal winter (such as the Hadley circulation, the subtropical jet, and the polar vortex in the lower stratosphere) as well as aspects of the regional and large-scale circulation (the Pacific Walker Circulation, the Indian monsoon, the North Atlantic Oscillation, NAO, and the Pacific North American pattern, PNA). For the past decades we calculate the indices from different reanalyses (NCEP/NCAR, ERA-40, JRA-25, ERAInterim). For the first half of the 20 th century the indices are statistically reconstructed based on historical upper-air and surface data as well as calculated from the Twentieth Century Reanalysis. The indices from all these observation-based data sets are compared to indices calculated from a 9-member ensemble of "all forcings" simulations performed with the chemistry-climate model SOCOL. After discussing the agreement among different data products, we analyse the interannual-to-decadal variability of the indices in the context of possible driving factors, such as El Niño/Southern Oscillation (ENSO), volcanic eruptions, and solar activity. The interannual variability of the Hadley cell strength, the subtropical jet strength, or the PNA is well reproduced by the model ensemble mean, i.e., it is predictable in the context of the specified forcings. The source of this predictability is mainly related to ENSO (or more generally, tropical sea-surface temperatures). For other indices such as the strength of the stratospheric polar vortex, the NAO, or the poleward extent of the Hadley cell the correlations between observations and model ensemble mean are much lower, but so are the correlations within the model ensemble. Multidecadal variability and trends in the individual series are discussed in the context of the underlying anthropogenic and natural forcings. While consistent trends were found for some of the indices, results also indicate that care should be taken when analysing trends in reconstructions or reanalysis data.