A statistical model is developed to reconstruct atmospheric surface data for the period 1902-1998 to force a coupled sea ice-ocean model of the Baltic Sea. As the response timescale of the Baltic Sea on freshwater inflow is of the order of 30-40 years, climate relevant model studies should cover at least century-long simulations. Such an observational atmospheric data set is not available yet. We devised a statistical model using a "redundancy analysis'' to reconstruct daily sea level pressure (SLP) and monthly surface air temperature (SAT), dew-point temperature, precipitation, and cloud cover of the Baltic. The predictor fields are daily SLP at 19 stations and monthly coarse gridded SAT and precipitation available for the period 1902 to 1998. The second input is a gridded atmospheric data set, with high resolution in space and time, based on synoptic stations, which is available for the period 1970-2001. Spatial patterns are selected by maximizing predictand variance during the "learning'' period 1980-1998. The remainder period 1970-1979 is used for validation. We found the highest skill of the statistical model for SLP and the lowest skill for cloud cover. For wintertime the dominant modes of variability on the interannual to interdecadal timescales of the reconstruction are discussed. It is shown that the wintertime variability of SLP, SAT, and precipitation is related to well-known atmospheric patterns of the Northern Hemisphere: the North Atlantic Oscillation, the Scandinavia pattern, the East Atlantic/West Russia pattern, and the Barents Sea Oscillation.