This study examines the level and origin of seasonal forecast skill of surface air temperature in northern Europe. The forecasts are based on an empirical methodology, canonical correlation analysis (CCA), which is a method designed to find correlated patterns between predictor and predictand fields. A modified form of CCA is used where a prefiltering step precedes the CCA as proposed by T.P. Barnett and R. Preisendorfer. The predictive potential of four fields is investigated, namely, (a) surface air temperature (i.e., the predictand field itself), (b) local sea surface temperature (SST) in the northern European area on a dense grid, (c) Northern Hemisphere 700-hPa geopotential height, and (d) quasi-global SST on a coarse grid, The design is such that four contiguous predictor periods (of 3 months each) are followed by a lead time and then a single predictand period (3 months long). The shortest lead time is 1 month and the longest is 15 months. The skill of the CCA based forecasts is estimated for the 39-yr time period 1955-93, using cross-validated hindcasting. Skill estimates are expressed as the temporal correlation between the forecasts and the respective verifying observations. The forecasts are most skillful in the winter seasons with a secondary weaker skill maximum during summer. During winter the geopotential height field produces the highest skill scores of the four predictor fields. The dominant predictor pattern of the geopotential height field is confined to the predictor period that is closest to a preceding core winter season and resembles the North Atlantic Oscillation (NAG) teleconnection pattern. The time series of the expansion coefficients of this dominant predictor pattern correlates well with a low-pass filtered rime series of an NAO index. The obtained skill is similar to what is found in the United States, both with regard to seasonal distribution and level of skill. The origin of skill is however different. In the United States it is the El Nino-Southern Oscillation (ENSO) with its predominantly interannual character that is the main source of skill in winter. In northern Europe it is instead the NAO that contributes the most, and especially the lower frequency part of the NAO (periods between 4 and 10 yr). Spatially sparse station data of surface pressure extending back to the middle of the nineteenth century suggests a nonstationarity in the NAO behavior. The implications of this nonstationarity for the obtained results of this study is briefly discussed. Because finely resolved field data are not readily available for this earlier period, the level of skill realizable for that period using a pattern relationship technique such as CCA remains an open question.