Statistical models for the seasonal prediction of iceberg severity in the Labrador Sea are developed from observational data for the years 1955-1984. The predictors are the monthly fields of sea level pressure and 700-mbar geopotential height over the region north is 15¿N. The primary statistical tools are gridded correlation patterns and multiple regression. Statistical significance is established by Monte Carlo simulations of the experimental procedure and by the use of a ''jackknife'' procedure for testing on independent data. Associations are found between iceberg severity and the circulation fields of the antecedent winter and spring. At lead times of 3--6 months, the mean absolute errors in the long-range forecasts of severity rank are approximately 4--5 (6--7) ranks in dependent (independent) samples of 30 years. These errors are significantly smaller than the corresponding chance level value of 9.6 (¿ &sgr;=1.2). The spatial patterns depicted by the predictor fields are dominated by the North Atlantic Oscillation and are synoptically compatible with eastern Canadian iceberg conditions in both extremes of severity.