This paper presents a verification study of the skill and potential economic value of forecasts of ice accretion on wind turbines. The phase of active ice formation has been associated with the strongest wind power production losses in cold climate, however, skillful icing forecasts would permit taking protective measures using anti-icing systems. Coarse- and high-resolution forecasts for the range up to day 3 from global (IFS and GFS) and limited-area (WRF) models are coupled to the Makkonen icing model. Surface and upper-air observations and icing measurements at turbine hub height at two wind farms in Central Europe are used for model verification over two winters. Two case studies contrasting a correct and an incorrect forecast highlight the difficulty of correctly predicting individual icing events. A meaningful assessment of model skill is possible only after bias correction of icing-related parameters and selection of model-dependent optimal thresholds for ice growth rate. The skill of bias-corrected forecasts of freezing and humid conditions is virtually identical for all models. Hourly forecasts of active ice accretion generally show rather limited skill, however, results strongly suggest the superiority of high-resolution WRF forecasts compared to other model variants. Predictions of the occurrence of icing within a period of six hours are found to have substantially better accuracy. Probabilistic forecasts of icing based on grid-point neighbourhood ensembles show slightly higher potential economic value than forecasts based on individual grid-point values, in particular at low cost-loss ratios, i.e., when anti-icing measures are comparatively inexpensive.