A new version of the AWI Coupled Prediction System is developed based on the Alfred Wegener Institute Climate Model v3.0. Both the ocean and the atmosphere models are upgraded or replaced, reducing the computation time by a factor of 5 at a given resolution. This allowed us to increase the ensemble size from 12 to 30, maintaining a similar resolution in both model components. The online coupled data assimilation scheme now additionally utilizes sea-surface salinity and sea-level anomaly as well as temperature and salinity profile observations. Results from the data assimilation demonstrate that the sea-ice and ocean states are reasonably constrained. In particular, the temperature and salinity profile assimilation has mitigated systematic errors in the deeper ocean, although issues remain over polar regions where strong atmosphere-ocean-ice interaction occurs. One-year-long sea-ice forecasts initialized on January 1st, April 1st, July 1st and October 1st from 2003 to 2019 are described. To correct systematic forecast errors, sea-ice concentration from 2011 to 2019 is calibrated by trend-adjusted quantile mapping using the preceding forecasts from 2003 to 2010. The sea-ice edge raw forecast skill is within the range of operational global subseasonal-to-seasonal forecast systems, outperforming a climatological benchmark for about two weeks in the Arctic and about three weeks in the Antarctic. The calibration is much more effective in the Arctic: Calibrated sea-ice edge forecasts outperform climatology for about 45 days in the Arctic but only 27 days in the Antarctic. Both the raw and the calibrated forecast skill exhibit strong seasonal variations.

Accelerated loss of the sea-ice cover and increased human activities in the Arctic emphasize the need for skillful prediction of sea-ice conditions at sub-seasonal to seasonal (S2S) time scales. To assess the quality of predictions, dynamical forecast systems can be benchmarked against reference forecasts based on present and past observations of the ice edge. However, the simplest types of reference forecasts –persistence of the present state and climatology– do not exploit the observations optimally and thus lead to an overestimation of forecast skill. For spatial objects such as the ice-edge location, the development of damped-persistence forecasts that combine persistence and climatology in a meaningful way poses a challenge. We have developed a probabilistic reference forecast method that combines the climatologically derived probability of ice presence with initial anomalies of the ice-edge location, both derived from satellite sea-ice concentration data. No other observations, such as sea-surface temperature or sea-ice thickness, are used. We have tested and optimized the method based on minimization of the Spatial Probability Score. The resulting Spatial Damped Anomaly Persistence forecasts clearly outperform both simple persistence and climatology at sub-seasonal timescales. The benchmark is thus about as skilful as the best-performing dynamical forecast system in the S2S database. Despite using only sea-ice concentration observations, the method provides a challenging benchmark to assess the added value of dynamical forecast systems.