A Multimodel Approach for Improving Seasonal Probabilistic Forecasts of Regional Arctic Sea Ice

We formulate seasonal probabilistic forecasts of Arctic sea ice concentration from a multimodel (MM) ensemble constructed from six state-of-the-art climate models. Trend-adjusted quantile mapping is applied to postprocess individual model forecasts prior to MM combination, and a comparison is made against two benchmark MM ensembles: one uncorrected and another where individual models are adjusted for mean and trend bias. Focusing on September hindcasts over 2000-2015 initialized monthly from April-August, calibration improves forecast skill for all models tested, but greatest skill is achieved by the calibrated MM ensemble. Compared against a climatology adjusted for trends, skill is seen over most of the Arctic for all MM formulations and at all lead times, with highest skill obtained by the calibrated MM ensemble. Furthermore, despite the overall skill for each MM formulation, we find evidence of the "spring predictability skill barrier" for forecasts initialized before June.