A data-driven forecasting method called kernel analog forecasting (KAF) is applied to the challenging and relevant problem of Arctic sea ice prediction on seasonal and regional scales, using data from the CCSM4 climate model. KAF encodes climate dynamics information into its prediction, providing a promising avenue to predict changes due to natural variability seen in a multi-century control climate model simulation.
KAF predictions in many cases show superior predictive skill over a benchmark persistence forecast, while also exhibiting limitations at known predictive barriers in Arctic sea ice. KAF also shows promise in being able to predict an unobserved quantity (e.g. predict sea ice thickness while only observing current sea ice concentration).
Analog forecasting is the idea of identifying a similar climate state (analog) in a historical record to a current climate state, and making a forecast past on the analog's future state. KAF extends this idea to identify an ensemble of historical analogs, weighted by a kernel function in such a way as to give preference to analogs with similar dynamics. Moreover, an out-of-sample extension technique is employed in the method that takes into account similarity between states at different scales. This method is applied to Arctic sea ice anomalies from a CCSM4 pre-industrial control simulation, utilizing ocean, atmosphere, and sea ice variables in the predictions. By refining to seasonal and regional scales, we can identify when and where we have predictive skill, and in many times exceed a benchmark damped persistence forecast by 3-6 months. Connections to sea ice reemergence phenomena are also explored, which aid KAF predictive skill. Pushing the method to its limits, we explore predicting an unobserved quantity, sea ice thickness, while only observing sea ice concentration, finding limited success.