Probabilistic forecasts of interannual September Arctic sea ice extent with data-driven statistical models

he widespread impacts of declining Arctic sea ice cover necessitate accurate and reliable predictions of Arctic sea ice. Up to now, much emphasis has been placed on either predictions at the sub-seasonal to seasonal timescales, or projections at the multi-decadal time scales, and less so on predictions at the seasonal to interannual time scales that are key for planning and infrastructure upgrade. Internal variability is a dominant source of uncertainty in predicting Arctic sea ice on seasonal to interannual timescales. However, initialized predictions conducted with dynamical climate models are of little use today, since these models exhibit biases and long-term drift that lead to poor skill beyond the seasonal time scale. In this study, we test and develop several statistical models in the form of transfer operators and neural networks to forecast probabilistic state transitions of the internal variability in Arctic September sea ice extent. Both the transfer operators and neural networks are trained on a large database of state transitions available from the CMIP6 archive. The models show comparable skill to other numerical and statistical models included annually in the Sea Ice Outlook for the predictions of September sea ice extent initialized in June, July, and August. While both statistical model types are able to make accurate and reliable predictions for many initialization months, the model performance is characterized by the spring predictability barrier and decreases for predictions initialized in March--May. The statistical models show skill beyond simple persistence when it comes to predicting sea ice extent trends at the interannual time scale. In particular, predictions initialized in July 2000 are able to reproduce the 2000-2010 accelerated decline in September sea ice extent, and predictions initialized in July 2012 capture the 2012-2024 slow-down in sea ice decline.