Probabilistic Forecasts of September Arctic Sea Ice Extent at the Interannual Timescale With Data‐Driven Statistical Models

• Published in: Journal of geophysical research. Machine learning and computation Vol. 2; no. 3
• Main Authors: Hoffman, L., Massonnet, F., Sticker, A.
• Format: Journal Article
• Language: English
• Published: 01.09.2025
• DOI: 10.1029/2025JH000669
• ISSN: 2993-5210

The widespread impacts of declining Arctic sea ice necessitate accurate and reliable predictions. While much focus has been placed on subseasonal to seasonal forecasts or multidecadal projections, seasonal to interannual predictions—crucial for planning and infrastructure—have received less emphasis. Internal climate variability is a dominant source of uncertainty on these timescales, yet initialized dynamical climate model predictions have limited usefulness due to biases and long‐term drift that leads to poor skill beyond seasonal timescales. This study develops statistical models—transfer operators (TO) and neural networks (NN)—to forecast probabilistic state transitions of Arctic September sea ice extent (SIE) internal variability. Trained on 24,420 transitions from the CMIP6 archive, these models make accurate and reliable predictions across multiple initialization months. At interannual timescales, they outperform simple persistence in predicting SIE trends. At seasonal timescales, their skill is comparable to other numerical and statistical models in the Sea Ice Outlook. While TO performance declines for spring initializations, NNs incorporating information about the area of thick ice can overcome the spring predictability barrier for March–May initializations. For the next decade, the TO suggests that September SIE will likely remain above the projected forced trend, while the NN predicts it will likely be lower. However, both models predict that the trend in September SIE will likely be higher (65%–96% chance) than the CMIP6 projected forced trend over the next 3 years, suggesting near‐term stability. These results highlight the potential of statistical approaches for improving Arctic sea ice predictions on critical planning timescales. The year‐to‐year variability in Arctic sea ice cover is challenging to predict, but is of great importance due to its widespread environmental, geopolitical, and logistical impacts. We show that statistical models can make skillful and reliable predictions of Arctic sea ice extent (SIE) by learning from variability in the sea ice state simulated by climate models. We compare the performance of two different statistical models in hindcasting September Arctic SIE at seasonal timescales, for known years of rapid and slower sea ice loss, and for years of extreme sea ice loss. These models are then used for future forecasts, where they predict an anomalously high Arctic SIE over the next 3‐year period. However, the different model types disagree in predictions over the next decade: one model predicts Arctic SIE will be anomalously high, and one predicts it will be anomalously low. This work is a novel example of statistical models providing skillful and reliable predictions of Arctic sea ice on timescales further out than one season. Statistical models can skillfully and reliably predict Arctic September sea ice extent (SIE) state transitions with a time lag of up to 5 years A reliable transfer operator predicts an 86% chance that September SIE will exceed the forced trend for the 2024–2028 period A neural network using information about the area of thick sea ice shows skill when initialized before the spring predictability barrier