Nine-year February fits cut September Antarctic ice test error by 40% against a mean in 2004-2024, with uncertain gain

We tested whether short rolling fits learn the February-September Antarctic sea ice relation in time to improve later forecasts. We used Sea Ice Index Version 4 daily data for 1989-2024 and 21 chronological forecasts for 2004-2024. A nine-year February fit reduced mean squared error by 40.1% against a mean from the same training years. RMSE fell from 0.641 to 0.496 million square kilometres. The conditional 95% three-year block interval for MSE reduction was -4.2% to 64.5%. The point gain was 62.6% in 2016-2024 and 29.9% when 2023 was excluded from scoring. Adding February to a nine-year trend gave a smaller 9.7% reduction, with an interval of -28.1% to 38.5%. All 12 full-period model comparisons had intervals that included zero. This extends prior campaign correlation and expanding-forecast tests with a test of short-window adaptation. It does not establish a reliable forecast advantage, physical cause, or regime shift. Short records, changing conditions, final data versions, and conditional uncertainty limit the result. Code, monthly inputs, source hashes, every forecast, an independent audit, and a figure are included.

Antarctic sea ice · seasonal prediction · rolling regression · forecast baseline · uncertainty