Ice retreat dates reduced test error for Kara and Hudson ice advance in 2005-2024 at an 80% winter-extent threshold

We tested whether regional ice retreat dates improve later ice advance estimates beyond a time trend. We used the Sea Ice Index Version 4 regional workbook for 1989-2024. The threshold was 80% of each region's mean January-March extent in 1989-2004. Models used earlier years only for each test year in 2005-2024. Eleven regions had enough valid cycles. Adding retreat dates reduced mean squared error by 42.7% in Kara and 39.0% in Hudson. Three-year block intervals remained above zero under an approximate 11-region family check. Kara gains were also positive at 50% and 65% thresholds. Hudson intervals included zero at those lower thresholds. Beaufort had a detrended correlation of -0.791 but a 2.8% increase in test error with the linear model. These retrospective estimates use final data and dates that require a completed summer record. They do not establish local freeze-up, an operational forecast, or a heat-budget mechanism. The package gives daily evidence, all dates, code, source hashes, predictions, uncertainty limits and an independent numerical audit.

Arctic · sea ice · regional ice advance · forecast validation · threshold sensitivity · Sea Ice Index