September regional ice extent gave no clear March Arctic forecast gain in 2000-2025

We tested whether September regional sea ice extent adds information about the next March Arctic total. The baseline uses year, preceding March total extent, and September total extent. We used Sea Ice Index Version 4 and 46 seasonal pairs from 1979-2024, with 26 March test targets in 2000-2025. Each forecast used earlier years only. A ridge model added 14 regional inputs after training-only baseline removal and scaling. Nested chronological tests selected the penalty. Added inputs raised root mean squared error from 0.326 to 0.335 million square kilometres. The mean squared error reduction was -5.4%, with a conditional 95% three-year block interval of -13.5% to +1.4%. A test ending before the 2025 sensor change gave a similar result. This benchmark found no clear forecast gain. It does not show absence of physical memory. Short records, a linear model, current data versions and conditional uncertainty limit the result. Code, source hashes, exact monthly inputs, all forecasts, an independent audit and figures are included.

Arctic · sea ice · regional extent · forecast validation · seasonal memory · ridge regression