Added winter SST cut northeast Atlantic forecast error in expanding fits for 1981-2025, but rolling fits were uncertain

We tested whether previous winter sea surface temperature (SST) adds next-winter forecast information after local summer SST and a linear time term. We used two fixed North Atlantic boxes, ERSSTv5 and HadISST1, and 45 test years from 1981 to 2025. Each fit used earlier years only. In the northeast Atlantic, added winter SST cut mean squared error by 19.19% in ERSSTv5 and 26.86% in HadISST1. Conditional 97.5% three-year block intervals were -0.03% to 33.09% and 7.24% to 41.96%. Sargasso gains were 15.30% and 11.01%, with intervals that included zero. With rolling 30-year training, northeast Atlantic gains fell to 6.87% and 6.42%, with uncertain signs. Thus the point gains depend on the training history. This is a forecast benchmark for known winter recurrence, not proof of heat storage below the surface. Revised data, shared source observations, two selected boxes, linear models, and uncertainty conditional on fitted forecasts limit the result.

North Atlantic · sea surface temperature · reemergence · seasonal forecast · training history