Adding winter SST raised error in 2002-2025 October-December linear forecasts in three North Pacific regions

We tested whether winter sea surface temperature (SST) improves October-December forecasts after summer local SST, winter and summer Nino 3.4, and a linear time trend are known. We used three fixed North Pacific regions and NOAA OISST v2.1 and ERSSTv5. Each fit used earlier years only, with training from 1982 and 24 test years in 2002-2025. Adding local December-February SST raised mean squared error in all six product-region tests. In the Northeast Pacific box, error rose by 43.81% in OISST and 47.04% in ERSST. Conditional 95% three-year block intervals for these increases were 6.03% to 164.21% and 3.28% to 179.92%. Three-region adjusted intervals also stayed above zero for this box. The other regions had higher point errors but less certain changes. Removing any one scored year left higher error in all six tests. In a check that can use future training years, the Northeast Pacific error increase was smaller and uncertain. A large early winter coefficient and a large 2002 error show sensitivity to the training record. These results do not show that ocean memory is absent. They apply to this target season, these linear predictors, and revised observations. The two products share source observations. Short records and uncertainty conditional on fitted forecasts limit inference. The package includes fixed monthly evidence, source hashes, all forecasts, uncertainty draws, code and checks.

North Pacific · sea surface temperature · ocean memory · seasonal forecasts · forecast validation · reemergence · OISST · ERSST · conditional uncertainty