Underestimated Extended Seasonal Hindcast Skill in Sparsely Observed Periods Revealed Through Hybrid Machine-Learning Initialization

The North Atlantic is a key source of seasonal-to-interannual climate predictability, as Subpolar Gyre (SPG) sea surface temperature anomalies (SSTAs), coupled with the North Atlantic Oscillation (NAO), modulate surface air temperatures over Europe and North America. However, model biases in North Atlantic dynamics and ocean–atmosphere coupling limit the skill of initialized hindcasts. While data assimilation partially constrains these errors using observations, hindcasts initialized during periods of sparse observational coverage may underestimate the true predictive potential of the system. Here, we reassess North Atlantic-driven extended seasonal predictability for the period 1960-2020 using a hybrid machine-learning (ML) assimilation approach, trained during periods with abundant observations (2004-2020) and applied to reconstruct North Atlantic Ocean temperatures during sparsely observed periods (1960-2004). Relative to standard initialization, the hybrid ML approach leads to stronger ocean–atmosphere coupling and a more robust NAO-like atmospheric response. As a result, we find enhanced winter and spring SSTA skill in the SPG during the first lead year in sparsely observed periods, along with improved surface air temperature skill over northwestern North America, southern Greenland, and central to northern Europe. Our results suggest that initialized prediction systems may systematically underestimate North Atlantic-driven predictability, and that initialization improved by hybrid ML can unlock greater forecast credibility than is implied by current standard hindcasts.