NYU Abu Dhabi team proposes a records-only sea ice forecast as the baseline for seasonal models

RAP draws on one input. It reads the historical record of Arctic sea ice extent and nothing else, while the physics-based models it was measured against simulate the atmosphere, the ocean and the ice ^[6]. So any skill it shows has to come from patterns already present in that single series, with no wind or ocean data behind it ^[6]. Extent is worth forecasting because ice reflects solar energy that darker open water absorbs, and changes in the ice can shift atmospheric and ocean patterns well outside the Arctic ^[13].

Because RAP uses so little, the authors offer it as a benchmark for physics-based and AI-driven models alike ^[8]. "If a much more complex model cannot outperform such a simple approach, that tells us something important about how much additional predictive information that complexity is providing," said Francesco Paparella, inaugural director of Mubadala ACCESS at NYU Abu Dhabi and the study's senior author ^[10][11]. We think the logic holds. If a coupled model only ties a records-only baseline, it has not shown that its physics added information about extent.

The comparison needs a careful read. NYU Abu Dhabi's account says RAP's error on September extent was comparable to that of 34 models used for seasonal forecasting ^[1]. Paparella said the approach "performs competitively with much more complex forecasting models" ^[2]. Both describe parity with those models and stop there ^[1]. Nine months is the furthest the method reaches ^[3]. The release did not disclose error values, the lead time behind the September comparison, or how skill falls off as the horizon stretches toward nine months. A forecast that held up at two months and faded at eight would still fit every sentence in the announcement.

The cost case for planners is real but narrow. Paparella said the team sees potential for RAP "to support the UAE's growing polar and Arctic research activities by providing a simple, low-cost tool for seasonal sea ice forecasting" ^[14]. He was also clear about where it stops: "The value of RAP is not that it replaces more sophisticated models, but that it gives us a clear standard against which they can be tested" ^[9]. Matching a group of models on a September score is a benchmark result. The release describes forecasts of total extent ^[3]. We'd expect anyone routing ships or scheduling fieldwork to need the location of the ice edge as well.

The built-in uncertainty is the feature we'd expect to be most useful outside the lab. Each forecast comes with the spread of its ensemble ^[7]. A user can therefore see when the record holds similar past years that disagree about what came next. Paparella called sea ice forecasting "an increasingly important one as the Arctic continues to change" ^[12]. An analog method needs the past to contain states like the present. We would want to see how its spread behaves in the most recent seasons, where close analogs should be hardest to find.