NOAA's 69% record El Nino call lands where forecast models are weakest

Divide the record probability by the very-strong probability and you get the number nobody quotes: conditional on the very strong event arriving, at most about 77% of the probability sits on "bigger than anything measured since 1950" ^[1]. That comparison window is 76 years long ^[2]. The central case, in other words, is an event outside the range the instrumental record contains.

This is inconvenient for the forecasting method being pushed hardest at the places with the thinnest margin. AI systems do not integrate the physics; they learn a statistical mapping from the current state of the atmosphere to a later one ^[5], which makes training expensive and running the finished model cheap enough for ordinary hardware ^[6]. Models such as GraphCast do beat physics-based systems on speed and general accuracy ^[7]. General accuracy is an average over ordinary weather. In the tails, the Oxford comment cites studies finding that AI systems underestimate both the intensity and the frequency of record-breaking heat, cold and wind relative to a leading physics-based model ^[8], and the author notes their reliability has not been properly vetted for extremes under a changing climate ^[9].

The bias has a map. The author's formulation is that the data gap is most dominant precisely where the vulnerability gap is greatest ^[11]. Kenya's coastal counties carry the added exposure of storm surge and coastal erosion on top of the rainfall ^[4].

The recent record also suggests lead time was not the only thing missing. In September 2023 Storm Daniel collapsed two dams at Derna, Libya, killing at least 4,000 people in a single night ^[12]. In the summer of 2024 the Arba'at Dam in eastern Sudan burst, destroying 20 villages and affecting 50,000 people already living through a civil war ^[13]. That April the UAE logged its heaviest rainfall in 75 years and Dubai stopped ^[14]. A sharper five-day forecast helps an evacuation order; it does not raise a spillway. Kenya's own sequence of 1997-98, 2006-07, 2015-16 and 2023-24, each of which killed hundreds ^[15], puts the last El Nino flood season roughly two years back ^[3], which is recent enough that the list of unfinished drainage work still exists somewhere.

Two caveats on the material. It is a single expert comment from Oxford, which is itself building an AI forecast product, Forecast4Africa, with AfriClimate AI ^[16]; the warning about extremes is the author's own, which is more than most product literature concedes. And 69% leaves 31% ^[4]. An agency that designs only for a record season is designing for the wrong one about three times in ten, which argues for triggered thresholds rather than one planning scenario.