Bias-correction wrapper doubles an AI weather model's subseasonal accuracy

In February 2021, three back-to-back winter storms, including Uri, knocked out electricity and heat for millions of people in Texas and contributed to 246 deaths ^[1]. Gizmodo argues the toll might have been lower if utilities had been given longer to build up natural gas supplies, strengthen infrastructure and alert customers ^[17]. That extra time would have to come from two to six weeks out, where forecasts are weakest. Errors accumulate as models step forward, and both ECMWF's numerical models and Google DeepMind's WeatherNext struggle there ^[2].

The tool in the Nature paper is a post-processing wrapper called probabilistic bias correction, or PBC ^[3]. It takes the output of an existing model and uses machine learning to correct the errors that are systemic to that model ^[3]. Judah Cohen, a climatologist and research scientist at MIT, built it with Lester Mackey of Microsoft over more than a decade ^[5]. Cohen told Gizmodo: "It's model agnostic, and the paper shows that whether it's a more traditional model or a newer AI model, we can correct these errors that are systemic to the model" ^[4].

The work is a correction layer applied to forecasts that already exist; Gizmodo's headline describes it as AI improving extreme-weather warnings ^[16]. We think that design helps the operator: nobody has to choose between a numerical model and an AI model before using it.

The hit rate on the numerical model is high but not complete. Skill did not improve on 9% of pressure targets or 8% of temperature targets ^[14][15].

Cohen considers the extremes result the most significant finding ^[8]. "Interestingly, previous bias correction schemes actually made forecasts of extreme weather worse, and our model showed clear improvement of extreme weather forecasts," he said ^[13]. For forecasts made three weeks ahead, the wrapper improved the numerical model on extreme heat, extreme precipitation and unusually high sea level pressure ^[9].

Uri and its companion storms were winter events ^[1]. The three-week extremes Gizmodo lists are heat, rain and high pressure ^[9]. Gizmodo's account does not describe any utility using the forecasts.

The cost of looking is low. The code is open-source and does not need a supercomputer ^[11]. Cohen hopes it will lead to more actionable extreme-weather forecasts ^[12].

Whether to ask a forecast provider about this on Monday turns on two facts. The first is how many weeks ahead the action has to start: buying gas, reinforcing equipment, warning customers ^[17]. The second is whether the hazard is one the paper tested at that range. Heat, heavy rain and high sea level pressure at three weeks have results ^[9].

Crossing the two gives four cases. A short lead sits outside the window the paper addresses ^[2]. A lead of three weeks or more with one of the listed hazards justifies asking for a back-test on your own region. A long lead with an unlisted hazard leaves the provider to supply the evidence. Texas 2021, as Gizmodo tells it, sits in that last case, because winter storms are not among the three-week extremes it lists ^[1][9].