MIT-led team uses machine learning to sharpen weather forecasts weeks ahead

Weather forecasts are fairly reliable for about two weeks. After that, skill drops off sharply, a stretch MIT research scientist Judah Cohen calls a "predictability desert."

A team that includes Cohen, a Harvard College undergraduate and Microsoft researchers reports in the journal Nature that it has made real progress in that window, known as the subseasonal range, according to MIT News.

How it works

Rather than build a new forecast model, the researchers built a machine-learning layer that sits on top of an existing one. It studies the model's historical errors and shifts new forecasts to compensate. It works with probability-based forecasts and keeps updating as fresh data arrive.

The team says ordinary bias correction tends to nudge forecasts toward the average, which can make extreme events harder to predict. Their method was designed to avoid that.

What they found

• Applied to the European Centre for Medium-Range Weather Forecasts (ECMWF) model, it doubled accuracy in the two-to-six-week range.
• A combined entry using the method won an ECMWF forecasting competition last year.
• It improved land forecasts across most regions and other models it was tested on.
• Predictions of extreme weather improved by a statistically significant margin.

The tool was built on MIT's Engaging computing cluster but does not need a supercomputer to run, the team says.

Caveats

The MIT article does not say how big the extreme-weather gains were, and the method depends on long historical records that some regions lack. Real-world use is just beginning: Kenya's weather service has asked about applying it to monsoon forecasts, but no results have been reported.

Better outlooks several weeks out could help farmers, energy suppliers and emergency managers plan ahead.