Teaching a Model to Fix the Errors in Existing Two-to-Six-Week Forecasts

Forecasts beyond two weeks are notoriously unreliable, which leaves a tight window for energy operators, farmers and emergency response officials who need to get fuel, water and crews into place before the weather arrives.

A team including MIT researchers has built a machine-learning model that doesn't forecast from scratch. It learns to correct the errors in existing forecasting models and improves their output over the two-to-six-week range. Applied to a gold-standard forecasting model, the technique doubled forecasting accuracy over that window and finished first in a prominent international forecasting competition.

The distinction is worth holding onto: the improvement is measured against systems agencies already run, so this is a correction layer riding on top of established forecasts rather than a new model competing with them.