Five-day Atlantic Hurricane forecasts in 2017 were as accurate as two-day forecasts in 1998 thanks to more powerful supercomputers and more precise models
Five-day forecasts are as good now as two-day forecasts were in 1998. — Back in 1998, the European Center for Medium-Range Weather Forecasts housed … Tweets: @arstechnica Tweets: @arstechnica : “...the location predicted by the hurricane center for a given storm was just 155 nautical miles away from the actual position of the storm 5 days later. What is incredible about this: back in 1998, this was the average error for a 2-day track forecast” http://arstechnica.com/...
Context & Ripple Effects
This 2018 data point marks the payoff of decades of compute-driven forecasting: a five-day Atlantic hurricane track error of 155 nautical miles at the European Center for Medium-Range Weather Forecasts had shrunk to what a two-day forecast missed in 1998. The arc since has only steepened — NOAA's $505M supercomputer upgrade tripled US modeling capacity in 2022, and AI entrants have begun competing directly with physics-based models.
By 2025, [[a:889534|Google DeepMind's Weather Lab was outforecasting traditional models on Hurricane Erin's path up to 72 hours ahead]], while Microsoft-backed Aurora claims accurate 10-day forecasts at smaller scales — meaning the supercomputer advantage this article documents is now being challenged and complemented by a different kind of model.
First-order effects
- Emergency managers and coastal residents gain actionable lead time: a five-day forecast as reliable as a 1998 two-day one effectively triples the warning window for evacuations and staging decisions.
Second-order effects
- Tech companies are now direct competitors to national weather agencies — DeepMind's Weather Lab and Aurora are positioning AI models as complements to, and benchmarks against, the supercomputer pipelines built by NOAA and ECMWF.
Third-order effects
- Forecasting is consolidating into hybrid stacks where physics-based supercomputer runs and AI models are blended, while NOAA research showing meteorologists still outperform algorithms in adverse conditions suggests the human forecaster remains the verification layer rather than being displaced.
The trend: Weather prediction is shifting from a supercompute arms race among national agencies to a hybrid regime where AI models from Google and Microsoft set the accuracy frontier and agencies must integrate or fall behind.