Forecasting science
How forecast accuracy improved
In 1985, a 3-day forecast was mostly guesswork. Today it's more accurate than a next-day forecast used to be. Here is what changed.
The metric that matters
Meteorologists measure forecast accuracy with 'skill' — how often the forecast beat a naive baseline like climatology or persistence.
In 1955, a 1-day forecast had about 60% skill. Today, a 5-day forecast has that same skill level. Forecast skill has increased by roughly one full day per decade for 40+ years — one of the great engineering achievements of the 20th and 21st centuries.
The turning points
1950s — First computer forecast
ENIAC ran the first numerical weather prediction in 1950. Crude but proved the concept.
1960 — TIROS-1
First weather satellite. Suddenly we could see storms globally instead of only where balloons launched.
1970s — Global data
World Weather Watch started. Data from every corner of the planet available in near real-time.
1980s — Ensemble forecasting
Instead of one forecast, run many with slightly different starting conditions. Uncertainty becomes measurable.
1990s — WSR-88D Doppler
Radar could see rotation. Tornado warning lead time jumped from 3 min to 11 min.
2000s — Regional high-res models
RAP, HRRR — updates every hour instead of every 6.
2010s — Warn-on-Forecast
Experimental: warn based on forecast rotation before radar sees it.
2020s — AI models
GraphCast, Pangu, Aurora. Machine-learning models are catching up to traditional physics-based models — and running 1,000× faster.
What still doesn't work
- Days 10+ still barely beat climatology on average.
- Convective initiation timing (exactly WHEN storms fire) — still hard.
- Ice storm placement — often bust.
- Snow amount forecasts — accurate for total but hard for exact location.
- Hurricane intensity change — dramatically improved but still imperfect.
The AI model era
Google DeepMind's GraphCast (2023), Huawei's Pangu (2023), and Microsoft's Aurora (2024) are machine-learning models trained on 40+ years of forecast data.
- They're not physics simulations. They're neural networks trained to predict future weather from present state.
- They already beat traditional models on some medium-range forecasts.
- They run in seconds instead of hours.
- They struggle with extreme events they haven't seen before.
- NWS uses them as inputs to human forecast decisions, not as the final product.
What is next
- Convective allowing forecasts out to 10 days.
- Warn-on-Forecast for tornado warnings.
- Sub-seasonal to seasonal (S2S) forecasts becoming useful for planning.
- Regional AI models trained on local geography.
- Continuous refresh instead of discrete run times.