Predict the storm.
Protect the block.
A governed research environment built to evaluate storm prediction errors, identify dangerous local impacts, and test whether new variables improve decisions before the next event.
Historical storm explorer
Select any Atlantic tropical cyclone from 1950–2025. Map displays best-track observations.
Loading…
What makes forecasts right—or wrong?
Compare documented inputs, expected failure modes, and the data needed for a genuine apples-to-apples experiment.
| Candidate variable | Potential forecast value | Failure we would investigate |
|---|
Train a constrained, regularized mixture of available model predictions using past errors, lead time, geography and storm regime. Freeze weights before testing on later seasons.
Separate bad inputs, poor initialization, incorrect physics, track miss, intensity miss, drainage-model miss and missing local exposure.
Only use source runs completed before the decision deadline. Never fit to a storm's eventual damage, track or final intensity.
References: NOAA HAFS physics and coupling; NHC SHIPS/LGEM rapid-intensification research; NHC P-Surge. HAFS ↗ · P-Surge ↗ · Verification ↗
Blind-year benchmark
Each historical prediction uses only earlier seasons; adaptive weights also use only previous errors.
Break the forecast on purpose.
Replay a transparent persistence baseline with data corruption, lead-time and input perturbations.
Awaiting replay
Perturbation is applied to a naive persisted wind estimate. This does not recreate an operational numerical forecast.
One flooded street is still a disaster.
Coupling rainfall and coastal surge with drainage, road geometry, critical facilities and vulnerability is a separate high-resolution modeling problem.
—
Not a flood depth, probability, property-specific forecast or evacuation trigger. A real risk engine requires surveyed land elevation, drains, rainfall nowcasts, soils, river stages, tide gauges and buildings.