Our forecasts, verified
No promises — measurements. Every day, yesterday’s forecast is compared with what the satellite actually observed. Here is the method, and the numbers — including the ones we miss.
No promises — measurements. Every day, yesterday’s forecast is compared with what the satellite actually observed. Here is the method, and the numbers — including the ones we miss.
100% of our "clean water" forecasts (calm season, 2026-10-08 → 2026-10-28) proved correct (over 1930); the rare alerts stay flagged low-confidence until the data confirms them.
Period: 2026-10-08 → 2026-10-28
In calm season there is almost no real beaching to confirm: this rate counts clean pixels wrongly flagged, not a 76% alert reliability.
100% of our "clean water" forecasts (calm season, 2026-10-08 → 2026-10-28) proved correct (over 1930); the rare alerts stay flagged low-confidence until the data confirms them.
From October 8 to October 28, 2026, 2,205 forecasts were checked against satellite observations across 21 beaches.
Accuracy is measured on our longest-running network — 20 Caribbean beaches (Martinique & Guadeloupe), where the verification archive is deepest. The exact same model (pipeline v3) powers the Punta Cana forecasts.
A forecast counts as a “hit” when the predicted status (clean, moderate, avoid) matches the status the satellite observed on the target day. Mean absolute error on the AFAI index: 0.016.
| Horizon | Hit rate | Checks | Displayed confidence |
|---|---|---|---|
| Day +1 | 99% | 420 | 60% |
| Day +2 | 98% | 399 | 60% |
| Day +3 | 99% | 378 | 56% |
| Day +4 | 99% | 357 | 50% |
| Day +5 | 99% | 336 | 44% |
| Day +6 | 100% | 315 | 40% |
The model also publishes its own confidence: high at day +1, low at day +6. When it is low, read the forecast as a trend, not a certainty.
When a beach is already loaded, we keep it 'red' until data (a clear satellite drop or a ground report) confirms it cleared: beached sargassum doesn't leave without collection. A conservative choice — it can make us over-alert on a beach that was cleaned in the meantime.
21 of 21 beaches score above 90%. The hardest ones to predict — published anyway:
Exposed beaches whose state hovers around a threshold (clean / moderate) remain the hardest to predict.