Right now
Active storms: the machine and the forecasters
Is it any good?
2017–2025: the auto estimate and the forecasters, against the final best track
The 2026 season
Every storm, auto vs official
Every storm has its own page: the auto estimate and the official intensity at each NHC, CPHC or JTWC update, what each voice read, how much the filter trusted it, and each voice's share of the number. Click a name, or browse all 2026 storms.
How it works
Voices. Only measurements and algorithms: the UW-CIMSS ADT (every storm, every basin, roughly every 30 minutes); CIMSS SATCON (from NHC's fix files, and from JTWC's prognostic reasoning, which quotes it); the new CIMSS machine-learning estimates (AIDT, D-PRINT, D-MINT) when JTWC quotes them; the climatlas satellite estimator on NOAA's GMGSI imagery (July 2021 on) and on GridSat-B1 (2017 to June 2021, calibrated separately); and aircraft — the larger of 0.9 × the 700-hPa flight-level wind and the SFMR surface wind, as reported in the f-decks. No subjective Dvorak number, no working best track, no agency advisory enters.
Calibration. Each voice is inverted through its measured response to aircraft truth (errors-in-variables: the final best track within 2 h of an aircraft centre fix, Atlantic and eastern Pacific 2017–2025), so a reading that saturates at the top of the scale is stretched back and its error bar widened, and each is anchored per basin to the final books of 2017–2024. Calibration for the backtest is fitted on odd years and scored on even years, and the reverse; the live system uses all years.
Filter. A damped-trend Kalman filter (the CBT's model) run forward only: every value on this page uses only what was known at that hour. The band is the filter's 90% interval. A storm is flagged when the official intensity sits outside that band for 12 hours or more.