Model registry
Only registered model versions publish numbers. Each version’s file is recorded in the registry with its SHA-256 hash, so any change to a model shows, and every promotion and retirement is listed with its reason.
Read from the test results bound to each published file.
Combines where large earthquakes happen in the long run with how they cluster after recent ones, plus GEAR1’s long-term rate from crustal strain, counted most where recent earthquakes are few. On 2023–2025 (a second look at those years), it ranked a cell that went on to have an M6+ earthquake above one that did not 97% of the time across the globe, and 77% among cells with recent earthquakes, the harder question.
For storms the National Hurricane Center covers, we publish NOAA’s own DTOPS guidance: in a test written down in advance and run once on the 2025 season it beat NOAA SHIPS-RII, NOAA RIOC consensus and our previous model (v8.2), ranking a forecast cycle that went on to intensify rapidly above one that did not 93% of the time. Elsewhere we publish our v8.3 model. Beside it, our satellite model J1 is recorded and, for West Pacific and South Indian storms, shown while it is tested on new forecasts. Our v10.1 model is shown beside DTOPS, labelled experimental, while it is tested on new forecasts. Our track and intensity forecast, TC1, is shown beside NHC’s official forecast.
Scores every thunderstorm NOAA tracks over the contiguous US for the chance it produces a tornado in the next hour. On every storm of 2025, scored once, it ranked a storm that went on to produce a tornado above one that did not 97% of the time.
Published: produces the number on the site. Fallback: takes over when an input feed is down. Experimental: shown beside the published number while it is tested on new forecasts. Shadow: recorded on every live forecast and never shown as one; listed here when its test could change what is published.
| Model version | Forecasts | Status | Test | AUC [95% CI] | Skill | What it is |
|---|---|---|---|---|---|---|
eq_operational_S2_gear1_v2-ecb87f020f99 | Earthquake (30 days, M6+) | published | 2023–2025 | 0.969 [0.964, 0.973] | IG +2.61 nats per M6+ cell-window | Boosted trees on long-term seismicity and aftershock-style clustering, plus GEAR1’s strain-rate model |
hurricane_ri_stack_v1-35938d6c09b7 | Hurricane RI (NHC areas) | published | 2025 | 0.928 [0.890, 0.966] | BSS +0.311 | NOAA DTOPS as issued (SHIPS-RII fallback) |
hurricane_ri_v10-8a5ba5896731 | Hurricane RI (NHC areas, our model v10.1) | experimental | 2022–2025 (each season out of sample) | 0.930 | LL 0.144 vs DTOPS 0.165 | LightGBM exceedance curve; in prospective verification, shown beside DTOPS |
hurricane_ri_v9-ce02551062e7 | Hurricane RI (NHC areas, v9.1) | shadow | 2026 (read once) | 0.924 | LL 0.132 vs DTOPS 0.150 | LightGBM on the early guidance, gated; claim not met on 2026. Its prospective test decides whether it replaces DTOPS as the published number |
hurricane_ri_v8_3 | Hurricane RI (other basins) | published | 2025–2026 (registered test; the basins where it publishes, unseen cycles, best-track inputs) | 0.852 | LL 0.1659 vs v8.2 0.1656 | GBT + logistic ensemble, Newton-calibrated; v8.2’s recipe on de-duplicated rows, adopted by amendment 15’s rule |
hurricane_ri_j1-6ffdec7356c1 | Hurricane RI (other basins, our satellite model J1) | shadow | 2024–2025 (registered hindcast, every JTWC cycle pooled) | 0.898 | LL 0.1336 vs v8.2 0.1412 | LightGBM on v8.2’s probability and inputs plus GMGSI satellite infrared; recorded beside the published v8.3, shown for West Pacific and South Indian storms. Its prospective test against v8.2 decides whether it replaces v8.3 as the published number |
hurricane_tc1-589cd13ad9b6 | Hurricane track and intensity, to 5 days (NHC areas) | published | 2023–2025 (pre-registered) | — | track 90.7 n mi vs NHC 92.1 | Online-weighted consensus of NHC’s real-time guidance; shown beside NHC’s official forecast |
tornado_v3-05a06c843c87 | Tornado within 60 min | published | 2025 | 0.970 [0.964, 0.976] | BSS +0.107 | LightGBM, 339 trees, 30 inputs |
tornado_v3-9005905f08b7 | Tornado within 60 min, when the NWS feed is down | fallback | 2025 | 0.969 [0.963, 0.975] | BSS +0.097 | LightGBM, 339 trees, 28 inputs |
tornado_v3-c45167a49056 | Tornado within 30 min | published | 2025 | 0.974 [0.968, 0.979] | BSS +0.084 | LightGBM, 301 trees |
tornado_v3-58208efc4bb4 | Tornado within 90 min | published | 2025 | 0.968 [0.961, 0.974] | BSS +0.111 | LightGBM, 312 trees |
tornado_v3-a4413fadb32c | EF2+ tornado within 60 min | published | 2025 | 0.994 [0.989, 0.997] | BSS +0.061 | LightGBM, 371 trees |
IG: information gain over a uniform map, in nats per M6+ cell-window: per 2° cell and 30-day window in which an M6+ earthquake occurred (higher is better). BSS: Brier skill score (above 0 beats the average rate). LL: log loss (lower is better).
Changes since the current testing programme began, each with the result that decided it.
| Date | Change | Model version | Reason |
|---|---|---|---|
| Oct 2026 | promoted | tornado_v3-05a06c843c87 | Pre-registered programme, 2025 scored once: AUC 0.970 [0.964, 0.976]; against NOAA ProbTor +0.031 [+0.022, +0.042] (paired by day) |
| Oct 2026 | retired | tornado_storm_v1_0 | Published from March 2026 until v3 replaced it on 3 Oct 2026; its live record is on the track record page |
| Oct 2026 | superseded | tornado_gbt_v2 | Overtaken by v3 before it published: on 2025, v3 − v2 +0.007 [+0.003, +0.010] AUC on the same storms and event |
| Oct 2026 | retired | hp-tornado-coherence-v1 | Its published AUC of 0.894 was measured on a 5:1-downsampled split with labels on the wrong clock (docs/AUDIT_2026-10-01.md) |
| Oct 2026 | promoted | eq_operational_S2_gear1_v2-ecb87f020f99 | Adds GEAR1’s long-term rate to C0, weighted by how active each cell is, decided by a test written down in advance on 2021–2022: +0.014 [+0.002, +0.026] nats per M6+ cell-window over GEAR1 with one weight; on 2023–2025 (a second look at those years) +2.61 nats per M6+ cell-window, AUC 0.969 |
| Oct 2026 | superseded | eq_operational_S1_gear1_v1-d409c534bb4c | GEAR1 with one weight; replaced by the activity-weighted stack. It scored +2.60 nats per M6+ cell-window on the same test |
| Oct 2026 | superseded | eq_operational_C0_v1-d28c62a2e35b | Now the base of the published model rather than published itself; it scored +2.59 nats per M6+ cell-window on the same test |
| Oct 2026 | retired | eq_coherence_v1_0 | Replaced by the pre-registered operational forecast (docs/EARTHQUAKE_FORECAST_PROGRAM.md); its live record is on the track record page |
| Oct 2026 | promoted | hurricane_ri_stack_v1-35938d6c09b7 | NOAA DTOPS published for the NHC areas: 2025 season scored once, AUC 0.928; beat NOAA SHIPS-RII, NOAA RIOC consensus and our previous model (v8.2) |
The full registry, with file hashes (JSON) · The same, through the API
