Track record
What Calchis keeps, and how its numbers have held up
Three records, all built from US federal sources and all public. The archive keeps what agencies said, as they said it, verifiable by hash. The ledger sets every warning against the survey that followed. The hindcast runs the live loss estimate over three decades of storms and shows where it was wrong. Nothing on this page is typed by hand.
The archive, as received
Since 2026-10-28, every product the platform has read from fifteen federal feeds, kept once per distinct body with the time it was issued and the time it arrived: 654 bodies across 296 polls. Append-only: the database refuses an update or a delete. A body can be fetched by its SHA-256 and hashed again. (Counts from the Data API; byte totals arrive with migration 216.)
| Source | Bodies | Polls | Size | First · last |
|---|---|---|---|---|
| NWS watches and warnings (six feeds) | 52 | 98 | — | — |
| USGS earthquake revisions | 311 | 46 | — | — |
| NHC current storms | 1 | 7 | — | — |
| NHC ATCF best-track files | 0 | 0 | — | no poll yet |
| SPC Day 1 outlook polygons | 5 | 46 | — | — |
| NIFC fire perimeters | 109 | 1 | — | — |
| NWS damage survey features | 167 | 15 | — | — |
| USGS high-water marks | 1 | 5 | — | — |
| USGS ShakeMap products | 4 | 47 | — | — |
| USGS ShakeMap contours | 2 | 10 | — | — |
| NHC wind swath bundles | 2 | 10 | — | — |
| NHC forecast cone bundles | 0 | 0 | — | no poll yet |
| NHC surge files | 0 | 0 | — | no poll yet |
| FIRMS hotspot rows | 0 | 11 | — | — |
| OpenFEMA declaration rows | 0 | 0 | — | no poll yet |
Counts: /api/v1/archive/summary. A body by its hash: /api/v1/archive/products/{sha256}. Two sources depart from “as received” on purpose and say so in the body: high-water marks lose surveyors’ names; OpenFEMA rows lose a timestamp the dataset restamps on every reload.
Warnings against surveys
Every tornado the Weather Service surveyed, against the warning in effect where it touched down: warned or not, how many minutes ahead, which office on each side, and how many homes were under the warning. Recomputed hourly as surveys are revised; method tornado-1. The record begins 2026-10-09.
- October 2026: surveyed
- 33
- warned
- 25 (75.8%)
- median lead
- 8 min
- homes under warnings
- 421,135
By office: RAH 9/9 · CAE 5/5 · CHS 5/5 · FGZ 1/3 · ABQ 0/2 · ILM 2/2 · MAF 0/2 · MHX 2/2 · GJT 0/1 · ILX 0/1 · TAE 1/1.
Replayed over 2020–2025 from the public archives (NWS damage surveys, the Iowa Environmental Mesonet’s warning polygon archive): 6,397 surveyed tornadoes, warned 56.1%, median lead 13 min, 60.4% with ten minutes or more; EF2 and stronger: 73.7% warned, median 15 min. Lead there runs from first issuance, the NWS measure; method tornado-iem-1, run 2026-10-29.
By year: 2020 54.4% (743) · 2021 55.8% (959) · 2022 55.1% (905) · 2023 50% (982) · 2024 58.6% (1550) · 2025 59.5% (1258).
Rows and summary: /api/v1/verification/tornado-warnings?month=2026-10. The readable report: case-studies/tornado-warnings/2026-10.
Modeled loss against what storms cost
The residential wind loss estimate Calchis freezes when a live storm comes ashore, run unchanged over every continental US hurricane strike 1996–2024 and set against NCEI’s total cost of each storm. 53 of 56 estimated. Modeled ÷ total cost, 2010–2024: 0.27 (p10–p90 0.09–1.44), n=23. Below one by construction; the spread is the signal.
Furthest below — surge and rain, which the model does not estimate
- Irene 20110.01×
- Sandy 20120.03×
- Katrina 20050.06×
Furthest above — small intense storms, size not in the best track
- Charley 20049.33×
- Dennis 20055.90×
- Bonnie 19984.50×
Every storm, both bands and what the misses say: methodology. Run 2026-10-28.
Live, as it happens
Each Atlantic storm that threatens the US gets a case study the moment a watch is posted, with the loss estimate frozen at the crossing before any reported figure exists, and the airport observations a day later. Case studies · severity forecast validation record.