For state insurance commissioners

An audit-defensible calibration credential under NAIC Bulletin 30.

The NAIC Model Bulletin on the Use of AI Systems by Insurers (December 2023) requires governance, testing, monitoring, and validation procedures for AI-driven decision systems. The Bulletin is silent on which specific calibration metric qualifies as adequate validation for catastrophe-model AI systems. That gap is filled today by vendor attestation — which is unverifiable by your office and unauditable in a rate-filing record.

What Forecast Registry provides. A neutral, third-party calibration credential standard: SHA-256 locked at forecast creation, OpenTimestamps-anchored on Bitcoin before resolution, scored by Brier and Brier skill score per peril class, preserved on an append-only public registry. Adoption requires no regulatory rulemaking — it requires only that your office cite the standard in rate-filing review questions and AI-disclosure templates as one acceptable form of model-validation credential.

What we built into the methodology for your office

Peril taxonomy keyed to your residual-market exposure

Atlantic basin tropical cyclone for FL/NC/SC/LA/TX. California wildfire for CA. Pacific basin for HI. European windstorm for international filings that reinsure your domiciliaries. Eight canonical classes total in methodology v1.0 §3.

Resolution sources already in regulatory practice

NOAA IBTrACS, CAL FIRE, USGS ANSS, NOAA SPC, USDA NASS, PERILS Industry Loss Index. You can verify Brier scores against the same datasets you already accept as authoritative.

Stale-model benchmark methodology

Methodology §4 specifies a sealed annual exceedance-probability snapshot that you can require from any vendor used in a rate filing — scored at end of policy year via our open scoring infrastructure. Vendors retain commercial confidentiality during the policy year; calibration credibility is established at year-end.

Conviction filter matches superforecaster reporting

|p − 0.5| > 0.20 for headline metrics. Hedging to the center is not rewarded. The full unconvinced set is preserved in audit and disclosed in the slice tuple.

What we are asking your office to consider

  1. Cite Forecast Registry methodology v1.0 in your next NAIC Bulletin 30 implementation guidance, AI-disclosure template, or rate-filing review checklist as one acceptable form of model-validation credential.
  2. Adopt the sealed annual exceedance-probability snapshot pattern (methodology §4) for vendor cat models used in domestic rate filings.
  3. Participate in the 60-day public comment period on methodology v1.0 — your office's input directly shapes the standards-board version of v1.1.

What we are not asking

Reference forecaster track record

Founding contributor's reference forecaster (Chorus Public Ledger) currently logs 1,963 SHA-locked predictions, 82 resolved, mean Brier 0.0399 on the 18-entry high-conviction in-headline forward subset (FiveThirtyEight state-level-call tier).

Per-peril retrodictive backfill is complete for the Atlantic basin: n=20 records (CSU April issuances 2020–2024 vs IBTrACS v04r01 realized), mean Brier 0.2002, BSS +0.199 vs climatology. The IBTrACS resolution dataset SHA-256 and the per-record breakdown are on the ratings page. Backfills for the other seven peril classes are scoped in the data-acquisition plan with explicit engineering and access-cost detail.

See the live ratings   Read methodology v1.0

Verified counterpart leadership (as of 2026-06-05)

Contact

, founding contributor. hello@forecast-registry.org (provisional). Methodology comments: comments@forecast-registry.org.

Last reviewed: 2026-06-05. Counterpart names verified against NAIC committee announcements, FLOIR commissioner profile, Citizens Property Insurance Corporation board page, and FHCF team page on review date.