California's CDI cleared AI wildfire catastrophe models for rate filings in 2026, giving P&C actuaries parcel-level risk scores that diverge 3x to 5x from 20-year historical ZIP loss data in high-hazard zones. Three models cleared the PRID process, Verisk in July 2025 and Karen Clark and Company and Moody's Analytics in August 2025. The methodology for turning those scores into approvable rate exhibits does not exist in codified form.

Key Takeaways

  • Model output runs 3x to 5x above 20-year historical ZIP experience in high-hazard zones, a gap that credibility weighting was never designed to arbitrate.
  • Z-FIRE flagged 94% of the Palisades affected area and 87% of the Eaton area as high or very high risk before January 2025, against a loss record that could not flag either.
  • Wildfire insured losses are growing at roughly 12% annually, so each decade's observation window systematically excludes the upper tail of the next decade's experience.
  • The FAIR Plan's April 2026 filing moved territories from about 78% down to above 300% up. CDI approved it, which proves the scale can clear prior approval but sets no template for admitted carriers.
  • The first five SIS filers each filed 6.9% average statewide increases, well below what the models indicate for the highest-hazard territories.

What the Models See That the Record Cannot

Swiss Re's Sigma sets the context: wildfire insured losses are growing at roughly 12% annually globally, and the January 2025 Los Angeles fires produced approximately $40 billion of insured losses, the largest wildfire event on sigma records. Those fires were also the out-of-sample test the admitted market had been waiting for. ZestyAI's Z-FIRE flagged 94% of the Palisades affected area and 87% of the Eaton affected area as high or very high risk before either event, and currently scores more than 1.5 million California structures at high or very high wildfire perimeter risk.

The validation result is the source of both the opportunity and the problem. The model was right where the loss record was structurally silent: a 20-year history that does not contain the Palisades or Eaton fires cannot identify those areas as extreme tail concentrations, because the tail had not yet materialized inside the observation window.

That is not a data quality issue. It is what 12% annual loss growth implies. Each decade's window excludes the upper tail of the next decade's experience, and averaging that window produces rates anchored to an understated expected cost.

Exposure growth widens the same gap from the other side. Housing units in California's wildland-urban interface rose 39% from 1990 to 2020, expanding the pool faster than the loss record could describe its risk distribution. The models adjust for current exposure density and current fuel load; the historical average does not. The 3x to 5x divergence carries both effects at once.

Credibility Was Not Built for a Structural Disagreement

Credibility theory weights an insurer's own experience against a complement drawn from a broader population. The complement earns weight by being statistically more credible, but it is the same kind of data: observed losses from the same observation period, differing in sample size rather than in kind.

That is not the situation here. When a modelled expected annual loss for a high-hazard Sierra Nevada ZIP code comes in four times the insurer's own 20-year average for the territory, the gap is not variance around a shared mean. It is the model asserting that the historical record omits events carrying material probability weight, which is a disagreement about what the mean is, and more data of the same type does not resolve it.

The usual complements fail for the same reason. Industry experience in the same territory carries an identical tail-exclusion bias. A statewide average dilutes exactly the geographic signal the model exists to supply. Regulation 2644.9 asks that supplemental data be as specific and appropriate to the insurer as possible, a standard written for traditional complements rather than for forward-looking models that diverge for identifiable structural reasons.

So the filing has to carry an argument rather than a weight. The actuary has to establish why model output deserves independent credibility rather than functioning as a supplement, and document that the historical experience is structurally understated rather than statistically thin. Under ASOP 38 that reliance also has to rest on validation against independent test data rather than on auditing internal logic, because event set construction, vulnerability functions, and fire progression calibration remain proprietary in all three cleared models. The Palisades and Eaton results are strong evidence of retrospective calibration; they do not independently establish forward-looking conditional probabilities at a given parcel.

Aggregation is the second unsolved piece. The models score individual properties; California filings express geography through territorial factors. Averaging parcel scores within a territory discards the distribution, and two territories with the same average expected annual loss can hold very different loss patterns, one uniformly moderate and one with a concentrated cluster of extreme parcels. Those two territories carry different reinsurance needs and different probable maximum losses, and a territorial average treats them as identical.

The FAIR Plan's April 2026 filing showed the scale involved, ranging from reductions near 78% in parts of the Central Valley to increases above 300% in the highest-hazard Sonoma and Sierra Nevada zones, and CDI approved it. It operates under different ratemaking authority, so it establishes feasibility rather than method.

Approval Is Rationed, and the Gap Compounds

Section 1861.05 requires CDI to find a rate neither excessive, inadequate, nor unfairly discriminatory, and the Department applies that with public interest considerations around rate shock in markets where homeowners cannot easily move. Models producing indicated increases of 200% to 300% in the highest-hazard WUI zones meet the same prior-approval process as everything else.

The filed evidence suggests carriers are pricing to that constraint rather than to the indication. The first five SIS filers, Mercury, CSAA, USAA, Pacific Specialty, and California Casualty, each filed 6.9% average statewide increases, and in wildfire-exposed territories reviewed across consecutive cycles the gap between indicated and approved change has run roughly 35 to 40 percent. The Center for Climate Integrity estimated approved 2026 requests could add about $1,015 to the average homeowner's premium against 2023.

Sequencing is what turns that into a compounding problem. A carrier approved for 60% of its indicated change generally cannot refile in most personal lines formats for at least 12 months, while losses compound at 12% annually and conditioning continues pushing expected annual loss upward. Every cycle where approved trails indicated carries an adequacy gap into the next season.

Meanwhile the pricing advantage sorts the market. A carrier filing model-supported rates prices the highest-hazard tail of each territory out of its book at renewal while retaining the lower-hazard properties the model rates at or below the historical average. A carrier still pricing off the 20-year average writes the whole distribution undifferentiated and accumulates what the first carrier shed. That is the same mechanism that ran through the last exit cycle.

Surplus lines homeowners policies went from roughly 50,000 in 2023 to about 320,000 by the end of 2025 and the FAIR Plan's residential structure exposure grew from $153 billion to $458 billion between 2020 and 2024. The SIS framework ties model use to an 85% market share writing obligation in wildfire-distressed ZIP codes, which closes the gap only as adoption spreads.

Further Reading

Sources