AI Wildfire Models in California Rate Filings: Credibility, Territory, and Certification Challenges in H2 2026

California's CDI cleared AI wildfire cat models for rate filings in 2026, giving P&C actuaries access to parcel-level risk scores that diverge 3x to 5x from 20-year historical ZIP loss data in high-hazard zones. Carriers writing California homeowners now face a credibility problem that traditional actuarial methodology was not designed to handle.

The three models cleared through the PRID process (Verisk in July 2025, Karen Clark and Company in August 2025, and Moody's Analytics in August 2025) provide expected annual loss estimates rooted in forward-looking climate conditioning, fuel load dynamics, and parcel-level vulnerability functions. ZestyAI's Z-FIRE, which CDI recognized as the first AI model ever approved in a California rate filing, predates the PRID approvals and serves the market for risk segmentation and rate differentiation at the individual property level. Together, these tools give admitted carriers the forward-looking modeling infrastructure that California's backward-looking Prop 103 framework denied them for three decades. The actuarial methodology for converting that infrastructure into approvable rate exhibits does not yet exist in a codified form, which is the condition carriers are writing into as the first wave of SIS rate filings reaches CDI examiners in H2 2026.

What AI Wildfire Models Are Showing California Actuaries

Swiss Re's Sigma 1/2026 set the macro context: wildfire insured losses are growing at roughly 12% annually globally, and the January 2025 Los Angeles fires generated approximately $40 billion in insured losses, the largest wildfire loss event on sigma records. Those fires also produced the most comprehensive out-of-sample test of the AI wildfire models that California's admitted market had been waiting for. ZestyAI's Z-FIRE flagged 94% of the Palisades fire affected area and 87% of the Eaton fire affected area as high or very high risk before the events occurred, with more than 1.5 million California structures carrying high or very high wildfire perimeter risk scores in the model's current output (ZestyAI, 2025).

That validation performance is the source of both the opportunity and the actuarial problem. The model was right where the historical loss record was structurally silent: a 20-year loss history that does not include the Palisades or Eaton fire events cannot flag those areas as extreme tail-risk concentrations, because the extreme tail had not materialized within the observation window. This is not a data-quality problem; it is what Swiss Re's 12% annual wildfire loss growth rate implies. Each decade's historical window systematically excludes the upper tail of the next decade's actual loss experience, and backward-looking averaging of that window produces rates anchored to an understated expected cost.

The WUI housing stock concentration compounds the gap further. Housing units in California's wildland-urban interface increased 39% from 1990 to 2020 (Milliman, 2025), expanding the exposure pool faster than historical loss data could capture the risk distribution. AI models adjust for current exposure density and current fuel load. The historical average does not. The 3x to 5x divergence between AI model outputs and historical ZIP code experience in high-hazard zones reflects the accumulated effect of both the tail-exclusion bias and the exposure-growth gap. That divergence is the signal the filing actuary must certify and CDI examiners must evaluate.

The Credibility Blending Gap: When Forward-Looking Models Diverge from Historical Data

Traditional credibility theory provides a framework for weighting an insurer's own loss experience against a complement drawn from a broader population of similar risks. The complement earns weight because it is statistically more credible than thin individual experience, but it remains fundamentally the same type of data: observed historical losses from the same observation period, differing from the individual experience in sample size rather than in kind.

That framework was built for a world where the gap between an insurer's own experience and the industry complement is explained by statistical variance, not systematic directional divergence rooted in a different modeling philosophy. When a Verisk EAL calculation or a ZestyAI Z-FIRE score produces a loss indication for a high-hazard Sierra Nevada ZIP code that is four times higher than the insurer's own 20-year loss average for that territory, the gap is not statistical noise that credibility weighting resolves. It is the model's assertion that the historical record excludes events with material probability weight at that location, and that the 20-year average therefore understates the true expected annual loss for reasons that more data would not fix.

Standard credibility complements do not help here. An insurer with thin California WUI experience could supplement with industry loss experience in the same territory, but that industry experience carries the same tail-exclusion bias for the same structural reasons. It could supplement with a statewide average, but that average dilutes the geographic signal the model exists to provide. The credibility blending question in an AI wildfire model rate filing is qualitatively different: the actuary must decide how much weight goes to a forward-looking probability model and how much to a backward-looking loss record, when the difference between them is not variance around the same mean but a structural disagreement about what the mean is.

California's Regulation 2644.9 framework acknowledged this tension obliquely. Milliman's analysis of the regulation noted that supplemental data must be "as specific and appropriate to the insurer as possible" when used to complement credibility (Milliman, 2025). That standard was written for traditional credibility supplements, not for forward-looking probabilistic models whose outputs diverge from historical data for identifiable structural reasons unrelated to data volume. An actuary filing AI model-based indications in H2 2026 must build an explicit rationale for why model output deserves independent credibility weight rather than functioning as a supplement to the backward-looking record, and document why the historical experience is structurally understated rather than statistically thin. CDI's conditions of use require carriers to present model output alongside supporting EAL documentation, giving examiners the ability to evaluate credibility methodology across multiple filers simultaneously. Significant carrier-level divergence in credibility approach will generate examiner questions that did not exist in California homeowners ratemaking before this year.

Parcel Scores and Territory Relativities: The Aggregation Problem

AI wildfire models operate at the individual property level. ZestyAI's Z-FIRE system analyzes vegetation density, roof materials, defensible space, topography, and structural characteristics of individual structures across more than 1.5 million California properties it currently rates at high or very high wildfire perimeter risk (ZestyAI). The PRID-certified cat models from Verisk, KCC, and Moody's Analytics produce expected annual loss estimates derived from parcel-level vulnerability functions applied to stochastic event sets, then aggregated to produce a modeled loss cost by geographic unit.

California homeowners rate filings express geographic risk differentiation through territorial rating factors: multipliers applied to all properties within a defined territory, with within-territory variation captured through individual risk characteristics like construction type and fire-resistive building materials. The territories are drawn to capture meaningful geographic variation in expected loss, but they are necessarily coarser than parcel-level precision. Bridging the two requires an aggregation methodology that California rate filings have never needed to specify before.

The approach of averaging parcel scores within each territory and deriving a relativity from that average loses the distributional information that creates actuarial precision at the parcel level. Two territories with the same average EAL can have very different loss distributions: one where risk is uniformly moderate across all properties, and another where a concentrated cluster of extreme-risk parcels drives the territorial average upward while most properties carry moderate scores. Those two territories have different aggregate loss patterns, different reinsurance structures, and different probable maximum loss profiles, but a simple territorial average treats them identically in the rate exhibit.

The FAIR Plan's April 2026 rate filing illustrated the magnitude of model-based territorial redistribution in practice. Territory changes in that filing ranged from reductions near 78% in parts of the Central Valley to increases above 300% in the highest-hazard Sonoma and Sierra Nevada zones, the first prior-approval filing in California to produce territorial relativities of that scale (CDI, April 2026). CDI approved the filing. That approval established one data point that model-driven territorial restructuring of this magnitude can clear the prior-approval process. It does not establish a methodology template for admitted market carriers, because the FAIR Plan operates under distinct regulatory obligations and a distinct ratemaking authority that do not translate automatically to the SIS framework.

The actuarial methodology for aggregating parcel-level AI scores to defensible territorial rating factors, while preserving the distributional information that drives rate adequacy, is the kind of problem that typically requires actuarial guidance or CDI rulemaking to resolve in a codified form. Neither existed as of July 2026. The first carriers to file model-supported indications with explicit parcel-to-territory aggregation methodology documentation will set the practical template for subsequent filings by default, which concentrates both the compliance burden and the precedent-setting value on the earliest SIS filers.

Certifying Model Outputs When the Model's Inner Logic Is Proprietary

California's PRID process required CDI's actuarial staff to conduct a six-month evaluation of each submitted model's scientific grounding, output consistency, and documentation sufficiency for a prior-approval rate filing. That evaluation established a validated basis on which the filing actuary can rely when certifying that the model supporting the rate indication is actuarially sound under ASOP 38 (Use of Models Outside the Actuary's Area of Expertise) and ASOP 56 (Modeling). It also published a CDI Wildfire Catastrophe Model Checklist that specifies the documentation structure for model filings, giving admitted carriers a procedural template even where actuarial methodology remains uncodified (CDI, December 2025).

What the PRID evaluation did not establish is full technical transparency. Verisk, Karen Clark and Company, and Moody's Analytics each maintain proprietary elements in their model architectures: event set construction methods, vulnerability function parameterizations, and fire progression calibration procedures that constitute trade secrets protected outside the scope of the PRID review. The CDI evaluation examined output consistency and methodological defensibility at a documentation level. It did not require vendors to make the model's inner architecture publicly auditable.

ZestyAI's Z-FIRE presents the same pattern in a more acute form. The system operates through a deep-learning architecture applied to satellite and aerial imagery, drawing on records from more than 2,000 historical fire events to produce its property-level risk scores. CDI recognized Z-FIRE as "the first AI model ever approved as part of a rate filing" in California (ZestyAI), an approval that predates the PRID framework and was conducted under a different CDI review process for individual filings. The feature weights and model structure that produce a specific parcel's Z-FIRE score are not public documentation any filing actuary can independently replicate.

The ASOP 38 obligation in this environment requires the filing actuary to validate model outputs against independent test data rather than auditing internal model logic, since the full logic is not accessible. The ZestyAI LA wildfire validation performance (94% of Palisades affected area rated high or very high risk before the event; 87% for Eaton) provides exactly the kind of out-of-sample validation data that supports an ASOP 38 reliance opinion. Post-event calibration against the largest wildfire loss event on sigma records is a strong evidence base for output reliability.

The limitation of post-event validation is that it confirms retrospective calibration for events that occurred. It does not independently establish the model's forward-looking conditional probability estimates for current exposures given today's fuel load, climate trajectory, and defensible space conditions at specific parcels. The actuary must make that judgment with access to the model's documented outputs and validation results, but not its inner logic, and certify it under ASOP 38 and ASOP 56. A certified actuary can reconstruct a loss development factor selection from first principles given the triangle data; they cannot reconstruct ZestyAI's parcel score for a specific address without the model. That asymmetry is inherent to proprietary AI, and it is precisely the condition the California filing environment has formally created for the first time.

CDI's Prior-Approval Standard and the Sequential Rate Filing Trap

California Insurance Code Section 1861.05 requires CDI to find that a filed rate is not excessive, inadequate, or unfairly discriminatory before approving it. The Department's application of that standard consistently incorporates public interest considerations around rate shock, particularly in markets where homeowners face concentrated wildfire exposure and limited ability to migrate to another carrier. That constraint is not unique to wildfire; it is how California's prior-approval system operates across personal lines. Its effect in wildfire ratemaking is specific: AI models may produce indicated increases of 200% to 300% above current approved rates in the highest-hazard WUI zones, changes that are actuarially defensible but politically constrained by the same prior-approval process that governs every other indication.

In wildfire-exposed territories where rate filing indications have been reviewed over multiple consecutive filing cycles, the gap between actuarially indicated rate changes and CDI-approved levels has averaged roughly 35 to 40 percent. AI model-supported indications will face the same regulatory friction even when the underlying model methodology is defensible and the ASOP 38 certification is clean. The first five SIS filers (Mercury, CSAA, USAA, Pacific Specialty, and California Casualty) each filed 6.9% average statewide rate increases (CDI, 2025-2026), a figure that almost certainly reflects deliberate conservative structuring rather than the full AI model output for high-hazard territories. The Center for Climate Integrity estimated that approved 2026 rate requests could add approximately $1,015 to the average California homeowner's premium relative to 2023; the gap between approved and actuarially indicated levels is larger.

The sequential filing structure this creates is the defining constraint on the pace of market correction. A carrier that receives CDI approval for 60% of its AI model-indicated rate change cannot file for additional rate in most personal lines formats for at least 12 months. Over a period when wildfire losses are compounding at 12% annually and climate conditioning continues pushing EAL upward in WUI geographies, each filing cycle where approved rate trails indicated rate adds to the adequacy gap the carrier carries into the next event season. Filing in deliberate sequential steps, a conscious strategy rather than a failure, requires the actuary to structure the initial filing around the portion of the indication most defensible as a standalone rate level: defensible not just actuarially but on the public-interest balancing test CDI applies under Section 1861.05.

The Adverse Selection Mechanism for Model Adopters and Holdouts

Carriers that file AI model-supported rates early and obtain approval for a material share of their indicated change gain a pricing precision advantage that holdouts pricing off 20-year historical averages cannot replicate. When a PRID-certified cat model or ZestyAI parcel scores produce a materially higher EAL for a specific WUI territory, the filing carrier raises rates to reflect that EAL and in doing so prices the highest-hazard tail of that territory's risk distribution out of its admitted book at renewal. Lower-hazard properties within the same territory, where the model-indicated rate may be at or below the historical average, tend to remain. The carrier's book shifts toward lower-hazard risk within each territory as model-based pricing takes effect at renewal.

Holdout carriers pricing the same territory at historical-average-based rates write the full risk distribution without pricing distinction. Over successive development cycles, adverse selection concentrates the highest-hazard risks in the books of carriers least equipped to recognize them in the rate. That dynamic is structurally identical to what accelerated California's market exit sequence between 2022 and 2024, when non-renewals from State Farm (May 2023), Allstate (2022), and a succession of smaller carriers drove surplus lines homeowners policies from roughly 50,000 in 2023 to approximately 320,000 by end of 2025 (CDI). The FAIR Plan's residential structure coverage grew from $153 billion to $458 billion between 2020 and 2024, concentrated in WUI geographies where admitted capacity contracted fastest.

SIS adoption by six admitted carriers as of June 2026 is a beginning. It is not yet sufficient to set the market price for WUI risk across the full admitted book: the holdout population is still large enough that adverse selection remains an active pricing risk for early adopters who price out better risks in a territory while holdouts continue to compete for them at historical-average rates. That dynamic narrows as SIS adoption grows, which is precisely what the SIS writing commitment is designed to ensure by tying model use to a 85% market share writing obligation in wildfire-distressed ZIP codes. In the interim, carriers that have completed defensible model-supported filings hold a first-mover advantage that compounds as subsequent renewal cycles confirm the adequacy of model-indicated rates.

The Actuarial Stakes for H2 2026 Rate Filings

The first wave of AI wildfire model-supported rate filings to reach CDI examiners in H2 2026 will define the practical methodology templates that subsequent filers reference. Actuaries who explicitly document the credibility methodology for forward-looking model output (including a reasoned explanation for why historical experience is structurally understated rather than statistically thin), specify the parcel-to-territory aggregation approach and its distributional rationale, and build the ASOP 38 and ASOP 56 validation package from out-of-sample test data will produce filings that survive examiner scrutiny and establish the methodology precedent. Carriers that insert model EAL into a traditional indication formula and assert actuarial adequacy without methodology documentation will face CDI examiner questions that delay approvals and generate public rate hearing exposure for issues that could have been addressed in the filing exhibit.

The broader stakes are adequacy in the compounding sense. With wildfire losses growing at 12% annually (Swiss Re Sigma 1/2026) and AI model indications running 3x to 5x above historical averages in high-hazard territories, each year of sequential filing that leaves an adequacy gap is a gap the next event season will widen. The PRID approvals created the legal basis for accurate pricing. The actuarial work of the next 18 months will determine whether California's admitted carriers use that basis well enough to avoid repeating the market exit cycle they emerged from just two years ago.