Merced Property and Casualty Company was founded in 1906 by a group of Merced County farmers who could not buy fire insurance because the area had no firefighting infrastructure.1 By the time the Camp Fire ignited on the morning of November 8, 2018, the company had been writing California homeowners' coverage for one hundred and twelve years. AM Best held it at A-.1 Twenty-five days later, on December 3, the Superior Court of California, County of Merced entered an Order Appointing Liquidator following an expedited petition by the California Department of Insurance filed on November 30.2 Merced was insolvent. One hundred and twelve years of underwriting cleared in less than a month from a single event.
That detail belongs in the front of every regional carrier CEO's mind right now, before the AI vendors come in with their next presentation about how machine learning will sharpen the rating curve and grow the book.
The Mechanism, Stated Plainly
Camp Fire was the deadliest and most destructive wildfire in California history at the time it burned. Eighty-five people died. Nearly 19,000 structures were destroyed, most of them in the town of Paradise.3 Total insured losses from the Camp Fire alone were estimated by Munich Re at approximately $12.5 billion.3
Merced's exposure to Paradise was concentrated. The carrier had only about 200 policies in Butte County, but it had built that book under a rating model and capital base that did not assume the entire town of Paradise would be a total loss in a single event.4 The company had cat reinsurance, the cat reinsurance attached and paid, and the loss above the treaty layer plus the retention below it exceeded the carrier's surplus. There was no second tower of capital to call on. The Department of Insurance moved quickly because the math was settled before Thanksgiving.
The specific mechanism is worth holding in mind. A long-tenured, well-rated carrier failed in less than a month from a single peril that hit a geographically concentrated book hard enough to overwhelm the combined ceiling of retention plus surplus. No fine, no regulatory finding, no reserving misstatement appeared anywhere in the chain. The reinsurance ran, and the surplus underneath it was not built to absorb what was left on its own.
What AI Rating Models Do to Concentration
Now overlay AI on the same mechanism.
An ML rating model reads historical loss experience and proposes prices. Its job is to identify under-priced risks the prior actuarial framework was missing. When it finds a cluster of zip codes whose recent loss history looks favorable relative to the rate currently being charged, it lowers the indicated rate. The book grows in that cluster because the rate is competitive, the agent appointment network notices, and the binds compound month over month.
The carrier's overall loss ratio looks fine through the soft years. The growth shows up in the board pack as a win. The geographic concentration line on the cat report ticks up, and the cat report's own model, also trained on historical experience, tells the underwriting committee that the aggregation is within the cat XL treaty's modeled return period.
The training data is the part of this picture that ages quickly. Per Aon's 2023 climate report, secondary perils — the wildfires and severe convective storms and atmospheric rivers that don't carry hurricane names — accounted for roughly 86% of global insured natural catastrophe losses in 2023, with primary perils (tropical cyclone, earthquake, European windstorm) making up only 14%.5 Models that learned from a 2010-2020 loss window were trained on a climate that no longer fully exists. A model can be technically correct on the data it was given and still systematically understate the aggregation exposure being built underneath it.
In California specifically, the longstanding interpretation of Proposition 103 has restricted the use of forward-looking catastrophe models in rate filings. Carriers writing in the state have had to rely on twenty-year historical loss experience for prior-approval rate-making, even when their internal underwriting models can see a different distribution forming.6 (Insurance Commissioner Lara's Sustainable Insurance Strategy is reforming this, conditioning the use of forward-looking cat models on insurer commitments to write more in wildfire-distressed areas, but the legacy regime is what shaped the books on the ground today.)
The Scenario Worth Walking Through
Picture a California-focused homeowners writer with $350M in direct premium. The carrier's AI rating model, refit annually, identifies a Sierra foothill cluster across a handful of zip codes where recent fire experience has been benign. The model recommends a competitive indication. The book in that cluster grows from 4,200 policies in 2023 to 11,500 policies by mid-2026. The cat XL treaty was placed on the prior aggregation profile.
The 2026 fire season opens dry. A first event in late August attaches the cat treaty and runs through the first two layers cleanly. The recovery process begins. A second event in October hits the same Sierra foothill corridor. The remaining treaty layer absorbs part of the loss and exhausts. The remainder, somewhere between $180M and $260M depending on how the loss develops, lands at the carrier's net retention.
The company's surplus going into the season was $310M. The combined hit from the two events plus IBNR development takes surplus below the regulatory threshold by Q1 2027. The carrier looks like Merced did in November 2018. The difference is that the geographic concentration this time was not a function of agent relationships or historical book inertia. It was a function of an algorithm finding a cluster the historical data said was favorably priced and growing the book into it.
The model in this scenario did exactly what it was optimized to do, which was find under-priced clusters in the historical loss data and grow the book into them. The optimization target did not include "the climate the model was trained on is no longer the climate the book is exposed to," because that constraint is hard to write into a loss function and even harder to enforce after the fact when the binds have already happened. By any technical measure of a rating model's performance, the system worked. The portfolio it produced is the part that did not survive contact with the actual peril distribution.
The Elementum Example
The pattern of historical-data wildfire models needing re-fitting against current climate conditions has already shown up in the catastrophe modeling and ILS markets. Elementum Advisors, an insurance-linked securities manager, told Insurance Journal in February 2024 that it had refined its wildfire model after analyzing data from almost two million US wildfires.7 The firm's prior third-party model, in their words, was "benchmarked to historical trends and not to today's climate."7
Elementum is a sophisticated ILS investor with strong actuarial and modeling capabilities. They identified the staleness, did the work to re-fit, and published their conclusion. The implicit message to cedents using off-the-shelf cat models or ML-driven rating tools is that the same staleness problem applies to them, and very few carriers have the internal capability to detect it from their own results until the book has already been built into the wrong shape. By the time the loss ratio shifts the conversation, the geographic concentration is already on the books for the next renewal cycle, and the cat XL treaty above it was placed against an exposure profile that no longer matches the carrier's actual aggregation.
What Actually Lives in the Treaty Binder
A regional carrier that lost a quarter of surplus to a wildfire in a region the rating model said was attractive is going to face two conversations.
The first is with the state regulator, which will ask whether the rate-making process complied with prior-approval requirements and whether the model documentation supports the indicated rates. That conversation is procedural. It produces orders and findings.
The second is with the reinsurer at June 1. Standard catastrophe excess-of-loss treaty wordings include warranties around disclosure of changes in underwriting practices, rating philosophy, and aggregation exposure. We covered the mechanics of how AI model retraining interacts with those warranties in a companion piece on why the reinsurance treaty is where AI risk becomes existential. The short version is that a reinsurer reviewing a cession that took an unexpected loss will read every commit message the cedent's data science team produced, and the disclosure history the broker documented in placement, and form a view about whether the cedent told them what was changing inside the model.
A carrier with reinsurance support behind it may survive a single bad season. A carrier facing a treaty non-renewal at three times the prior cost the following June, after a season that already cost it a quarter of surplus, is in a different conversation entirely, and the conversation rarely ends with the company still writing new business.
The CEO Question
The question to ask the chief actuary, the head of underwriting, and the data science lead, in that order, is short.
How concentrated is our AI-rated book in any single peril zone, and when did the rating model last see today's climate?
The first half of the question requires producing an aggregation report by peril, by territory, with the AI-rated portion of the book broken out separately from the legacy actuarial book. Most carriers can produce a version of this report. Few produce one that distinguishes the two sources of the rate.
The second half requires a model card or its equivalent, dated, with training data window stated explicitly, and a comparison against the most recent NOAA, CAL FIRE, or USGS climate baseline relevant to the peril. A model trained on 2010-2020 wildfire data, when the relevant climate baseline has shifted measurably since, has documented staleness, and the documentation belongs in front of the reinsurer at the next renewal rather than in front of the conservator after the next event.
Merced was an A-rated, 112-year-old California carrier. The Camp Fire ignited on a Thursday morning. Twenty-five days later the company was being liquidated. The mechanism that closed Merced is older than AI, faster than regulation, and waiting in every concentrated, model-rated homeowners book in the state right now. The CEOs who have already asked the question above are buying themselves a second conversation with their reinsurer next June, and a much shorter call with the Department of Insurance the season after that.
Related
- The Reinsurance Treaty Is Where AI Risk Becomes Existential
- The 89-Day Death Clock: What FedNat Teaches Every Carrier Building AI
- The One Treaty Clause Your AI Strategy Lives or Dies On
Footnotes
-
"The company was formed in 1906 by a group of farmers in Merced County who found it difficult to obtain fire insurance because the area had no firefighting abilities, according to the company's website… Still, the company carried a solid A-minus rating from insurance consultant A.M. Best before the fire." — United Policyholders: Insurer goes bust from Camp Fire with millions in claims unpaid ↩
-
"On November 30, 2018, California Insurance Commissioner Dave Jones filed a petition to place Merced Property and Casualty Company ('Merced') into liquidation. On December 3, 2018, the Superior Court of the State of California, County of Merced [entered] an Order Appointing Liquidator, in the case entitled Insurance Commissioner of the State of California v. Merced Property and Casualty Company, Case No. 18CV-04739 ('Liquidation Order')." — California Department of Insurance: Regulator takes control of small failing insurer (Press Release release141-18) ↩
-
"The Camp Fire caused 85 fatalities, displaced more than 50,000 people and destroyed over 18,000 structures… Munich Re estimated the fire caused $12.5 billion in covered losses and $16 billion in total losses." — Wikipedia: Camp Fire (2018) (citing Munich Re's NatCatSERVICE estimate) ↩
-
"A spokesperson for the Department of Insurance says the company had 200 policies in Butte County, but doesn't know how many people impacted by the fire held insurance with the company." — ABC30 Fresno: State takes over Merced insurance company unable to pay out claims after Camp Fire ↩
-
"Aon pegged insured natural catastrophe losses at $118bn for 2023, with perils traditionally considered as primary or peak — tropical cyclone, earthquake and European windstorm — accounting for just 14 percent of the total." — The Insurer: Aon estimates 2023 insured nat cats at $118bn as protection gap rises to 69% ↩
-
"Since the passage of California's last major reform of insurance industry regulations, Proposition 103, insurers have only been allowed to use historical modeling, looking at the past 20 years' worth of climate data when justifying increases to customers' premiums." — California Department of Insurance: Proposition 103 and ABC10: California insurance regulation reforms take step forward ↩
-
"Elementum Advisors LLC, a $3.6 billion investment manager specializing in cat bonds and other ILS products, says it's had to devote considerable time and resources to refining the wildfire model it licensed just a few years ago. It was 'benchmarked to historical trends and not to today's climate,' said Jake Weber, Elementum's head of data and analytics. After analyzing data from almost two million US wildfires, Elementum saw a 'statistically significant, higher frequency of areas that were burned in northern California' than the model indicated." — Insurance Journal: Risk Models Behind Insurance-Linked Securities Are Getting a Lot Harder to Crack (February 26, 2024) ↩
