LexisNexis Scores Home Risk Across Six Perils: The Validation Problem Actuaries Should Ask First

LexisNexis Risk Solutions launched Location Intelligence for Home on May 14, 2026, an AI model that scores homeowners properties across six perils and reports a 20-fold gap in claim frequency between its top and bottom scoring properties.

The interesting number is not the 20x. It is that non-weather water made up 24% of 2025 home claims against 4% for weather-related water, and it is the peril no exterior inspection or aerial image can see.

Key Takeaways

• A 20-fold blended claim frequency gap across six perils is a single statistic covering hail, wind, weather water, non-weather water, freeze, and collapse or falling object.
• 24% of 2025 home claims were non-weather water, against 4% weather-related, on LexisNexis's own book. That is the segmentation gap the product is aimed at.
• The six perils split into two families. Hail and wind sit next to catastrophe modeling and existing imagery tools; freeze, collapse and non-weather water are attritional and have no exterior signature.
• Referral and rate classification are different bars. A score can be good enough to trigger an inspection and still fail the actuarial soundness test for a filed rating variable.
• Policy years 2018 through 2025 are covered by the NAIC's nationwide homeowners data call, which will publish peril-level, ZIP-level results in early 2027.

A Score Aimed at the Peril Underwriting Tools See Least Well

The model sits inside Smart Selection, the existing automated service that issues inspection flags and configurable business rules at new business and renewal. It uses neural network techniques trained on industry-wide home claims data, location signals and historical loss patterns to produce one property-level score across hail, wind, weather-related water, non-weather-related water, freeze, and collapse or falling object.

"Rising loss costs and shifting risk patterns are making it harder for home insurers to rely on traditional underwriting approaches alone," said George Hosfield, vice president of home insurance at LexisNexis Risk Solutions. The company says it will file the model as a predictive model in multiple states for underwriting and rating use.

The bundling is the design decision that matters. Hail and wind live next to catastrophe modeling, where roof age, exterior condition and geocoded storm exposure already drive most carriers' tools, including the aerial imagery platforms from ZestyAI, EagleView and Moody's Cape Analytics.

Freeze, collapse and non-weather water are a different family. They are attritional, plumbing and structure driven, and they leave nothing for an inspector or a satellite to observe. LexisNexis's own split makes the point: 24% of 2025 claims were non-weather water against 4% weather-related. Roughly a quarter of claim volume sits in a peril the industry's existing tools were not built to see.

A Blended Lift Across Six Perils Is Not a Rating Variable

Frequency, severity and loss-cost volatility differ enormously across those six perils, so one combined lift figure cannot support a pricing decision. A model that separates hail exposure well and non-weather water poorly still produces an impressive blended number if hail dominates the score's variance, while doing nothing about the plumbing problem the product is positioned to solve. The reverse holds too.

What an actuary can test is narrower and more useful: six lift curves, one per peril, each on a holdout sample the vendor did not fit on. That is also where the regulatory expectation sits, since state model governance rules built on the NAIC's AI taxonomy push toward peril-specific and use-specific validation rather than a headline statistic.

For non-weather water the test has a specific form, because that is where the model's theoretical edge is real. A model trained on industry claims can in principle learn neighborhood plumbing vintage from subdivision-era construction codes, water pressure and soil interactions with pipe corrosion, or appliance failure clustering by manufacturer cohort and region. Traditional rating reaches all of that only through home age, prior loss count and visible inspection findings.

So the question is whether the non-weather water lift survives once the carrier's own plan already controls for home age, a plumbing material proxy where available, and prior loss count. If it collapses, the score is re-deriving existing variables at a higher price. If it holds, there is a real new variable, and the next decision is where it goes.

That placement decision carries different burdens. A referral rule needs directional validity and consistent application across territories and underwriters. A filed rating variable has to be predictive, not unfairly discriminatory, and stable enough that policyholders are not repriced on noise. A score can clear the first and fail the second on cohort instability alone.

The Data Call Will Make the Territory Pattern Public

The timing runs into a separate NAIC initiative. State regulators issued a nationwide homeowners market data call in spring 2026 covering policy years 2018 through 2025, requiring insurers writing at least $50,000 of relevant premium to report peril-level claims, premiums, deductibles, cancellations and non-renewals down to ZIP code.

The June 15, 2026 deadline drew a request for a one-month extension to July 15, and the NAIC plans a public report on the aggregated findings in early 2027.

It was not built with any vendor's model in mind, and that is what makes it consequential here. It will produce the first peril-level, ZIP-level public dataset against which anyone can check whether AI underwriting tools are delivering the segmentation gains claimed, or reallocating coverage away from ZIP codes correlated with income or demographic characteristics.

A carrier deploying a third-party score should expect its own territory-level actions to become comparable against that dataset once the 2027 report lands, whether or not it intended the comparison. Non-renewal patterns that cluster geographically will be visible within a reporting cycle or two, and the peril-level validation file is what separates explaining that pattern from discovering it alongside a market conduct examiner.