Measure/Property data for underwriting
For insurers and reinsurers
Property data for water damage underwriting
Surface water arriving at a foundation is a building-scale phenomenon, and the terrain that produces it does not show up in a flood zone, which is drawn for riverine and coastal inundation. We deliver, by API and at book scale, per-structure ground slope, upstream contributing area, modeled surface flow and standing water — as measurements, with their uncertainty, and with the quantities we declined carried in the same payload.
The important sentence first, because it decides whether we are worth a call. This is not a risk score, and it has not been validated against loss experience. It is a feature set. The score is yours to build, against your own claims data, and we think that is the only honest arrangement.
What is in the payload
| Field | Definition |
|---|---|
| Per-wall grade | Slope on each wall's own outward normal, measured 0.75–3.25 m out and combined so no single reading decides it. Sign convention stated. Target: 5% fall over 3 m. |
| Grade range and direction call | The spread produced by seventeen footprint positions, and a direction that is allowed to read too close to call. |
| Contributing area | m² draining to each wall, sampled at nine points beside it and re-sampled at every displaced position, published with its range and distribution — never as a bare maximum. |
| Standing water | Depth and extent within 20 m of the footprint, outside every building. Under 3 cm is measurement noise and is not counted. |
| Identity provenance | Which grade of identity evidence answered — and whether identity was established at all. A proximity answer locates rather than identifies and is labeled as such. |
| Declined quantities | Undecided walls, unreadable profiles, and addresses whose ground data failed the quality gate — named rather than omitted. |
The uncertainty is a field, not a footnote
A published building outline is a trace of a roof, and it can sit a meter from the wall it claims to be. On a real Pittsburgh footprint, shifting it north in 0.2 m steps moved one wall from +5.4% to −23.5%, crossing zero inside 1.4 m — less than the outline error the source data's own limits concede.
So each footprint is displaced one meter in each of sixteen directions and every wall is re-measured. Seventeen positions in all, and a direction is emitted only when all seventeen agree on the sign. Magnitude agreement is not required, because a meter of movement on steep ground can legitimately swing a grade from −52% to −13% without saying anything about which way water goes.
For a modeling team this is the difference between a feature you can fit on and a feature that silently encodes footprint vintage.
Address-to-building matching, with the failure modes
At book scale nobody inspects an aerial to catch a wrong structure, so how the match was made is a first-order property of the dataset rather than an implementation detail. It ships as a field, on every row.
| Method | What the claim is | How the row is marked |
|---|---|---|
| Direct identification | The address is tied to this outline in particular | Identified |
| Address point contained by the outline | Independent evidence puts the address inside this outline and no other | Identified |
| Nearest outline — the guess, flagged as one | The nearest outline is all there is | Located, not identified |
| Nothing fired | No evidence identifies a structure | Refused, and never billed |
So a modeling team can stratify on identification strength, or drop the located rows entirely, without taking our word for anything. We publish no headline match rate: one number across a book would hide exactly the variation you would want to condition on.
The caveats belong in the same paragraph. Where the underlying records are thin the evidence is thin too, and the thin cases are not randomly distributed — they concentrate in rural addresses, new construction and recently renumbered streets, which is to say the silence concentrates exactly where the resolver is guessing. Nothing is imputed to cover it.
And not every disagreement is a different building at all: some are our source and somebody else's tracing one roof differently. The two failures look the same from outside and only one of them matters.
What we have not shown, stated in full
No loss validation. These features have never been tested against insurance loss experience. We have no lift curve, no Gini, no AUC to offer you, and we will not construct one from a proxy.
The one predictive test we ran came back null. Measured terrain against Chicago 311 basement-water complaints did not separate them. Read carefully that mostly indicts the proxy — an owner who knows their grading is bad calls a contractor rather than the city, and when a third of a block reports water in one hour the cause is the pipe under the street. It is no evidence that grade does not matter. It is also not evidence that it does, and it does not generalise beyond one flat city.
No claim about any structure. Ground falling toward a foundation raises the risk of water reaching it, which is ordinary building science. That a particular insured structure has water, or will, is not measured and is not claimed — and no output of ours should appear in an adverse action or a non-renewal as though it were.
Not measured at all: soil and infiltration, water table, subsurface drainage, plumbing and appliance condition, roof discharge, sewer capacity, rainfall and return period.
Why the honesty is the product
A feature set that quietly drops the rows it found hard will make your model look better in development and worse in production, because the dropped rows are not a random sample of your book. Ours names them. Walls that did not survive the uncertainty pass, addresses that failed the data quality gate — the median step between adjacent elevations must clear 0.004 m — and identities that could not be established all arrive as data.
A refusal is never billed. If the building is wrong, the property is replaced free. Both are policies we can hold because the engine measures its own uncertainty and only releases what clears it.
Commercial shape
Enterprise underwriting runs on a custom API plan priced by volume: whole-book scoring rather than one address at a time, dedicated capacity that never queues, and re-runs when new survey data lands. That is a different product from the self-serve plans on the pricing page — Solo at $99 a month for 25 lookups, Pro at $399 for 150 — which exist for people looking up individual properties.
Talk to us about a book
Bring a sample of addresses, including the awkward ones. We will run them, show you the refusals as well as the answers, and be specific about where coverage does and does not support the question.
Or run a single address free to see the shape of the output first.
Is this a water damage risk score?
No. It is a measured feature set. The score is yours to build against your own loss data.
Have the features been validated against loss experience?
No, and we say so. The one predictive test we ran used a 311 proxy and came back null — a result that mostly indicts the proxy and still leaves us with no lift figure to offer.
How are refusals handled in a response?
As explicit fields, not as dropped rows. A refusal is never billed.
How is the address-to-building match reported?
As a field on every row. Identity is settled by graded evidence, strongest first, the row carries the grade that answered, and where none is strong enough the address is refused rather than matched to the nearest roof. That lets you stratify on identification strength, or drop the merely located rows, without taking our word for anything.