FloodVuln Global is a vendor-neutral vulnerability layer that converts flood intensity into a probability distribution of physical damage — specific to building archetype, geography and the local evidence you can actually produce.
A flood model can now resolve inundation to a few metres. The step that turns that water into money is still, in most portfolios, a single curve mapping one depth to one expected damage ratio — and the observations say that curve is not what happens.
Observed NFIP claims for single-family, one-storey, slab-on-grade dwellings. Bars are the observed P10–P90 within each depth bin; dots are observed bin means. The published JRC curve and the fitted curve are drawn over the same data.
A curve fitted on one country’s building stock, moved to another, carries its construction, code era and repair economics with it — invisibly.
Calibrating on settled claims without normalising deductibles and limits teaches the model your policy terms and calls the result vulnerability.
Out-of-domain depth, occupancy or geography returns a number with the same confidence as an in-sample one, and nothing on the screen says so.
Five independently versioned layers. Every production prediction reproduces from a model version, an evidence snapshot, a calibration dataset ID and a parameter set — so model governance is a property of the system, not a document written afterwards.
A universal translation layer between flood intensity and financial damage — designed to plug into the hazard and catastrophe-model ecosystem you already run, and strengthen the vulnerability component inside it.
We do not rebuild stochastic event sets, terrain models or policy financial modules. Bring depths from Fathom, JBA, Moody’s RMS, Verisk, a government map or your own hydrodynamics — the vulnerability layer is the product.
Choose an archetype, a depth and a region. The engine returns a bounded damage distribution fitted to 1,861,852 US flood claims, and separates the width that is the world from the width that is the thinness of your evidence. Only the second one responds to calibration.
Region drives the local sample size, the evidence grade and the confidence score. The value on the right is yours, not ours: the engine returns a damage ratio and the money is your number.
Calibration Level — Local EvidenceEvery response carries its archetype, calibration level, evidence grade and model version. Pin the version and the result reproduces next year, after the model has moved on.
Live engine. Coefficients are the maximum-likelihood fit to 1,861,852 OpenFEMA NFIP claims, with hierarchical offsets per archetype and region and Laplace posteriors behind the confidence score. Model version 1.0.0.
Local data update the global prior in proportion to their quality and sample size, and the fitted model shows it: Florida (n=63,780) carries a large offset, Utah (n=4) is shrunk to within 0.02 of the archetype. What that buys is a correctly located answer and a far smaller epistemic term — not a narrower range of outcomes, and not, on this evidence, a better forecast for an event unlike the ones already seen.
Measured on a withheld region, eight resamples per point. The unpooled refit is competitive from about ten local records; tracing that to an offset carrying the level of the curve but not its slope led to a shape offset that restores the pooling advantage below about fifty records. Both are in the evidence section.
| Level | Minimum Evidence | Model Behaviour |
|---|---|---|
| G0 | Location, sector, depth | Global / continental prior |
| G1 | Building characteristics | Archetype prior |
| G2 | Structured local expert evidence | Informative prior modifiers |
| G3 | Tens of depth–damage pairs | Hierarchical posterior, shrunk |
| G4 | Hundreds of quality records | Market-specific posterior |
| G5 | Carrier portfolio history | Client-specific layer |
We elicit distributions, not curve edits — the proportion of masonry in the local stock, typical first-floor height, basement prevalence, code era, repair practice. Expert statements are stored as auditable priors with source, date and confidence, and they move the answer by exactly as much as they should.
The same versioned engine answers the underwriter, the pricing actuary, the accumulation team, the model-validation function and the claims unit — which is the point. Today those five desks usually disagree because they are quietly running different vulnerability assumptions.
Both properties sit in the same 1-in-100 depth band, and a depth-only curve prices them identically. The archetype split is where the underwriting decision actually lives: the same 0.90 m of water does about twice the proportional damage to one as to the other, and the fitted model says so from 1.86 million claims rather than from judgement.
Not a score. A severity distribution, the archetype it was drawn from, the evidence grade behind it, and a warning if the depth sits outside the calibrated domain.
| Risk | Archetype | Mean DR | P90 | P(>50%) | Grade |
|---|
Depth 0.90 m above the lowest floor; fitted engine output.
Expected loss comes off the mean; the risk load comes off the tail. A point estimate gives you the first and forces you to guess the second, usually through a flat regional factor that is wrong in both directions at once.
Because ground-up damage is kept strictly separate from policy terms, the deductible and limit apply downstream of the vulnerability layer — so a change in attachment does not silently rewrite the physics.
Each country arrives with its own national classification and its own local functions. The taxonomy crosswalk makes the books comparable; the evidence grade stops that comparability from being mistaken for equal confidence.
| Territory | Dominant Archetype | Exposure | Mean DR | AAL (bps) | 1-in-200 | Evidence |
|---|
Illustrative portfolio. AAL expressed in basis points of total insured value; 1-in-200 is the modelled portfolio ground-up loss at that return period.
This is where most first engagements start, because independence is an asset to a validation team rather than a threat. FloodVuln Compare benchmarks your incumbent vendor curve, your own internal curve and the harmonized global families against the same observations, on the same hold-out.
You get the disagreement, the size of it, and the evidence behind each side — in a form a regulator will accept as model-risk documentation.
Mean signed error in damage ratio, averaged over eleven named events each withheld from training in turn. Negative is under-prediction. Measured, not illustrative — the underlying table is in the evidence section.
Post-event observed depth plus the exposure file gives a probable severity for every affected location within hours. Not to settle claims — to sequence them, staff them and reserve for them while the water is still in the street.
Because the output is a distribution, the reserving question “what is the chance this book runs past our booked figure” is answerable on day two, not month three.
Banks are asked to quantify physical climate risk on collateral they did not underwrite for flood, using hazard layers that stop at the water’s edge. The missing step is the same one insurers are missing: what a given depth does to a given building, with the uncertainty stated. Damage on collateral is what moves valuation, LTV, loss given default and, ultimately, expected credit loss.
Hold vulnerability fixed and move the hazard, or evolve adaptation and hold the hazard — the two are separable, so a supervisor can see which assumption is doing the work.
Damage distributions feed a forward-looking haircut on collateral value, and therefore LGD, with provenance a model-validation function can trace back to source evidence.
Quantified avoided loss from elevation, floodproofing or utility relocation turns a resilience retrofit into a priceable adjustment rather than a marketing claim.
When the lender and the insurer of the same building disagree about its flood vulnerability, one of them is mispricing and neither can prove which. A shared, versioned, independently validated vulnerability layer makes that disagreement visible and testable — which is a precondition for insurance-linked lending terms, resilience-linked mortgages and any credible transfer of physical risk between the two balance sheets.
Output from a reproducible run against public data: 1,927,948 NFIP claim records and the published JRC global depth–damage database, with the full benchmark programme run out of sample.
The JRC North America residential curve is the standard fallback where no local function exists. Against 1.86 million US claims it runs high by roughly 0.2 to 0.3 in absolute damage ratio between 0.5 and 3 m — the size of the correction a calibration layer exists to supply.
A portfolio priced on an uncorrected global curve carries a systematic, measurable bias — here toward over-reserving. Direction and size are both testable.
| Vulnerability Assumption | MAE | RMSE | Bias | CRPS | 80% Cov | Cal. Slope |
|---|---|---|---|---|---|---|
| FloodVuln archetype prior | 0.2169 | 0.2695 | -0.0460 | 0.1489 | 0.786 | 1.065 |
| FloodVuln region calibrated | 0.2193 | 0.2775 | -0.0599 | 0.1540 | 0.753 | 1.052 |
| FloodVuln global fit (no archetype) | 0.2324 | 0.2827 | -0.0487 | 0.1583 | 0.789 | 0.826 |
| Single depth-damage curve fitted on training data | 0.2305 | 0.2859 | -0.0727 | 0.2305 | n/a | 0.829 |
| Constant mean (naive) | 0.2469 | 0.2986 | -0.0712 | 0.2469 | n/a | n/a |
| JRC North America residential (published) | 0.2493 | 0.3031 | 0.0910 | 0.2493 | n/a | 0.360 |
All models scored identically; a deterministic curve is a degenerate forecast, so its CRPS is its mean absolute error. Lower is better for MAE, RMSE and CRPS; 80% coverage should be 0.80 and calibration slope 1.0.
| Vulnerability Assumption | MAE | RMSE | Bias | CRPS | 80% Cov | Cal. Slope |
|---|---|---|---|---|---|---|
| FloodVuln archetype prior | 0.1912 | 0.2439 | 0.0164 | 0.1237 | 0.836 | 0.881 |
| FloodVuln global fit (no archetype) | 0.2111 | 0.2584 | 0.0353 | 0.1332 | 0.840 | 0.759 |
| Single depth-damage curve fitted on training data | 0.2028 | 0.2560 | 0.0108 | 0.2028 | n/a | 0.764 |
| Constant mean (naive) | 0.2141 | 0.2618 | 0.0134 | 0.2141 | n/a | n/a |
| JRC North America residential (published) | 0.2695 | 0.3239 | 0.1657 | 0.2695 | n/a | 0.339 |
This is the zero-local-data test: the held-out state contributes nothing to training, so it measures what a new market gets on day one.
| Local Records | Pooled MAE | Pooled CRPS | Local-Only MAE | Local-Only CRPS |
|---|---|---|---|---|
| 0 | 0.2745 | 0.1888 | — | — |
| 5 | 0.2714 | 0.1811 | 0.2598 | 0.2185 |
| 10 | 0.2718 | 0.1831 | 0.2483 | 0.1792 |
| 25 | 0.2697 | 0.1765 | 0.2324 | 0.1551 |
| 50 | 0.2679 | 0.1738 | 0.2303 | 0.1526 |
| 100 | 0.2666 | 0.1681 | 0.2272 | 0.1485 |
| 250 | 0.2666 | 0.1648 | 0.2244 | 0.1462 |
| 500 | 0.2674 | 0.1641 | 0.2274 | 0.1456 |
| 1000 | 0.2674 | 0.1640 | 0.2276 | 0.1456 |
| 2500 | 0.2673 | 0.1636 | 0.2281 | 0.1454 |
The pooled model is never worse than the prior it starts from. An unpooled local refit on the same records is worse until it has enough data to earn its variance, which is the entire argument for partial pooling.
| Source | Title | Licence | Status in This Build |
|---|---|---|---|
| JRC_GLOBAL_2017 | Global flood depth-damage functions: methodology and the database with guideli | Open | Ingested |
| NFIP_CLAIMS_V3 | NFIP Redacted Claims v3 (OpenFEMA) | Open | Ingested |
| HAZUS_FLOOD | Hazus Flood Model Technical Manual (depth-damage function library) | Open | Not Obtained |
| USACE_NACCS | North Atlantic Coast Comprehensive Study physical depth-damage functions | Open | Not Obtained |
| HOWAS21 | HOWAS21 object-specific flood damage database | Reference Only | Not Obtained |
| MCM_UK | Multi-Coloured Manual property and economic loss functions | Licensed | Not Obtained |
| INSYDE | INSYDE: a synthetic, probabilistic flood damage model based on explicit cost a | Open | Benchmark Only |
Two public documents refused automated retrieval; two more require a licence or data agreement. None were scraped around.
A vulnerability layer that cannot be independently tested is worth nothing to a model-risk function, whatever it scores. The benchmark programme is pre-specified, reproducible and open to review — and the marketing language is held to what it has actually passed.
| Benchmark | Design | Primary Metrics |
|---|---|---|
| Curve reproduction | Harmonized function vs source implementation | transform error, datum QA |
| Hold-out event | Exclude an event, predict its losses | MAE, RMSE, bias, CRPS |
| Leave-region-out | Train elsewhere, predict a held-out region | transfer error, calibration |
| Leave-country-out | Zero-local-data global coverage test | vs JRC / transferred baseline |
| Learning curve | n = 0, 5, 10, 25, 50, 100, 500 local records | error reduction per record |
| Portfolio blind test | Insurer withholds event and supplies allowed inputs only | portfolio error, severity calibration |
| Uncertainty | Evaluate the probability forecasts themselves | PIT, CRPS, coverage, slope |
The API returns a scientifically documented answer anywhere on earth. What changes by geography is the evidence grade attached to it — and we would rather show you a D than dress it up as an A.
“Globally harmonized, locally calibratable, probabilistic flood vulnerability, built for transparent benchmarking.”
“The most accurate flood model in the world.” “Uncertainty eliminated.” “Guaranteed accurate at every property.”
A first engagement replaces nothing. You supply a historical sample — property attributes, event depth or an agreed hazard layer, ground-up or normalised loss, and the policy metadata needed to strip out financial terms. We run blind against your current vulnerability assumption.
Does FloodVuln reduce blind portfolio loss error and improve probabilistic calibration relative to your current vulnerability assumption — while staying explainable enough for model governance?
| Workstream | Deliverable |
|---|---|
| Data audit | Coverage, missingness, depth datum, loss normalisation, leakage checks |
| Baseline reproduction | Your incumbent curve’s performance on the agreed hold-out |
| Global prior | FloodVuln with no client-specific calibration at all |
| Sparse calibration | Learning curve at n = 10, 25, 50, 100, 500 |
| Full calibration | Client-specific posterior on the approved training subset |
| Blind evaluation | Location-level and portfolio-level predictive metrics |
| Governance pack | Methodology, evidence, limitations, version IDs, reproducibility bundle |