Manaz Afzal
Insurance · Claims triage

Finding the large flood losses on the day they are lodged

A first-notice triage score for a property insurer’s claims team, fitted on 2019 to 2023 flood claims and tested on 96,255 claims from the 2024 and 2025 hurricane seasons.

0.87AUC at first notice, before inspection
92%of large losses found at the rule’s caseload; rule 76%
81%of the top-tenth tranche is large
72%of paid dollars sit in large losses

Executive summary

Scoring new flood claims at first notice finds 92% of large losses at the caseload where a coverage-and-zone rule finds 76%.

0.87AUC at first notice, before inspection
92%of large losses reach the large-loss desk, against 76% for the rule
72%of paid dollars sit in losses of US$100,000 or more

Situation

After a major flood, a property insurer’s claims team receives a surge of new claims; the minority that become large losses carry most of the money and wait in the same queue.

Approach

A first-notice score from policy and lodgement fields, fitted on 2019 to 2023 US flood claims and tested on 96,255 claims from 2024 and 2025, against a coverage-and-zone desk rule.

Result

AUC 0.868 without inspection, 0.883 once water depth is recorded. At the rule’s caseload the score finds 92% of large losses against 76%.

The problem

After a flood, an insurer’s claims team faces thousands of property claims at once, and the largest losses are mixed in with minor ones. In the 2022 east-coast floods, insurers used flood-depth imagery and policy data to send assessors to the worst-hit homes first. Most claims, though, are still triaged on what the customer says at lodgement.

I measured how well a score from the policy and the first notice of loss finds the large losses on the day they are lodged, against sorting by the amount the customer estimates.

Team
Claims operations, large-loss and catastrophe desk, reserving actuaries, panel adjusters
Scale
Home and small commercial property books with tens of thousands of claims in a catastrophe year
Systems
Guidewire ClaimCenter, Duck Creek Claims or Sapiens, with policy data from the policy admin system
Measure
Large losses found in the first review tranche; caseload on the large-loss desk

What I built

A first-notice score for each new property claim, computed from the policy and the notice itself: flood zone, building cover and replacement cost, contents cover, occupancy, construction year, foundation and basement type, floors, elevation, cause of damage, whether the loss belongs to a named event, location and the days between loss and notice. A second version adds water depth inside the building once the customer or a desk adjuster records it.

Claims are ranked on arrival. The top band goes to the large-loss desk with an early reserve prompt; the middle band to standard handling; the low band to fast-track settlement with photo evidence.

First noticePolicy and noticefields at lodgementScoreLarge-loss likelihoodwith reason codesRouteLarge-loss desk, standardor fast-track queueWater depthScore refreshed whendepth is recorded
Reason codes show which fields drove the score, so the desk lead can overrule it.

Results

The score was fitted on claims from losses in 2019 to 2023 and tested on 96,255 paid claims from 2024 and 2025, a period that includes Hurricanes Helene and Milton and was never used to fit anything. It ranks a random large loss above a random smaller claim 87% of the time (AUC 0.868); adding water depth lifts that to 0.883.

0%25%50%75%100%0%20%40%60%80%100%FNOL fields + water depthFNOL fields onlyRandom orderTop fifth of claims:44% of large lossesShare of new claims sent to the large-loss desk, highest score first
Reviewing the top tenth of new claims finds 22% of large losses; 81% of that tranche is large. Large losses were 36% of test claims, so a random tenth would find a tenth of them. Data: OpenFEMA NFIP redacted claims

Against the coverage-and-zone rule

A rule sending every claim with building cover of US$250,000 or more in an A or V zone to the large-loss desk flags 56% of claims and finds 76% of large losses. The score, sending the same number of claims, finds 92%. The desk keeps the same caseload and misses about a third as many large losses.

Large-loss desk ruleClaims sent to the deskLarge losses foundDesk caseload that is large
Building cover US$250k+ in an A or V flood zone56%76%49%
Score, same caseload56%92%59%
Score, top fifth only20%44%80%
Score, top tenth only10%22%81%
Highest score bandmedian paid US$201k81%Score band 9median paid US$162k79%Score band 8median paid US$127k65%Score band 7median paid US$103k52%Score band 6median paid US$81k38%Score band 5median paid US$59k25%Score band 4median paid US$33k13%Score band 3median paid US$15k6%Score band 2median paid US$12k2%Lowest score bandmedian paid US$18k0%
Scores sort claims cleanly: the highest band is mostly large losses, the lowest almost none.

What the signals show

Water depth is the single strongest fact once it is known. Claims with 25 to 48 inches inside the building become large losses more than half the time; claims recorded at 49 inches and above show a lower rate, a pattern I would check against the insurer’s own claims before relying on it.

0 inches recordedmedian paid US$30k20%1-6 inchesmedian paid US$69k36%7-12 inchesmedian paid US$83k44%13-24 inchesmedian paid US$106k53%25-48 inchesmedian paid US$110k55%49+ inchesmedian paid US$37k24%
Depth is often unknown at lodgement, which is why the score works without it and is refreshed when it arrives.

Speed of notice matters too. Claims lodged within three days of the loss became large losses 32% of the time; those lodged more than 30 days later, 9%. Severe damage gets reported fast.

Reported on the day of loss28%1 to 3 days later32%4 to 7 days28%8 to 14 days19%15 to 30 days14%More than 30 days9%
Rated flood zonePaid claimsLarge-loss rateMedian paid
A (high risk)191,03230%US$46k
X / B / C (moderate-low)49,24216%US$26k
V (coastal high)8,45728%US$29k

Holding up in a surge

2024 was the heaviest year in the data, with US$8.8 billion paid. Inside each event the score still ranks claims well, which is what a catastrophe desk needs when every claim in the queue shares the same storm.

Large-loss rate, all test claimsAUC inside the eventHurricane Helene87%Hurricane Milton78%2025-04-Central U.S.-Flooding81%Hurricane Debby84%
Hurricane Milton’s claims are harder to rank: only 26% became large losses, against 50% for Helene.
201915% of claims largeUS$1.5bn20209% of claims largeUS$0.9bn202117% of claims largeUS$2.0bn202235% of claims largeUS$5.3bn202318% of claims largeUS$1.0bn202438% of claims largeUS$8.8bn202518% of claims largeUS$0.6bn
The large-loss share swings with the event mix, so the review threshold is set as a share of the queue, never as a fixed score.

Data and method

Data
OpenFEMA FIMA NFIP Redacted Claims (v3), US National Flood Insurance Program, public domain: 249,139 claims with a payment, losses in 2019 to 2025
Target
Building plus contents paid of US$100,000 or more
Split
Fitted on 152,884 claims from 2019 to 2023 losses; tested on 96,255 claims from 2024 and 2025 losses
Features
Policy and first-notice fields only; water depth in a second version; no payment, damage or adjuster fields
Method
Gradient boosting with native categorical handling; benchmark against a coverage-and-zone rule
Measures
AUC, average precision, large losses found at fixed review shares, results inside each named event

Limitations

How I would take it live

Financial impact

About 586 more large losses on the right desk in week one, per 10,000 event claims

The gain is large losses that reach a senior adjuster and an early reserve in the first week instead of waiting in the general queue. Event size and the cost of late handling are assumptions.

InputValueSource
Claims from one catastrophe event10,000 mid-size property book
Large-loss share36%Measured 2024 to 2025 test claims
Large losses found by the rule76%Measured
Found by the score, same caseload92%Measured
Cost of late handling per large lossA$3,000 accommodation, leakage, reserve strain

Calculation

  1. Large losses in the event3,623
  2. Extra found in week one586
  3. Late-handling cost avoidedA$1,758,454

Saving per eventA$1,758,454

Faster settlement for low-score claims through fast-track handling is a further gain and is not counted here.

Implementation and rollout

Six weeks from claims extract to live routing

  1. Weeks 1 to 2

    Extract and refit

    Three years of property claims with first-notice fields and final paid amounts from ClaimCenter or Duck Creek; score refitted on the insurer’s own book.

    Claims analytics, Manaz

  2. Week 3

    Set the tranche

    Large-loss desk share and fast-track threshold agreed with the claims lead and reserving.

    Claims lead, reserving actuary

  3. Weeks 4 to 5

    Shadow run

    Score runs on every new claim without routing; misses reviewed with the desk.

    Large-loss desk, Manaz

  4. Week 6 onward

    Live routing

    Score and reason codes on the claim screen; top tranche routed with an early reserve prompt; recalibrated after each event.

    Claims operations, Manaz on support

Tools

Guidewire ClaimCenter, Duck Creek Claims or Sapiens, Policy admin data and flood-zone or flood-height layers, Python, scikit-learn gradient boosting, Power BI or Tableau desk view

Team

  • Claims operations lead owns routing and desk capacity
  • Large-loss desk reviews scored claims and overrules with reasons
  • Reserving actuary early reserve rules for high scores
  • Manaz score, routing integration and recalibration

The timeline is a plan for a typical engagement; it depends on data access and how many systems are involved.

Risks and mitigations

What could go wrong in production, and the control for each

RiskMitigation
A storm unlike past eventsThreshold set as a share of the queue; recalibration after each event; desk overrides logged.
Water depth unknown at lodgementScore works without it and refreshes when depth is recorded.
Bias against older or lower-value homesScore used for routing priority only, never for coverage or settlement decisions; outcomes reviewed by segment.
US data differs from an Australian bookRefitted on the insurer’s own claims before any live use.

Key takeaways

Key takeaways

  • First-notice facts alone rank flood claims with AUC 0.87.
  • At the same desk caseload the score finds 92% of large losses; the rule finds 76%.
  • Water depth is the strongest single fact and refreshes the score when it arrives.
  • Results hold inside each named event, which is what a catastrophe desk needs.