Flood insurance keeps rebuilding the same homes — 227,000 repetitive-loss properties, 12,972 mitigated since 1989

What is happening?

The National Flood Insurance Program pays claims on the same buildings over and over. As of January 2026 there were 112,640 repetitive-loss properties insured by the NFIP, about 2.5 percent of the 4.58 million policies in force in December 2025 — and as of December 2021, 48 percent of NFIP claims by dollar value had been paid to properties with two or more losses.

The stock is growing, not shrinking. Unmitigated repetitive-loss properties went from about 131,000 in 2009 to 176,000 in 2018 to 227,000 in 2026, a rise of roughly 73 percent. Against that, FEMA mitigated 12,972 repetitive-loss properties across all of its grant programs from 1989 through 2025.

The bill lands on the Treasury. FEMA has borrowed about $38.5 billion since 2005 and still owes $22.525 billion, leaving $7.9 billion of borrowing authority. The current authorization expires 2026-09-30.

Whose problem is this?

Nobody in the Decides row pays the claim, and the party that pays the claim — the federal Treasury — cannot initiate a buyout. Acquisition runs through a state or local government and requires a willing seller, so the payer of last resort has no lever on the one action that would end the payments.

Where does this problem end?

The boundary matters because the program already knows which properties these are and counts them every year. This is not a measurement gap — it is a case where a well-measured concentration persists anyway.

What is the state now, and what should it be?

Now

Needs a new measurementThe target state. None of the sources read here states an official target for the number of repetitive-loss properties, a target mitigation rate, or a date by which the concentration should fall. The only quantified federal trajectory found is actuarial, not physical: GAO estimates that it will be 2037 before 95 percent of policies reach full-risk premiums, with a $2.7 billion premium shortfall in 2023 and $26.7 billion over the period.

How big is it?

The affected count is between 112,640 and 227,000 properties, and the two ends measure different things rather than bracketing one number. The low end is the set of repetitive-loss properties currently insured by the NFIP as of January 2026 — the exposure the program is paying for today. The high end is the set of unmitigated repetitive-loss properties as of 2026, and that count holds whether or not the properties in it still carry policies.

The unit here is structures rather than people, because the program counts insured buildings and not the households living inside them. The sources do not report occupancy, and the NFIP also covers non-residential buildings, so multiplying by an average household size would manufacture precision that no source supports.

Under what conditions does it arise?

Three conditions hold this pattern in place, and each one is a rule rather than an accident of weather.

1. The fast money rebuilds and the slow money moves. A claim pays out on the existing structure. An acquisition requires planning a project, applying for and receiving a grant, then purchasing and demolishing the home — a process that typically takes at least 2 to 3 years and often longer. In a FEMA analysis of 1998–2018 data, states took on average 16 months after a disaster declaration merely to submit Hazard Mitigation Grant Program acquisition applications. For Hurricane Helene, which struck in September 2024, FEMA had approved 82 acquisitions as of February 2026 while more than 575 further applications had been submitted and not approved. 2. Premiums are capped below risk by statute. FEMA is statutorily required to charge rates that do not fully reflect flood risk, and faces limits on how much it can raise premiums each year. So the property that floods repeatedly does not pay the actuarial price of doing so, and the gap is carried by the premium base and the Treasury. 3. Mitigation is voluntary and cost-shared, and its funding is unstable. The nonfederal cost share is typically 25 percent, which some communities cannot raise. Because owners participate voluntarily, acquisitions produce checkerboarding — some homes in a block go, others stay, and the community still services a thinned-out street. Funding itself moved twice in one year: FEMA announced in April 2025 that it was ending the BRIC program, and in December 2025 a court ordered that termination reversed.

What has been tried?

Both directions have been tried — buy the property out, or price the risk correctly — and neither has bent the count. As of February 2026, Congress had not enacted the means-based affordability assistance that GAO recommended in July 2023.

What was found?

These rows do not move together, and that gap is the whole finding. Mitigation works per dollar and is measured to work per dollar, yet the number of properties needing it rose by 96,000 between 2009 and 2026.

Why is it still unsolved?

Incentive inversion — the cheapest, fastest, and most certain response to a flood is to rebuild in the same spot at federal expense, and the response that would end the losses is slow, voluntary, locally cost-shared, and rationed.

The money that arrives quickly pays to put the house back where it stood, and the money that would move the household out of the floodplain arrives slowly or never. A claim is an entitlement under a policy the owner already holds. A buyout is a competitive grant that a local government must choose to pursue, fund a quarter of, and shepherd for years. Facing those two doors after a flood, the rational owner takes the claim — and having taken it, becomes more likely to take the next one.

The inversion is structural rather than a matter of anyone acting badly. FEMA cannot price the risk away because statute forbids charging the full-risk rate immediately and caps annual increases. Local governments hesitate because acquisition permanently removes land from the tax rolls and leaves gaps in the street grid. Owners decline because they may owe more on the mortgage than the buyout yields. Each of these is a defensible position, and their sum is a program that has borrowed $38.5 billion since 2005 to restore buildings to locations already known to flood.

Risk Rating 2.0 was meant to close the gap from the pricing side, and it does add a surcharge after a second loss. But the same reform appears to shrink the insured pool rather than the risk pool: new policies fell 11 to 39 percent and renewals 5 to 13 percent depending on the size of the increase. A property that drops coverage leaves the claims data without leaving the floodplain, which improves the program balance sheet and worsens the underlying exposure.

What observation would mean it is solved?

Candidates — (a) the count of unmitigated repetitive-loss properties falls year over year (b) the share of claim dollars going to properties with two or more losses drops well below 48 percent (c) NFIP debt to the Treasury falls without a write-off.

No single number closes this problem, because each candidate can improve for a reason that has nothing to do with mitigation. (a) falls if owners simply drop coverage or abandon the structure, and the insured count is the more easily gamed version of it. (b) is a ratio, so one catastrophic year paying out on first-time losses lowers it while no repetitive-loss property is touched. (c) needs its qualifier because the debt has fallen before by legislation rather than by repayment — Congress canceled $16 billion in 2017 — and in any case the balance tracks hurricane landfall more than program design. Read together, (a) measured on the unmitigated stock rather than the insured stock, plus a rising mitigation count, is the pair that cannot be faked by attrition.

What is it connected to?

Fills with researchlikely links to coastal development policy, mortgage lending in flood zones, disaster recovery funding, and the private flood insurance market. Relation type and evidence grade were not researched in this round.

What these sources do not say

• The dollar total ever paid to repetitive-loss properties. CRS states plainly that the OpenFEMA multiple-loss dataset carries claim counts per structure but not claim amounts, and that no online information allows that total to be calculated. The headline 48 percent is a share, not a sum.
• How many people are involved. None of these sources reports occupancy, tenure, or the residential-versus-commercial split of the 227,000 unmitigated properties. Owners and renters are not distinguished anywhere, though only an owner can accept a buyout.
• The state-level distribution of the 227,000. GAO maps expected premium change by state, which is a different quantity. Where the concentration physically sits was not obtainable from the sources opened here.
• The national size of the acquisition backlog. The figure of 82 approved against more than 575 submitted covers one storm in one state at one date. No source read here totals pending applications nationally, so the backlog cannot be sized.
• Whether the repetitive-loss surcharge changes behavior. GAO gives a worked example of a surcharge rising from $482 to $964 after successive claims, but no source reports whether surcharged owners mitigate, drop coverage, or file anyway.
• What becomes of a property after it exits the NFIP. Given that risk-based pricing is reducing uptake, this is the pivotal unknown, and none of these sources tracks whether such structures are sold, rebuilt, occupied, or left standing in the floodplain.
• Whether the headline counts hold outside one document. The 227,000 unmitigated and 112,640 insured repetitive-loss properties, the 48 percent claim-dollar share, the 12,972 mitigated repetitive-loss properties and the 95,762 total mitigations all rest on GAO-26-109045 alone. Six further sources were opened to test that dependence — three CRS products, two peer-reviewed studies of FEMA buyouts, and one FEMA final rule — and not one of them restates any of those five figures. They are not contradicted either. They are single-sourced, and nothing read here settles whether that is because the counts are unique to this testimony or because the round never reached the FEMA dataset underneath it.
• How long a buyout actually takes. Four figures from three groups of authors are on the table and this dossier does not choose among them. GAO puts an acquisition at typically at least 2 to 3 years. Researcher A and colleagues measure the average Hazard Mitigation Grant Program buyout at 5.7 years from the start of the disaster to project closeout, median 5.3, range 0.4 to 16.8. Researcher B and colleagues put the median at over 5 years from flooding to close-out. NRDC gives an average of five years. The three later figures sit at roughly double the first, but they are not three independent measurements — the two studies share an author and both read the same FEMA grant records. The windows are also not the same thing: GAO describes the acquisition, the studies measure the whole grant project from disaster to close-out. So the gap may be definitional rather than a disagreement. No source read here reconciles them.
• How many properties have been bought out in total. GAO gives 69,415 acquisitions across all FEMA grant programs FY1989 through 2025. Researcher A and colleagues count 43,633 FEMA-funded voluntary buyouts from 1989 to 2017, and limit that count to projects for which data are publicly available. The windows differ by eight years and the second figure is explicitly incomplete, so the two need not conflict — but no source read here reports the total on a single basis. Both values are kept, the first in block 7 and the second in block 13, and neither is presented as the national figure.
• What share of a buyout the community actually pays. The 25 percent nonfederal match stated in block 6 is the general rule, and Researcher B and colleagues confirm it for most federal buyout funds. Against that, a FEMA rule from 2021 records that Biggert-Waters 2012 raised the federal share for repetitive loss properties from 75 percent to between 75 and 90 percent and the federal share for severe repetitive loss properties to 100 percent. Both values are kept and neither is presented as the rate in force. That rule is a preamble describing an amendment, not the current regulatory text — the fetch that returned it stated the cost-share sections themselves were not in what was read — so whether those shares still stand in 2026, and which share the 12,972 repetitive-loss mitigations were funded at, is not reported anywhere read here.

See the evidence

The GAO testimony was read directly as a PDF, page by page, and every figure attributed to it above was taken from that document rather than from a summary. The NRDC fact sheet was likewise read directly. Rows one and two are the same GAO product — the testimony PDF and its highlights page — so these twelve rows carry eleven distinct documents. The last six were opened in a later round whose only purpose was to test how much of this dossier rests on that one testimony, and the answer is in block 12: the counts that name the problem are still single-sourced, while the financial, statutory and buyout-process claims around them are now confirmed two to four times over.

Four sources agree on the debt: CRS TE10122 and CRS R44593 both give $22.525 billion, CRS IF10988 gives the same figure against a $30.425 billion cap with $7.9 billion of authority left, and PGPF gives $22.5 billion as of February 2025 — consistent, because IF10988 records the most recent borrowing as $2 billion in that same month. Three sources agree on the 2017 cancellation of $16 billion, and two on the 2026-09-30 authorization date. Two spreads were left unreconciled because both are differences of base rather than of fact: policies in force are 4.58 million at GAO, over 4.57 million at CRS TE10122 and over 4.5 million at CRS R44593, all as of December 2025, which does not move the 2.5 percent share; and annual receipts are $4.3 billion of premium at PGPF against about $4.6 billion of premiums, fees and surcharges together at CRS R44593. Several sources could not be opened: fema.gov returned HTTP 403 for the OpenFEMA multiple-loss dataset page and the February 2025 borrowing press release, cbo.gov and pew.org returned 403, and federalregister.gov document pages redirected to a block page, which is why the FEMA rule was read through GovInfo instead. FEMA therefore speaks here in one voice of its own, the 2021 rule, and otherwise through GAO and CRS restating it with attribution. Property counts circulating in search results but not verified in an opened document were excluded, including a widely repeated figure of 12,953 mitigated properties that conflicts with the 12,972 stated by GAO. This is a Path A output, so observation_refs is empty and provenance_mode: press-derived.

People affected

Estimated range 112,640–227,000 As of 2026-01 ~ 2026

Sensitivity This is not a confidence interval and the two ends are not two estimates of one quantity. They are two differently filtered counts from the same source that overlap without either containing the other, so no arithmetic between them is meaningful and none is performed here. The unit is structures, not people. A person count was deliberately not derived: the sources report no occupancy, no owner-versus-renter split, and no residential-versus-commercial split, and the NFIP insures non-residential buildings, so multiplying by an average household size would invent precision no source supports. Both ends are also likely understatements of the physical problem, because a property only enters these counts by having held NFIP coverage at the time of its losses; structures with identical flood histories that were never insured are absent from FEMA statistics entirely. Direction of drift is known and adverse: the unmitigated count rose about 73 percent from 2009 to 2026 while 12,972 repetitive-loss properties were mitigated across all grant programs from 1989 through 2025.

Regional breakdown No opened source gives state-level counts of repetitive-loss properties alongside the national total. The state-level map published in GAO-26-109045 shows median percent change from December 2022 premiums to full-risk premiums, which is a pricing quantity and not a property count, so it cannot support a geographic decomposition. Splitting the national figure by state population is prohibited proportional allocation and would be especially baseless here, since repetitive-loss properties concentrate in coastal and riverine floodplains rather than following population