When Earthquake Loss Estimates Move for Weeks, Each Revision Measures Something Different

A magnitude 7.4 earthquake struck 5 km south of San José del Palmar, in Colombia's Chocó department, at 12:34:28 UTC on 10 August 2026. The U.S. Geological Survey placed the hypocentre at 110.3 km depth, recorded a maximum shaking intensity of MMI 8.0, and issued a red PAGER alert — its highest impact category. Within four days the reported toll moved from about 132 dead to 281 dead, 3,971 injured and 379 missing, with more than 12,000 homes destroyed and 74,000 damaged.

Numbers that move that fast invite a tempting reading: early estimates are wrong, later ones are right, and the honest figure is whichever arrives last. That reading is close enough to be useful and wrong enough to be expensive. The figures quoted after a large earthquake are not successive approximations of one quantity. They are distinct quantities, produced by different institutions, on different clocks, for different purposes. Much of what looks like revision is a reader swapping one measurement for another without noticing.

The Situation as of Mid-August 2026

Three things are true at once about the Colombian event, and they point in different directions.

The shaking was severe and deep. An intermediate-depth rupture at 110 km within the South American subduction zone spreads moderate shaking widely rather than concentrating it in a small footprint. Cotality's assessment noted roughly 1.5 million people at intensity VII or greater, while Bogotá and Medellín — where insured value sits — saw only light to moderate shaking.

The modelled loss is bounded but ambiguous. Cotality published an insurable loss range of US$2–4 billion, with the caveat that actual insured market losses should be only a small fraction of it, because damage concentrated in non-engineered rural construction. Mid-August market commentary pointed the same way: claims across commercial and residential property, but low penetration near the epicentre.

The public-sector number does not exist yet. The World Bank disbursed US$200 million on 13 August and launched a Global Rapid Post-Disaster Damage Estimation (GRADE) exercise. GRADE measures direct physical damage: not insured loss, not reconstruction cost, and not yet published.

Set that beside the season's other events. Verisk estimated insured losses from the magnitude 6.8 Kumamoto earthquake of 28 July 2026 at JPY 220–340 billion (roughly US$1.4–2.1 billion). Swiss Re put first-half global insured natural catastrophe losses at US$42 billion against about US$100 billion of economic losses; Munich Re's parallel estimate was US$44 billion. Two credible institutions, measuring the same six months under different definitions, land US$2 billion apart before any earthquake enters the picture.

Four Quantities That Get Called "The Damage Estimate"

Distinguishing them is most of the analytical work.

Number	Produced by	Measures	Typical first release
PAGER alert	USGS	Probable fatality and economic loss ranges from modelled shaking exposure	Generally within 30 minutes
Modelled insurable loss	Catastrophe model vendors	Simulated insured physical damage under stated inclusions	Days
Direct physical damage	World Bank GRADE and equivalents	Replacement cost of damaged assets, insured or not	About three weeks
Industry loss index	PCS, PERILS	Aggregated actual insurer paid and reserved losses	Six weeks to several months

PAGER deserves care because it reaches headlines first. Its alert bands are fixed: green is zero fatalities and under US$1 million; yellow is 1–99 fatalities or US$1–100 million; orange is 100–999 fatalities or US$100 million–1 billion; red is 1,000 or more fatalities or over US$1 billion. Those bands are wide by design. USGS documentation is explicit that PAGER preserves uncertainty measures so users can gauge whether an alert is over- or under-stated, and that alerts are re-issued as shaking maps improve. A red alert describes a probability distribution, not a figure.

Vendor estimates carry their own boundary. Verisk's Kumamoto release stated that its range excludes uninsured property, infrastructure, automobiles, business interruption, workers' compensation, marine and aviation risks, loss adjustment expenses, demand surge, and non-modelled perils including landslide and fire following. It attributed the range width to uncertainty in fault slip distribution, ground-motion intensity, damage estimation and take-up rates. A later figure including any excluded category is not a revision; it is a different measurement wearing the same label.

The Transmission Mechanism: How a Number Becomes a Cash Flow

These distinctions matter because reinsurance and insurance-linked securities contracts name a specific number and reporting cycle. Nothing pays out on a headline.

Index reporting cadences are published, not improvised

Verisk's Property Claim Services designates a US catastrophe when insured property losses exceed US$25 million and a significant number of policyholders and insurers are affected; above a US$250 million preliminary estimate, PCS resurveys until the estimate is final. Cadence differs by territory: C$25 million in Canada with 60-day resurveys; MX$300 million in Mexico with updates every 60 days above MX$950 million; US$2 billion for Japan and US$2.5 billion for the rest of Asia-Pacific, updated every 90 days; US$500 million for Latin America, and US$2 billion for Caribbean events.

PERILS, which covers European and selected other markets, publishes its first index value six weeks after an event, then updates at three, six and twelve months, with reporting closed in any case after 36 months.

Two consequences follow. An event can be economically severe and still generate no index number if it falls below the territorial threshold — the Kumamoto range straddles the US$2 billion Japan threshold. And the interval between updates is the interval during which an index-triggered instrument's value is unknowable from public data.

Trigger design decides how long the ambiguity lasts

Catastrophe bond structures differ sharply in settlement speed. Actuarial literature on insurance-linked securities describes four families. Parametric triggers key off measured hazard parameters such as magnitude or central pressure, delivering the shortest settlement and the most basis risk for the sponsor. Modelled-loss triggers run event parameters against a fixed exposure database. Industry loss index triggers spare the sponsor from disclosing its portfolio, at the cost of a gap between its own book and the index. Indemnity triggers key off actual losses, minimise basis risk, and require claim verification.

The difference is visible in real transactions. Hurricane Melissa made landfall in Jamaica on 28 October 2025; the World Bank announced a full 100 percent redemption of a US$150 million parametric catastrophe bond on 7 November 2025 — ten days later — after the calculation agent evaluated central pressure and storm path as reported by the National Hurricane Center. No claim file was opened to reach it.

Indemnity structures work on another scale entirely. Because loss and loss adjustment expense can take many years to resolve, indemnity bonds typically embed commutation provisions forcing final settlement on a specified date even if open claims and reserves remain. An illustrative structure in the actuarial literature pairs a three-year risk period with a commutation and final payment date two full years after coverage ends. That tail is a contractual admission that the number is not knowable when the risk period closes.

What the 1994 Northridge Loss Development Curve Actually Shows

The cleanest documented case of earthquake loss development remains Northridge, 17 January 1994. A National Academies review reproduced the PCS estimate sequence:

Date	PCS insured loss estimate
February 1994	US$2.5 billion
April 1994	US$4.5 billion
August 1994	US$7.2 billion
January 1995	US$10.4 billion
July 1995	US$12.5 billion

A fivefold increase across seventeen months, still moving at the last observation. The stated reason was mechanical: the full extent of earthquake damage is not known until reconstruction and repair are complete, so estimates grew monthly for more than two years.

New Zealand's Canterbury sequence of 2010–2011 extended the pattern. By 30 June 2016 — nearly six years on — private insurers and reinsurers had paid more than NZ$18 billion, with 93 percent of residential and 94 percent of commercial claims settled. In the twelve months to that date the Earthquake Commission transferred a further 1,620 properties to private insurers, 394 in the final quarter. Earthquake is often filed as short-tail; Canterbury says otherwise.

Two structural features drive the pattern. Concealed damage in foundations, shear walls and cladding surfaces only during repair. And demand surge — local inflation in labour and materials after concentrated reconstruction — is excluded from several vendor estimates at the outset, so it enters the number later by construction.

Two Cases Where the Number Did Not Keep Rising

The "estimates always climb" heuristic fails often enough to be unsafe. Hurricane Sandy is the plateau case: PCS issued a first estimate of US$11 billion in November 2012, raised it to US$18.75 billion in January 2013, then held that figure through the March and May 2013 resurveys. The initial move was large; everything after it was flat.

Windstorm Nils, event end date 13 February 2026, is the reversal case and the more instructive one. PERILS reported €586 million at six weeks, raised it to €767 million at three months, then reduced it to €695 million in the six-month report published on 17 August 2026. A fourth estimate is due 13 February 2027. Anyone treating the three-month figure as a floor was positioned against a 9 percent downward revision.

Revision direction depends on which measurement is being refined. Figures anchored on modelled exposure tend to understate, because concealed damage and demand surge sit outside the model. Figures anchored on incomplete claim reporting can overstate, because reserves set under uncertainty are released as files close.

Economic Damage and Insured Loss Are Not the Same Series

The largest single source of confusion is treating economic damage as a proxy for insured loss. The gap is structural and, in seismic regions, very large. California is the documented case. The state Department of Insurance data call for the 2025 experience year, published on 3 July 2026, recorded 1,591,648 residential earthquake policies against 12,754,225 residential dwelling policies — a take-up rate of 12.48 percent — on US$1.52 billion of earthquake written premium. In the best capitalised earthquake market in the world, roughly seven in eight residential structures carry no earthquake cover.

The same arithmetic runs harder elsewhere. Swiss Re noted that the Venezuelan seismic sequence in the first half of 2026 caused approximately US$20 billion in economic damage while no reliable insured-loss estimate was available, because low penetration implies a small insured share. Türkiye's February 2023 Kahramanmaraş earthquakes set the reference scale for the public-sector number: the World Bank's GRADE report, released 27 February 2023, three weeks after the event, put direct physical damage at US$34.2 billion, equal to 4 percent of 2021 GDP — 53 percent residential (US$18 billion), 28 percent non-residential (US$9.7 billion), 19 percent infrastructure (US$6.4 billion) — and stated that recovery and reconstruction costs would be much larger, potentially twice as large, with damage estimates liable to increase further.

A Colombian GRADE figure, when it appears, will therefore cover insured and uninsured assets alike. Reading it as an insurance market loss would overstate market impact by an unknown multiple.

Where This Argument Is Weakest

Several conditions blunt the framework, and they should be stated plainly.

• The distinctions collapse in high-penetration markets. Where take-up approaches full coverage and building stock is uniformly engineered, economic damage and insured loss converge and the early modelled estimate is a reasonable proxy. This framework earns its keep in mixed and low-penetration markets, not everywhere.
• Historical development factors are not stable. Northridge developed under 1990s claims handling, litigation practice and building codes. Satellite tasking, aerial imagery and remote adjusting have compressed part of that timeline. Assuming a fivefold multiple today because Northridge produced one is an error of the same family as trusting the first number.
• Direction of revision is genuinely uncertain in advance. The Nils case is not an exception to be explained away. Reserve release is a real mechanism, and index estimates fall as well as rise.
• Depth cuts both ways. A 110 km hypocentre spreads moderate shaking widely, which can produce large aggregate non-structural and contents claims across a broad footprint even where no building is destroyed. The rural-damage framing may understate the eventual commercial claim count.
• Aftershock sequences can reset the clock. Where a sequence rather than a single rupture is involved, event definition and hours-clause allocation become contested, and the eventual industry number depends on contractual interpretation rather than measurement.

What to Watch Next Week

• Publication of the Colombian GRADE assessment, and whether commentary distinguishes direct physical damage from insured loss. The Türkiye precedent suggests roughly three weeks.
• Any first modelled insured-loss range from a catastrophe model vendor. The relevant detail is the exclusion list, not the headline range; compare it line by line with the Kumamoto exclusions.
• Whether the event clears the documented US$500 million Latin America index threshold. If it does not, no industry loss index number will exist, and index-triggered instruments settle at zero regardless of humanitarian severity.
• The Kumamoto trajectory against the US$2 billion Japan designation threshold. A range straddling a threshold is where index and indemnity outcomes diverge most sharply.
• Second-half seasonality framing. Swiss Re noted that the second half of the year has historically carried roughly 58 percent of annual global insured natural catastrophe losses. A below-trend first half of US$42 billion is not evidence about the full year.

Concrete Framework — The Watch List

1. Label every figure before using it. Assign each published number to one of four categories: probabilistic alert, modelled insurable, direct physical damage, or settled and indexed insured loss. Numbers from different categories should never be plotted on one series.
2. Read the exclusion list before the headline. Business interruption, loss adjustment expense, demand surge, landslide and fire following are commonly excluded at first release; their later inclusion looks like deterioration and is not.
3. Check the territorial threshold first. Below it, there is no index number to trigger anything, however severe the event.
4. Diarise the cadence, not the news cycle. Six weeks, then three, six and twelve months for PERILS, closing at 36 months; 60-day resurveys in Canada and Mexico, 90-day in Asia-Pacific. Movement outside those dates is commentary.
5. Identify the trigger type before forming a view. Parametric can resolve in about ten days, as the Jamaica bond did; indemnity can carry a commutation date two years past the end of the risk period.
6. Hold a symmetric prior on revision direction. Sandy plateaued after one large step; Nils fell 9 percent between reports; Northridge quintupled. All three are normal.
7. Apply a penetration haircut explicitly. Where take-up data exists — 12.48 percent residential in California for 2025 — use it. Where it does not, treat economic damage as an upper bound of unknown tightness, not a scaled proxy.
8. Treat unscheduled revisions as scope changes. A figure moving outside its reporting window usually signals a change in event definition, which matters more than the size of the change.

The Honest Position

As of mid-August 2026 the insured cost of the Colombian earthquake is not known, and no published figure claims to know it. This framework does not shorten that wait. It says which figure answers which question, and when the next scheduled answer is due. Any narrative resolving the uncertainty faster than the reporting calendar is running ahead of the data.

This article is analysis of publicly reported figures and market structure. It is not investment, insurance or financial advice.