Modeled Loss

Yale Climate Connections' confirmation that Earth just recorded its hottest August on record — with an intensifying super El Niño expected to make 2026 the hottest year ever, and 2027 projected even hotter — is not a background fact. It is an active variable in the loss models. The standard hurricane and wildfire exceedance-probability curves are calibrated against historical frequency-severity distributions. When sea surface temperatures in the Atlantic and Gulf of Mexico are running at record levels, the frequency and intensity assumptions embedded in those models are almost certainly underestimating the current risk. The model is a hypothesis built on historical data; the climate is running an experiment on a different dataset.

The practical implication for the H1 2026 result that Theo and Margaret are celebrating is this: the underwriting income was earned in the off-peak season. The Atlantic peak runs August through October. We are now in the window where the model-to-actual gap is most consequential. A record-temperature August in a super El Niño year does not straightforwardly amplify Atlantic hurricane activity — El Niño actually tends to increase wind shear in the Atlantic basin, which suppresses hurricane formation — but it does amplify the severity of any storm that does form, particularly in terms of storm surge and rainfall-driven flood losses. The secondary perils embedded in those storms are precisely where the models have the worst track record. Demand surge post-landfall, inland flooding from rainfall, and the expansion of insured values into coastal zones all create an actual-to-modeled loss ratio that has consistently run above 1.0 in recent major events.

I want to address something Soren raised at the Cat Bond Desk: the Harbor Crest Re structure for Porch Group covers wildfire and fire-following-earthquake alongside named storm. In a record-heat August under a super El Niño, wildfire risk in the western U.S. does not diminish — El Niño's drying effects in the Southwest are well-documented. A multi-peril cat bond that bundles Florida named storm with California wildfire is holding correlated climate tail risk in the same structure. The aggregate-level expected loss of 2.5% across the outstanding market does not capture the correlation uplift when a single season delivers both a Florida storm and a California wildfire. The model assumes those perils are largely independent. The climate record is making that assumption less defensible each year.

The CatIQ revision from CAD 850 million to CAD 923 million — an 8.6% upward move at the three-month mark — is squarely within the documented pattern of secondary-peril loss development. Severe convective storms in Canadian prairie provinces are a peril where the event catalog is thin relative to the Gulf Coast hurricane record; exceedance-probability curves for Saskatchewan and Manitoba are built on a shorter and less uniform observational record, which means the model confidence interval is wider and the upward development risk is higher than users typically discount.

The NOAA confirmation that the contiguous U.S. just recorded its hottest August, hottest summer, and hottest year-to-date in history is the structural backdrop against which this specific event revision must be read. Record heat is not just a wildfire preconditioner for California — it is a moisture-flux driver that intensifies the convective available potential energy (CAPE) environment for Great Plains and Canadian prairie severe weather. The model is a hypothesis built on historical temperature-moisture relationships; if those relationships are shifting, the EP curve is already stale by the time the event season ends.

I want to flag something that Soren raised — whether the Saskatchewan/Manitoba figure triggers aggregate ILS structures. That is the right question, but the harder one is whether the models that priced those structures adequately captured the frequency distribution of multi-storm aggregate accumulations in the Canadian prairie corridor. My read from the structure of this event: probably not fully. The three-month development pattern here is consistent with demand surge (construction costs, labor scarcity in semi-rural prairie markets) on top of a modeled footprint that likely underestimated commercial property exposure in the affected Saskatchewan agricultural and industrial zones.

Washington State wildfires: $600M-plus in insured losses from a peril that most standard cat models treat as a secondary — meaning the hazard footprint, ignition probability, and suppression-cost curves embedded in vendor models were calibrated on a Pacific Northwest fire history that looks increasingly obsolete. Commissioner Kuderer's office is now fielding the claims adjustment and coverage-dispute load from a fire season that 'plagued Washington for nearly half the summer,' per Insurance Journal. That duration signal matters more than the point estimate: it suggests a persistent, aggregate accumulation event rather than a single ignition shock, exactly the mode that EP curves underweight because they're built around discrete peak-season events.

Hurricane Lowell approaching Hawaii is the corpus's live catastrophe signal, and it is worth calibrating carefully. Yale Climate Connections reports potential for isolated 16-inch rainfall totals, flash flooding, stream overflow, power outages, and storm surge. The independent model read tags this as Consensus on the meteorological forecast. What the corpus does not provide — and what matters enormously for the loss assessment — is the attachment of Hawaii property insurance coverage to the flood peril. Standard homeowners policies in Hawaii, as elsewhere in the U.S., exclude flood. NFIP penetration in Hawaii is not reported in today's corpus, but the historical pattern for island states with significant rainfall events is low flood-insurance take-up relative to actual exposure. The insured loss from Lowell may be a fraction of the economic loss, concentrated in the wind and storm-surge-classified damage rather than the rainfall-driven inland flooding. This is the secondary-peril gap made visible: a named hurricane produces losses that the naming convention suggests are covered, but the dominant damage pathway — freshwater flooding — sits in the uninsured bucket.

The Harbor Crest Re cat bond for Porch Group, which Soren flags, is the modeled-loss problem statement in structured form. A $100M tranche covering named storm, winter storm, severe weather, wildfire, and fire-following earthquake as a bundled multi-peril trigger means the aggregate exceedance probability is a convolution of five individual EL curves. The correlation assumptions between, say, a California wildfire year and a Gulf hurricane year are not stable under climate non-stationarity. The models that price that tranche were calibrated on a historical event catalog that is increasingly stale for the secondary-peril components. Margaret may be right that ILS spreads look attractive at roughly 2.2x market EL, but the 'market EL of 2.5%' is only as good as the models producing it.

Munich Re's cyber call from Golling at RVS is where I want to push back on both the cat-bond and cycle framing. Cyber is not a peril in the physical-damage sense; its loss distribution has essentially no validated empirical catalog. The 'protection gap' framing assumes we know what the expected loss looks like. We do not. Cyber cat-bond structures exist, but the model uncertainty is a multiple of what it is for Atlantic hurricane. Calling it a 'major business opportunity' without acknowledging that the EP curve for a systemic cyber event is largely an assumption rather than a calibrated estimate is the model error I most worry about in 2026.

Swiss Re's $200 billion premium figure is a market-sizing exercise, not a loss model. I do not dispute the capex super-cycle thesis — the data-centre buildout and the energy transition are real, and so is the insurance gap they create. But the modeling community is not ready for the perils that data-centre concentration produces, and the absence of a mature event catalog for this risk class is the most important fact missing from the coverage.

Margaret and Soren are debating capital capacity and cycle timing. I want to focus on what neither of them is pricing: the secondary-peril exposure embedded in data-centre risk. A Phoenix hyperscale campus faces extreme heat-driven cooling-system failure as a correlated loss driver alongside named-storm wind. An offshore wind farm in the Gulf faces both named-storm and subsea cable damage in ways that do not appear in any standard property cat model. The correlation of data-centre business interruption losses to physical damage losses — and the duration of those BI losses while a campus is reconstructed — is essentially unmodeled territory. The event catalog for AI infrastructure catastrophe does not exist.

The relevant historical parallel is the early 2000s offshore energy accumulation in the Gulf of Mexico before Katrina and Rita exposed the gap between engineer-designed loss estimates and actual insured outcomes. The model was a hypothesis. Katrina was the experiment. The gap was tens of billions of dollars. I would not wait for the equivalent data-centre Katrina to discover how far the EP curve is from the actual loss distribution on this new peril class.

The Insurance Journal's reporting that the 2026 Atlantic hurricane season is the most tranquil since 1941 is meteorologically coherent with the El Niño signature — dry air, elevated wind shear, and a stable thermocline in the Main Development Region. From a modeled-loss standpoint, suppressed seasonal activity does not rewrite the exceedance-probability curve; it samples the quiet tail of the distribution. The EP curve for Gulf and Southeast coastal wind still peaks where it always has. One quiet season is not a structural regime shift in the hazard, and I'd caution against underwriting decisions that treat 2026's outcome as signal rather than noise.

Hurricane Lowell is the more interesting story today for my purposes. Yale Climate Connections reports a former Category 5 — now tracking toward Hawaii's northwestern islands after a hard-right turn. Hawaii is chronically under-insured for wind and storm surge relative to the Gulf Coast; the residential insurance market there is thin, NFIP penetration in the northwestern islands is near zero for the relevant communities, and reinsurance treaty structures for Hawaii wind are not standardized across the primary market the way Florida programs are. If Lowell makes significant landfall on the main Hawaiian islands, the modeled versus actual gap could be material — not because the wind models are wrong, but because the demand-surge and claims-handling infrastructure assumptions are calibrated to Gulf Coast event responses, not Pacific island geographies.

Margaret is correct that the $1.34 trillion in Bermuda gross claims over a decade reflects a sector that priced the hard-market years properly. But I'd add a qualifier: the BMA figure spans 2016-2025, a period that included years where secondary perils — severe convective storms, wildfire, inland flood — consistently ran above model output. The 'record capital is comfortable capital' narrative holds only if we're confident the model inventory for those perils has closed its gap with observed loss experience. I am not yet confident of that, and a quiet Atlantic does nothing to validate or invalidate SCS or flood model accuracy.

Aon's $800 billion capital headline and the projected 10% property rate decline at January 1, 2027 deserve a peril-by-peril interrogation before anyone treats them as an all-clear. The capital abundance story is real — but capital is not uniformly distributed across risk types, and rate relief at the aggregate reinsurance level can mask very different conditions in specific peril corridors.

The ILS deal flow in this corpus is instructive. Armor Re II for American Coastal is a pure Florida named-storm placement. Harbor Crest Re for Porch Group stacks wildfire and fire-following-earthquake alongside severe convective storm and winter storm — a multi-peril structure that tells you something about where cedents are still hunting capacity. These are not perils where the model-to-actual gap has closed. Severe convective storm and wildfire remain the two largest sources of model underestimation in the current property cat environment. Annual SCS insured losses have run materially above long-run model expectations for several consecutive years, and wildfire loss distribution in California has shown non-stationary behavior that standard exceedance-probability curves were not calibrated to capture.

Soren Vaeth's Cat Bond Desk reads the 5.53% spread as still doing 'real work' against a 2.50% market-level expected loss. I do not dispute the arithmetic — but that market-level EL is an average across a portfolio where secondary-peril underestimation has been a persistent feature. If the models are carrying SCS and wildfire EL at, say, 0.8% when the actual run rate is 1.5%, the effective spread-over-realized-loss is narrower than the market believes. That is the gap the softening cycle will eventually expose.

The corpus contains a headline that deserves more attention than its low insurance-relevance score suggests: Tropical Storm Edouard delivered 'disproportionate flash flooding' to east Texas, with Yale Climate Connections explicitly flagging that the event 'punches above its weight' in terms of flood loss. This is a secondary-peril event — a named storm that underperforms on wind but overperforms on precipitation and inland flood. East Texas is precisely the geography where the modeled-vs-actual gap is widest: NFIP penetration is low, inland flood is poorly captured by standard named-storm peril models, and demand surge following any event that closes roads and disrupts construction supply chains amplifies the ultimate loss.

What makes this event modeler-relevant beyond its immediate loss footprint is the Lowell signal embedded in the same article: next week, Hurricane Lowell could veer sharply toward Hawaii. Hawaii is a region where the exceedance-probability curve has very limited historical event catalog support — almost no major hurricane landfalls in the modern record. If Lowell does make a Hawaii approach, the modeled loss will be driven almost entirely by parametric assumptions about building stock vulnerability, and the actual loss will test whether those assumptions hold in a high-value, high-density coastal market with severe access and demand-surge constraints. I do not have loss estimates for either event in this corpus, and I will not fabricate them. But the peril-region combination — east Texas inland flood plus potential Hawaii named storm — is a simultaneous multi-region stress test of secondary-peril model adequacy.

Soren Vaeth's point about the Armor Re II deal covering Florida named storm is worth connecting here: the same model uncertainty that plagues east Texas inland flood affects the attachment and exhaustion probabilities on Florida named-storm cat bonds. If the precipitation-driven loss component is systematically under-modeled, the 2.5% market-level expected loss figure on the outstanding ILS book may be understated. The model is a hypothesis; Edouard is an experiment.

A federal judge striking down New York's Climate Change Superfund Act deserves more actuarial attention than it is getting in the reinsurance press. The ruling — that the legislation conflicts with federal authority — removes a mechanism that would have required fossil fuel companies to contribute to climate-adaptation costs. From a modeling standpoint, this is not just a legal outcome; it is a statement about who absorbs the unmodeled residual of climate non-stationarity. When the legal system declines to assign those costs to emitters, they flow somewhere else — to state insurers-of-last-resort, to NFIP, to uninsured homeowners, and ultimately to the exceedance-probability tails that no vendor model has fully repriced for a non-stationary climate.

Soren's read on the Florida named-storm deals is correct as far as it goes, but I want to sit with the model uncertainty underneath it. The Armor Re II transaction for American Coastal on Florida named storm is priced against a vendor EP curve that reflects historical Atlantic storm climatology. The question the model cannot fully answer is whether the 2024-2026 Atlantic seasons — including whatever this season produces before November — represent a new distribution or a cluster within the old one. Cat bond investors are pricing a hypothesis. Each landfall is an experiment that either confirms or falsifies it.

The Zelenskyy warning to airlines and insurers about Russian airspace is a secondary-peril signal worth flagging. Aviation war risk is not a cat-model peril in the traditional sense, but it is an accumulation risk for aviation war-risk underwriters that sits entirely outside property-catastrophe EP curves. The corpus carries a cross-source count of 5 on this story, which suggests it is getting genuine traction in specialist markets. I cannot quantify the exposure from this corpus, but the directional signal — a state actor explicitly naming insurers as parties who should update their risk assessments — is not something to dismiss.

The Yale Climate Connections report on the fast-developing Gulf tropical system — described as 'small but potentially potent' over warm Gulf waters — is exactly the scenario where the gap between modeled and actual loss becomes the entire story. Fast-spin-up systems are structurally underrepresented in historical event catalogs. The standard Atlantic hurricane catalogs used by RMS, AIR, and KCC are built on historical storm tracks and intensification rates. Rapid intensification events — defined as a 35-knot or greater increase in maximum sustained winds in 24 hours — have increased in frequency in the Gulf of Mexico as sea surface temperatures have risen. The event catalog is a backward-looking instrument; the Gulf it was calibrated on is not the Gulf of today.

For the Texas-Louisiana landfall scenario, the modeled-loss framework faces additional complexity from secondary perils. Tropical systems that make landfall near the Texas-Louisiana border generate not just wind losses but storm surge and inland flooding — and the exceedance-probability curves for compound flood-wind events are significantly less well-characterized than either peril in isolation. The CMIP7 emissions scenarios that Carbon Brief covered today are directly relevant here: the scientific consensus embedded in those scenarios increasingly supports a non-stationary view of Gulf SSTs and hurricane intensification rates. A model calibrated on the 20th-century Gulf is running on a hypothesis that the 21st-century Gulf is actively disproving.

Margaret Ennis is right that a tropical storm producing modest insured losses is ambiguous for the pricing cycle. But from a modeling standpoint, the right frame is different: a small storm that produces outsized inland flood losses relative to its wind speed is exactly the kind of event that exposes secondary-peril undermodeling. The model says 'small storm, small loss.' The loss run may say otherwise.