Insured losses from natural catastrophes surpassed $100 billion for the sixth consecutive year in 2025. However, the absence of peak-zone hurricane landfalls offers a window to address critical gaps in catastrophe models.
In this article…
- How the LA wildfires revealed limitations in urban fire-spread modelling
- What the Myanmar earthquake tells us about California tail risk
- Why severe convective storm losses remain elevated at $50 billion annually
- The tropical cyclone flood component that models underweight
- Stress-testing strategies to prepare for inevitable surprises
Introduction
Global insured natural catastrophe losses reached approximately $105 billion by the end of Q3 2025, according to Gallagher Re estimates. That places the year below the ten-year average of $130 billion, despite opening with the costliest wildfire event on record. Claire Souch, independent consultant, gathered four senior practitioners to examine what the year’s events reveal about modelling assumptions and operational readiness.
The panel comprised of Cameron Rye, Director of Natural Catastrophe Analytics at Willis Re; Nalan Senol Cabi, VP, Catastrophe Research Engineer at Arch Insurance Group; Nigel Winspear, Head of Natural Catastrophe Analytics Research at Sompo International; and Robert Stevenson Director of Cat Risk and Reinsurance Analytics at Lloyd’s.
Urban configuration remains the critical wildfire modelling gap
The January LA wildfires caused approximately $30 billion in insured losses, making them the costliest wildfire event in industry history. Nalan Senol Cabi characterised the event not as a black swan but as confirmation of two converging trends. Accelerating development in the wildland-urban interface combines with persistent modelling shortfalls.
The biggest gap today is urban configuration. Current models are sophisticated at predicting how fire spreads through vegetation. But there’s room to improve with the transition phase when fire leaves the trees and enters the city.
The wildland-urban interface across the United States has expanded by 31% in recent decades. Some 44 million homes now sit in high-risk zones. Senol Cabi noted that vendors are not capturing more fire ignitions per se; rather, human settlement has encroached aggressively into fuel sources. The physics of structure-to-structure spread differ fundamentally from wildland fire behaviour. This could mean that models underweight the tail risk associated with dense urban burning.
Vapor pressure deficit, an atmospheric signal which means drier air, which dries out fuels and makes fires more likely and more intense, has been rising and offers one avenue for model improvement. Senol Cabi expressed optimism that artificial intelligence can ingest dynamic fuel changes and atmospheric signals more rapidly than previous generations of models.
Underwriters should scrutinise exposure concentrations in the wildland-urban interface and challenge vendors on structure-to-structure fire-spread assumptions.
The Myanmar earthquake as an analogue for California tail scenarios
A magnitude 7.7 earthquake struck Myanmar on 28 March 2025, causing approximately $1.3 billion in domestically insured losses and over 5,400 fatalities. Nigel Winspear used the event as an analogue for semi-to-full-length ruptures on the San Andreas fault system. Such ruptures remain in the tail of California’s earthquake rupture forecast at return periods exceeding 500 years.
“500 kilometre rupture. That’s approximately the distance between the southern San Francisco Bay area and the northern suburbs of Los Angeles. These are tail scenarios, but globally we have now seen one of these events.”
The Myanmar rupture exhibited supershear propagation. This occurs when fault slip velocity exceeds the shear wave speed and generates a ground shockwave analogous to a sonic boom. Winspear noted that vulnerability functions in catastrophe models do not account for supershear effects because empirical experience is limited.
The collapse of a 33-storey skyscraper under construction in Bangkok demonstrated the far-field impact of long-period ground motions. The building site lay over 600 kilometres from the southern end of the rupture. Soft soil amplification in Bangkok elevated shaking intensity to Modified Mercalli Intensity V. This should not damage properly engineered structures but proved sufficient to trigger resonance in the incomplete tower. Winspear directed attention to the Los Angeles basin, a large sedimentary structure with significant potential for similar long-period amplification.
Exposure managers should verify whether builder’s risk portfolios in seismically active regions incorporate long-period ground motion and resonance assumptions.
Source: Sompo International
Severe convective storms sustain $50 billion annual losses
Insured losses from US severe convective storms reached approximately $50 billion in 2025, the third consecutive year above that threshold. Cameron Rye presented data showing 2023 at $59 billion and 2024 at $54 billion, adjusted for inflation.
For many people in the industry, this sudden jump was surprising. Part of that was because severe convective storms had traditionally been viewed as a secondary peril.
Two factors dominate the trend: exposure growth and construction cost inflation. Hotspot regions including Tornado Alley and Dixie Alley have added roughly 100,000 new properties annually since 2000. Asphalt roof replacement costs have increased by approximately 250% since the turn of the century.
Research from Willis Reseach Network partnership with Columbia University reveals changing environmental conditions. Conditions conducive to severe tornado outbreaks have increased two to three times since the 1970s in parts of the southeast United States. Detection in observational records remains difficult because observations rely heavily on human reports and carry inherent bias.
Rye noted that most major vendors have released updated convective storm models in the past 12 months. These converge on a consensus that $50 billion represents a typical annual loss in the current environment. He cautioned that the industry should not be surprised by $70 billion or $80 billion years in the near future.
Syndicate and company exposure managers should benchmark their severe convective storm modelled losses against the new $50 billion industry baseline.
Tropical cyclone flood modelling has room to improve
Nearly 200 flood events were documented globally in 2025 through media reports, humanitarian organisations and satellite imagery. Insured losses remained limited because most events occurred outside key exposure centres or in low-penetration markets. Nalan Senol Cabi emphasised that absence of insured loss should not be mistaken for absence of risk.
The Camp Mystic tragedy in Texas illustrated the speed at which flash floods can develop. The Guadalupe River rose approximately 26 feet in under 45 minutes, a rate historically rare and extremely dangerous.
Hurricane Melissa struck Jamaica as a Category 5 storm, causing catastrophic flooding. Economic losses are projected at nearly 30% of GDP, with 1.5 million people affected according to UN estimates. Cyclone Alfred delivered Brisbane its wettest day since the devastating 1974 floods. Despite carrying weaker wind speeds than other historic cyclones, its slow pace and high moisture content extended the flood footprint. Damage reached well beyond the wind radius.
The panel noted that prolonged atmospheric conditions are causing storms to linger over land and deposit unprecedented rainfall totals. Rapid intensification research suggests that storms intensifying quickly before landfall may produce increased precipitation, compounding flood exposure.
Cat modellers should challenge vendor assumptions and methodologies when validating TC-induced flood and storm surge components of the models.
The 2025 Atlantic hurricane season: potential without impact
The Atlantic hurricane season produced three Category 5 storms for the first time since 2005. That year, Katrina, Rita and Wilma devastated the Gulf Coast. Sea surface temperatures remained near record highs, yet the United States recorded no hurricane landfalls for the first time in a decade.
Cameron Rye attributed the outcome to favourable steering flows. The Bermuda High positioned further east than average, directing storms into the open Atlantic rather than toward the US coastline. Melissa, Erin and Humberto all reached Category 5 intensity, demonstrating the basin’s capacity for major storms.
“The potential was there. We could have had large losses. We should remain aware that next year we could come back with large losses.”
The panel advised against interpreting the benign landfall record as a signal of reduced risk. Basin activity confirmed that thermodynamic conditions can support multiple extreme storms in a single season when steering patterns align unfavourably.
Reinsurers and cedants should resist pressure to reduce hurricane pricing based on a single benign season. Underlying sea surface temperatures remain elevated.
Operational discipline and the PRA DyGIST exercise
Robert Stevenson highlighted that Lloyd’s market messaging emphasised discipline in exposure management alongside pricing discipline. Accurate, timely and complete data form the foundation for all modelling applications; without operational rigour, sophisticated models cannot deliver reliable results.
If your exposure management is in good shape, if it understands its data, methods of placement and timeliness, that’s the level you need. You must answer questions from stakeholders and conform to the standard expected of entities like ourselves.
The Prudential Regulation Authority’s DyGIST exercise, scheduled for May 2026, will issue three consecutive weekly event scenarios requiring rapid loss estimation responses. Stevenson described it as a stress test of operational capability across property and non-property lines. Participants must demonstrate command of aggregates, probable maximum losses, deterministic scenarios and probabilistic modelling.
Stevenson referenced Lloyd’s publication of an extreme disaster scenario involving a Pacific Northwest earthquake with tsunami. The largest loss component arose from strikes, riots and civil commotion rather than physical damage. Secondary perils including business interruption and supply chain disruption may generate surprises that current frameworks underquantify.
Managing agents should audit exposure management workflows before May 2026. Ensure systems can deliver rapid, accurate loss estimates under time pressure.
Disaster scenarios: Hanoi and Hong Kong earthquake risk
Nigel Winspear presented two earthquake scenarios for consideration in stress-testing programmes. The first involves Hanoi, built on the Red River Delta soft soils at the terminus of the Red River fault system. This fault is a sister system to the one that ruptured in Myanmar. No significant earthquake has struck near Hanoi in recorded Vietnamese history. Yet the fault trace is clearly visible and maximum credible magnitude approaches 7.0.
Source: Sompo International
There are more than 430 high-rise buildings of 30 storeys or more in Hanoi. It’s unlikely they are designed against long-period ground motions.
The second scenario involves a repeat of the 1918 Nan’ao Island earthquake on the Fujian-Guangdong border, approximately 300 kilometres from Hong Kong. That magnitude 7.4 event remains the only earthquake to have caused damage in Hong Kong since 1842. The territory now contains over 9,000 high-rise buildings, many constructed before 1997 without seismic design requirements. Local amplification from soft sediments overlying granite bedrock introduces additional uncertainty.
Insurers with Asia-Pacific earthquake exposure should incorporate Hanoi and Hong Kong seismic scenarios into realistic disaster scenario libraries.
Data quality offers the highest return on investment
Nalan Senol Cabi argued that the current period of relatively low insured flood losses provides an opportunity to repair data quality. Flood is a location-specific, high-resolution peril where geocoding accuracy determines underwriting precision.
“If you’re sitting on unknown secondary modifiers or ZIP code level geocoding in 2025, you are underwriting blind. Cleaning that location data is the highest ROI activity you can do right now.”
She recommended using stochastic catalogues for active stress testing. Filter specifically for events that strike specific portfolio concentrations even where models assign low probability. Identifying the worst plausible day for a unique portfolio reveals vulnerabilities that aggregate statistics can obscure.
Cameron Rye echoed the theme. Adjusting historical convective storm losses using standard indices may understate true trends. Asphalt roof replacement costs have diverged significantly from consumer price inflation.
Chief underwriting officers should mandate location data audits and secondary modifier enrichment programmes before the next active loss year.
What’s next
Despite the lack of major hurricane losses, the events of 2025 confirmed that catastrophe models require continuous refinement. Urban wildfire spread, supershear earthquake propagation, tropical cyclone-induced flooding and secondary perils represent areas where vendor capabilities have not kept pace. Strikes, riots and civil commotion demand particular attention. The PRA Digest exercise in May 2026 will test whether London Market participants have the operational discipline to respond quickly.
InsTech’s Exponential Risk conference runs 10–11 March 2026 in the City of London. Sessions will examine coastline characteristics and storm surge, European natural catastrophe frequency trends and public-private partnerships to address protection gaps. Webinars in January 2026 will cover coastline risk and European market dynamics.
Join us for Exponential Risk London 26
➜ Register for Exponential Risk 2026 to continue this conversation with the panel
