Severe convective storms overtook hurricanes in 2025 to become the costliest insured peril of the 21st century. A question now running through carrier underwriting rooms is whether the models used to price that risk are current enough to keep up.
SCS events generated $61 billion in global insured losses last year, the third-highest on record, per Aon's 2026 Climate and Catastrophe Insight report. The Insurance Information Institute (Triple-I), meanwhile, put the US share at $51 billion, the third consecutive year that American SCS losses exceeded $50 billion.
Those figures have sharpened attention on cat modeling methodology. A new white paper published by Karen Clark & Company (KCC) argues that traditional statistical cat models are poorly suited to SCS risk. KCC is a Boston-based catastrophe risk modeler. A new generation of artificial intelligence (AI)-informed models, the paper contends, addresses that gap.
Carriers using older statistical models carry a less current view of the risk than those running physical or AI-augmented systems, according to the KCC paper. In a softening property market, that divergence is producing uneven pricing.
Some carriers are competing aggressively on SCS-exposed accounts, according to Moody's observations and the BCG 2026 Insurance Value Creators report. Others are holding higher retentions and coverage restrictions. BCG found that carriers with superior cat modeling are better positioned for strong returns, while those with weaker risk views face pressure when losses arrive. For a broker placing storm-exposed business, which model a carrier uses now matters.
Traditional statistical cat models fit historical data to distributions and generate hypothetical events from those distributions. That approach works reasonably well for hurricanes, which dissipate in predictable patterns after landfall. Severe convective storms behave differently: highly localized, rapid to develop, and unlike any prior event.
Extrapolation from historical data is unreliable for this peril, according to the KCC paper. Moody's has reached a similar conclusion, noting that inconsistency in historical SCS data makes traditional approaches less dependable.
A second generation of physical models emerged to address that gap. Rather than extrapolating from records, physical models simulate atmospheric conditions in real time through equations of mass and momentum.
KCC's SCS model ingests over 30 gigabytes of satellite, radar, and weather data each day. It produces hail and tornado intensity footprints by 7 AM ET.
Insurers use those footprints to estimate claims exposure from the prior day's storm activity. KCC said the process has run daily since 2018, generating eight years of verifiable output. Those eight years of output form the training foundation for a third generation of AI-informed models.
AI's role in next-generation cat models is to refine the physical structure continuously, rather than replace it. KCC said its scientists archived over 100 terabytes of atmospheric data alongside tens of billions of dollars of claims data since 2018. Machine learning systems trained on those archives would update model components on an ongoing basis, the paper argues. Model updates could shift from months to days.
KCC illustrates one application: training an algorithm to identify derechos using labeled radar imagery. Derechos develop their own internal dynamics and persist in conditions that would not otherwise support convective storms. Whether that approach produces better loss estimates in production is not independently verified in the paper, which is a vendor document.
The broader challenge extends beyond hazard science. Research from Gallagher Re found that economic and societal factors account for 80% to 90% of the long-term rise in U.S. SCS insured losses. Higher construction costs, labor shortages, and a more litigious claims environment drive most of that growth. Improved AI hazard modeling addresses one part of a larger problem.
A separate May 2026 Moody's study illustrated how AI-driven data enrichment shifts SCS risk estimates within existing portfolios. Adding property-level attributes such as roof condition, cladding type, and vegetation produced a modest average reduction in modeled loss across a sample portfolio.
Moody's, however, noted that nearly half of all properties saw a modeled loss change of more than 15% in either direction. The headline figure conceals substantial redistribution within books.
For brokers, the practical implication runs through to the placement conversation. Carriers that have adopted more current modeling tools price SCS-exposed risks differently from those that have not. Those pricing gaps can surface at renewal in ways that are not always explained by changes in the client's own exposure.
