Updated October 1, 2026.
Catastrophe models from RMS, AIR, and Verisk translate physical peril science into probabilistic loss estimates insurers use for pricing, reinsurance, and accumulation control. Each vendor builds stochastic event catalogs, damage functions, and financial engines that roll up to metrics such as average annual loss (AAL), probable maximum loss (PML) at chosen return periods, and occurrence and aggregate exceedance probability curves. The outputs are long-horizon views of tail risk that must be validated, calibrated to experience, and read alongside underwriting judgment.
Where RMS, AIR, and Verisk sit in the market
For decades, property cat risk in North America and many global markets has been shaped by two large independent model families: Risk Management Solutions (RMS) and AIR Worldwide (AIR). Both were acquired by Verisk and now sit under a single analytics umbrella while retaining distinct model lineages, data pipelines, and client bases. Many carriers run one vendor as a primary view, the other as a challenger, or blend outputs in internal aggregation systems.
Practical differences show up in event catalogs, peril coverage, vulnerability assumptions, and financial layering—not in the abstract goal. Each vendor documents peril modules (hurricane, earthquake, severe convective storm, wildfire, flood where offered, and other secondary perils), update cadence, and regional footprints. Underwriters and catastrophe analysts choose a vendor—or a multi-vendor stack—based on portfolio mix, reinsurance program design, and how well model losses reconcile with their own claims history.
Risk managers who do not run models in-house still encounter these tools through insurer submissions, broker analytics, and treaty pricing. Understanding what the vendor is actually quantifying prevents mistaking a modeled PML for a policy limit or a single scenario for a full tail distribution. For broader context, see the hub’s complete guide to catastrophe modeling and how modeled climate trends feed pricing.
How a catastrophe model is built
A cat model is three linked engines plus a roll-up layer. Vendors publish methodology papers; implementation detail varies by peril and region, but the architecture is consistent enough that risk teams can audit each stage.
Hazard and stochastic event generation
The hazard module defines how events are drawn from history, physics, and statistical extrapolation. For hurricanes, that typically includes track, intensity, and landfall logic tied to basin climatology. Earthquake modules use fault and seismicity models. Severe convective storm and wildfire modules use different generators—hail and tornado footprints versus fire spread and ignition patterns tied to fuels, weather, and topography.
The output of this stage is a large stochastic catalog: thousands to hundreds of thousands of events, each with frequency and spatial footprint. Event rates and severities are tuned so that modeled distributions align with historical catalogs and, where vendors offer them, forward-looking or climate-conditioned views that adjust selected peril parameters.
Vulnerability and exposure
Vulnerability functions (damage curves) map hazard intensity at a location to a damage ratio or damage state for a given construction and occupancy class. Exposure inputs describe values at risk: replacement cost, business interruption, contents, deductibles, limits, and secondary modifiers such as year built, roof type, or mitigation features when captured.
Small shifts in vulnerability assumptions can move tail losses materially for high-gradient perils (hurricane wind and storm surge near coasts, earthquake shaking in soft soil, wildfire in the wildland–urban interface). That is why model change management is a standing agenda item at renewals. Property owners see the other side of the same exposure data in policy structure—how Coverage A, B, C, and D apply.
Financial module and metrics engine
The financial module applies policy terms to ground-up damage: deductibles, limits, coinsurance, sublimits, waiting periods on time element coverage, and reinsurance in gross versus net runs. From simulated event losses, the engine builds occurrence and aggregate distributions and summary metrics.
Occurrence loss groups all locations hit by one event before policy terms. Aggregate loss sums all events in a period (often a calendar or treaty year). That distinction drives whether a program responds on a per-event excess-of-loss tower or an aggregate stop-loss layer—topics covered in depth in the hub’s reinsurance treaty guide and quota share versus excess-of-loss structures.
Outputs risk teams actually use
Average annual loss (AAL)
AAL is the mean loss per year across the simulated catalog, net of terms and optionally net of reinsurance. It is a central tendency metric useful for pricing expected loss load and comparing portfolios over time. AAL alone says little about tail risk; two portfolios with similar AAL can have very different PMLs if one concentrates coastal wind or wildfire exposure.
PML and return periods
Probable maximum loss (PML) in industry practice usually means a high quantile of the modeled loss distribution at a stated return period—often 100-year, 250-year, or 500-year occurrence, depending on peril and internal standard. The label “PML” is not interchangeable across organizations unless return period, occurrence versus aggregate basis, gross versus net, and peril scope are defined.
Occurrence and aggregate exceedance probability curves
An occurrence exceedance probability (OEP) curve plots the probability that a single event exceeds a loss threshold. An aggregate exceedance probability (AEP) curve plots the probability that total annual losses exceed a threshold. Reinsurance buyers use OEP views to size event caps; capital and aggregate treaty discussions lean on AEP. Tail Value at Risk (TVaR) captures the average severity of outcomes worse than a chosen PML point—not just the threshold loss.
Gross versus net and model blending
Insurers run gross models to design reinsurance, then net models to reflect ceded recoveries under treaty terms modeled with varying fidelity. Multi-vendor shops may take ensembled or bounded views when RMS and AIR diverge on the same book. The discipline is to document which view anchors pricing versus which view stress-tests the program.
Validation, calibration, and model change
Models are hypotheses backed by data and expert judgment. Validation compares modeled loss distributions and event rates to independent benchmarks: historical industry losses, company experience where credible, and peril-specific studies. Calibration adjusts components—event rates, vulnerability anchors, or financial flags—so that modeled AAL and selected return periods sit inside agreed tolerance bands.
When vendors release major updates, clients rerun baseline portfolios and trace drivers: new hazard science, exposure schema changes, vulnerability retuning, or financial engine fixes. Underwriting and actuarial teams align on whether to phase changes into pricing or absorb them in trend factors. Secondary perils that once sat in “other natural catastrophe” buckets now get explicit modules; that shift alone can reorder rank by peril for inland books.
Honest use of cat output treats the model as one input. It does not replace field underwriting, inspection data, or knowledge of local building code enforcement. It also does not replace reading the policy—modeled limits must match bound coverage, including exclusions that standard property forms leave open.
2026 model themes: climate-conditioned views and secondary perils
Vendor release notes in recent cycles emphasize two parallel tracks: forward-looking hazard views that reflect selected climate science pathways for certain perils and regions, and expanded treatment of perils that drive growing attritional and tail loss outside traditional hurricane and earthquake zones.
Climate-conditioned hurricane and flood-related parameters, where offered, adjust aspects of severity or frequency relative to baseline catalogs. Users choose whether to run baseline, alternate, or blended views; the important operational point is to label results clearly in reinsurance submissions and internal capital discussions so year-over-year comparisons are not apples-to-oranges.
Secondary and tertiary perils—severe convective storm (hail, tornado, straight-line wind), wildfire, and inland flood where modeled—now receive dedicated event generators and vulnerability suites on major platforms. That matters for mid-continent carriers and any account with wildland–urban interface exposure. It also shapes property pricing when modeled tail risk rises even if recent loss history is quiet—see the hub’s overview of hard versus soft market cycles and coverage pricing.
Earthquake, hurricane, and wildfire modules continue to receive regional refinements (landfall logic, fire spread, liquefaction and site amplification where applicable). Risk managers should read vendor release documentation for their perils and territories rather than relying on headline press summaries.
How insurers and risk managers apply the outputs
Property pricing and portfolio steering
Commercial and personal lines insurers use modeled AAL and tail metrics to set rate adequacy targets by territory and construction class, cap aggregate deployment in high-return-period zones, and justify deductibles or sublimits. Large insureds with scheduled locations may see cat analytics reflected in capacity placement and minimum attachment points.
Reinsurance structure and renewal
Reinsurance brokers and ceding insurers use occurrence EP curves to layer event excess-of-loss programs and AEP views to size aggregate covers. Modeled net PML after reinsurance supports discussions of retention, co-participation, and reinstatement economics. When modeled losses shift at a fixed return period, attachment points and rates on excess layers move even if the prior year’s experience was benign.
Accumulation management
Cat models power aggregation by event: sum insured within a hurricane footprint, fault rupture scenario, or wildfire burn probability swath. Risk teams set caps by zone, monitor roll-ups against tolerances, and stop binding when modeled aggregate PML or TVaR breaches guidelines. Corporate risk managers use the same logic indirectly when insurers aggregate locations on declarations and require updated schedules after acquisitions or new construction.
Corporate and captive perspectives
Organizations retaining risk through deductibles, captives, or parametric triggers still live in a market priced with vendor cat analytics. Modeled tail loss informs retention sizing, collateral discussions, and whether to buy incremental limit above primary. The same renewal discipline applies: know what the model assumes about values and locations.
Frequently asked questions
What is the practical difference between RMS and AIR models?
Both vendors produce stochastic event catalogs, vulnerability functions, and financial roll-ups, but they differ in peril modules, historical data blends, and damage assumptions. Many insurers treat one as primary and the other as a benchmark; large divergences at a chosen return period usually trace to hazard footprints or vulnerability curves, not to a single “correct” number.
How should I interpret AAL versus PML?
AAL is the average modeled loss per year and supports expected-loss pricing. PML is a high quantile at a stated return period on an occurrence or aggregate basis and describes tail risk. Use them together: AAL for central tendency, PML and EP curves for reinsurance and capacity limits.
When do I use an occurrence EP curve instead of an aggregate EP curve?
Use the occurrence EP curve when a single event drives recovery—typical for per-event excess-of-loss reinsurance. Use the aggregate EP curve when multiple events in one treaty year stack toward an aggregate limit or stop-loss. Mixing bases without relabeling is a common source of mis-sized towers.
What does model validation and calibration actually check?
Validation compares modeled loss distributions and event rates to credible external references and company experience where data allows. Calibration adjusts model components so agreed metrics—often AAL and selected return periods—fall within tolerance. It does not guarantee the next event will match the catalog; it keeps the tool honest against history and stated assumptions.
What changed in recent vendor updates around climate and secondary perils?
Release cycles have added or expanded climate-conditioned hazard views for selected perils and regions, plus dedicated severe convective storm and wildfire modules with their own event generators. Users should note which view they ran—baseline versus forward-looking—and expect secondary peril contributions to rank higher on inland and wildland–urban interface books.
How do cat models feed reinsurance buying decisions?
Modeled occurrence losses size event excess layers; aggregate losses size stop-loss and aggregate covers. Net runs after treaty terms show remaining tail retention. Brokers and cedents anchor attachments and limits to EP points and TVaR where internal policy requires, then negotiate rate and capacity against that modeled net profile.