Published October 2026.
Direct Answer: Model blending is the disciplined practice of combining outputs from two or more catastrophe models—or the same model under explicit alternative assumptions—into a single view of loss that no single vendor run can justify on its own. Professionals weight models by peril and portfolio fit, stress vulnerability curves, event rates, and financial terms, and document every choice so underwriters, reinsurers, and auditors can see why the renewal number is what it is—not a black-box average.
Two public facts, not a new blending method. Florida’s 2023 law lets property insurers average two or more commission-approved hurricane models; before that, a filing had to pick one accepted model (Insurance Journal, June 12, 2025). Separately, Aon’s 2025 survey of its own clients and prospects — a survey, not a census, and not a license-dollar market share — found that about half of respondents license no vendor and lean on a broker. Moody’s RMS and Verisk are the two predominant licenses. Verisk is wider among U.S. respondents. Moody’s is preferred in UK/EMEA. Reinsurance placements in that survey are built on one of those two or a blend of them, because that is what the reinsurer will re-run. No percentages are printed here: the document does not state how many people answered.
When your catastrophe report shows one expected loss, one probable maximum loss, or one tail metric, the practical question is rarely which model is right. It is whether any one model, frozen at one version and one set of assumptions, should carry the full weight of a placement, a retention decision, or a board narrative. Model blending and sensitivity testing make combination and challenge explicit: how blends differ from naive averaging, how weights vary by peril, what to stress, and how to record the work so the dashboard number has a defensible story. For how cat modeling fits the wider renewal stack, see the catastrophe modeling complete guide for 2026.
Why One Modeled Number Is Never Enough
Portfolios sit at the intersection of exposure quality, peril mix, and contract language. No single module excels on every peril and region. Blending does not fix bad exposure; it acknowledges that each model is a structured opinion on physics, statistics, and damage—not a tape measure.
Risk managers and CFOs often receive one modeled AAL or 1-in-250 year loss on a slide. Blending reconciles parallel model worlds before the market does opaquely; sensitivity testing shows which assumptions move pricing, attachment, or capital. That pairing is the discipline behind the dashboard.
What Model Blending Means in Practice
Model blending assigns weights to two or more modeled loss distributions—or key metrics—and produces a composite view across vendors, versions, or approved overlays. Outputs include blended AAL, merged exceedance curves, or fixed return-period PML. Weights must be reproducible and tied to peril rationale.
Reinsurance panels often expect proof you did not blindly adopt the lowest or highest vendor quote. This workflow is not EP-curve reading or vendor shootouts; it is the arithmetic and governance of combination itself.
Metrics Versus Full Curves
Metric blending forms a weighted sum of AAL or 250-year PML from two runs. Curve blending merges exceedance probabilities at each loss level, then reads off metrics. Metric blending is simpler to audit; curve blending preserves tail shape when weights are stable across return periods. If weights change by return period—more weight to Model A below retention and Model B above—you are curve blending, not averaging two PMLs at one point.
Blending Versus Simple Averaging
Simple averaging—50/50 between two AALs, or the mean of three vendor PMLs—looks neutral and rarely is. Equal weights imply equal credibility across perils, regions, and model generations, which almost never matches reality. Averaging also ignores correlation: two models built on similar catalogs and vulnerability libraries do not provide independent information; their mean can overstate certainty.
Weighted blending starts from explicit hypotheses—higher weight where catalog treatment, extratropical behavior, or occupancy mapping fits your book. A judgment layer is allowed only if defined, not smuggled in as rounding.
Worked Example: Weighted AAL Versus a Straight Mean
Take a $500M total-insured-value schedule with two vendor AALs for the same peril and territory: Vendor 1 reports $4.2M; Vendor 2 reports $6.8M. A straight average is ($4.2M + $6.8M) / 2 = $5.5M. Suppose your peril committee assigns 65% weight to Vendor 1 and 35% to Vendor 2 after a side-by-side on exposure mapping and flood sublimit handling. The blended AAL is (0.65 × $4.2M) + (0.35 × $6.8M) = $2.73M + $2.38M = $5.11M. The $390K gap versus the simple mean is documented relative trust, not noise. If sensitivity testing shows Vendor 2’s AAL moves more than $1M when wind deductibles follow manuscript wording exactly, you may shift weights at renewal.
Weighting Models by Peril
Peril-specific weights are the norm in multi-peril programs. Hurricane, earthquake, severe convective storm, wildfire, and flood draw on different modules and validation paths. A weight set for Florida wind is inappropriate for California earthquake without re-argument. Many organizations maintain a weight matrix: rows are perils or peril-region buckets, columns are models or scenarios, cells sum to 100% within each row.
Weights connect to historical fit, independence of catalog and vulnerability logic, update stability, and reinsurer pricing alignment. Document carve-outs when licensing limits leave part of the book single-sourced.
Regional Splits Within a Peril
Large portfolios sub-weight within hurricane: higher weight where storm surge and BI waiting periods match coastal wording, lower weight inland where models agree within a few percent. Earthquake blends split strike-slip versus subduction zones. Wildfire blends may down-weight a vendor where wildland-urban interface coding was mass-updated mid-term. Collapsing that heterogeneity into one national weight without narrative is what auditors push back on. Align weights with where models structurally disagree using hurricane, earthquake, and wildfire catastrophe peril analysis as peril context, not as a substitute for your matrix.
Sensitivity Testing: What to Stress
Sensitivity testing varies defensible inputs and records how key metrics move. Blending sets the central view; sensitivity defines error bars and escalation triggers. Standard domains are vulnerability, event rates, and financial terms—the levers underwriters manipulate when comparing options.
Vulnerability, Event Rates, and Financial Terms
Stress vulnerability curves, catalog frequency, and financial terms together. Upper-bound vulnerability for a dominant occupancy can move PML more than a small weight shift between vendors. Rate tests use approved scalers or discrete activity views to test stability, not to pick a climate storyline. Terms tests rerun the same events with manuscript-faithful deductibles, coinsurance, BI waiting periods, and sublimits—where gross AAL agrees but net PML diverges, terms—not physics—often drive the gap.
Worked Example: Sensitivity on a Single PML
Suppose blended 250-year hurricane PML is $48M on a $500M TIV program with a $5M per-occurrence deductible. Vulnerability high case: +18% on wind-driven damage for key occupancies yields PML about $56.6M. Event rate +15% on the catalog yields PML about $55.2M. A diagnostic run that applies deductibles per building instead of per occurrence might show $41M—terms sensitivity larger than vendor spread. For an excess layer attaching at $50M, vulnerability and rate move the metric across the attachment; deductible mechanics determine whether the layer responds in the model.
Governance, Documentation, and External Review
Blending without workflow becomes spreadsheet folklore. A minimal path: lock exposure snapshot and model versions; run baselines; agree peril-level weights with recorded sign-off from risk, underwriting, and analytics; compute blended metrics; execute sensitivity tests; publish a blend memo tied to renewal. Budget sometimes limits vendor runs—document single-sourced perils honestly; see catastrophe modeling costs and who pays for fee allocation, not weight logic.
Underwriters need weights by peril, central metrics, and sensitivity on attachment or named storm sublimits; flag any gap versus bound deductibles. Tie metrics to treaty form—aggregates versus cat excess—using reinsurance treaty structures: quota share, excess of loss, and aggregate. Memos should list snapshot date, versions, weight table, merge rule, blended AAL and tail metrics, sensitivity deltas, approvers, and internal versus submission use.
Auditors and reinsurers want repeatable weight rationale, synchronized terms, and change logs. Sensitivity should bracket the base case. Align narrative with catastrophe portfolio management, accumulation, and PML for reinsurance, and note whether vendor updates refreshed the blend or weights stayed fixed to isolate exposure growth—see climate risk pricing and catastrophe model updates in 2026.
Pitfalls and Platform Context
Do not treat correlated models as independent, blend gross with net, or let weights drift after sign-off. Side-by-side diagnostics inform weights without becoming a vendor contest; platform background sits in how RMS, AIR, and Verisk approaches quantify catastrophe risk, while your memo cites weights and outcomes only.
Before Renewal Sign-Off
Confirm peril-level credibility, sensitivity on vulnerability, event rates, and financial terms, and reconstructability of the central metric from the memo. Then the dashboard figure reflects governed blending—not a single oracle run.
FAQ
What does model blending mean in catastrophe risk?
Model blending is the explicit combination of loss results from two or more catastrophe model runs into one coherent view, using defined weights or merge rules rather than treating a single run as definitive. The inputs can be different vendors, different versions, or the same engine under approved alternative assumptions. Outputs may include blended average annual loss, exceedance curves, or return-period metrics such as PML. The purpose is to reflect peril-specific credibility and portfolio fit while keeping the method transparent for renewal and audit.
How is blending different from simple averaging?
Simple averaging assigns equal importance to every input, which assumes equal credibility across perils, regions, and model designs that rarely match reality. Weighted blending applies peril-specific or region-specific weights tied to documented criteria such as historical fit, exposure handling, or treaty relevance. Averaging also treats highly correlated models as if they were independent, which can narrow apparent uncertainty falsely. Blending makes those judgments visible instead of hiding them inside a neutral-looking mean.
How should models be weighted by peril?
Build a matrix where each peril or peril-region row has weights summing to 100% across models or scenarios used for that row. Tie weights to explicit factors: quality of exposure mapping for that peril, stability across recent updates, independence of catalog and vulnerability assumptions, and alignment with reinsurer pricing for the same peril. Avoid one national weight if coastal wind, inland storm, or seismic zones behave differently in your book. Document carve-outs where only one model ran because of data or licensing limits.
What does sensitivity testing cover?
Sensitivity testing varies defensible inputs and records how key metrics move. Standard domains are vulnerability or damage functions, event rates or catalog frequency treatments, and financial terms including deductibles, coinsurance, BI waiting periods, and sublimits. Teams usually report percent changes to AAL and to tail metrics such as 100-year or 250-year loss against a documented base blend. The goal is to show which assumptions can change placement decisions, not to replace the central blend with the worst case.
How should a model blend be documented?
Publish a blend memo with exposure snapshot date, model names and versions, peril-level weight tables, merge formulas, blended central metrics, and a short sensitivity summary with named approvers. State which metrics are for internal planning versus carrier or reinsurer submission and confirm financial terms match bound or quoted manuscripts. When weights change from prior renewals, log the reason. Store files with version control so mid-term edits do not overwrite the record shared at renewal.
What do auditors and reinsurers expect to see in a blend?
They expect repeatability: weights justified by criteria, synchronized policy terms, and a clear trail when assumptions change year over year. They look for sensitivity results that bracket the central estimate rather than a single unchallenged base case. Reinsurers compare your metrics to their own models; documentation explains intentional differences without appearing arbitrary. They also expect honesty about single-model perils, known data gaps, and whether major vendor updates were incorporated or deliberately held fixed to isolate exposure growth.