Insurance-linked securities
Catastrophe modeling
status: draft
Catastrophe modeling
Catastrophe models are the analytical backbone of ILS. Your entire investment thesis rests on model output. Understanding how models work, their limitations, and how to stress-test results is essential for ILS investing.
The three major modeling firms
| Firm | Strengths | Ownership | Market Position |
|---|---|---|---|
| AIR Worldwide | US hurricane, earthquake | Verisk | Largest market share |
| RMS | Multi-peril, international | Moody’s | Strong in Europe, Asia |
| CoreLogic | Flood, convective storm | CoreLogic | Growing presence |
Most sophisticated investors run multiple models and compare or blend results. Relying on a single model exposes you to that model’s biases and blind spots.
Boutique modelers
| Firm | Specialization |
|---|---|
| Karen Clark & Company (KCC) | Independent cat modeling, climate risk |
| Milliman | Mortality and longevity |
| Impact Forecasting (Aon) | Proprietary models for Aon-placed deals |
Boutique modelers offer alternative views that can validate or challenge the big three.
What models actually do
Cat models have three components that simulate thousands of years of potential catastrophic events:
1. Hazard component
Simulates event characteristics:
- Hurricane tracks, intensities, wind fields
- Earthquake fault ruptures, shake intensities
- Flood depths, storm surge heights
Models generate 10,000-100,000 simulated years of events based on historical patterns, physics, and climate science. Each simulated year contains zero, one, or multiple events.
2. Vulnerability component
Estimates damage to structures given event intensity:
- Construction type (wood frame vs. masonry vs. steel)
- Building height and age
- Roof type and anchoring
- Foundation and flood defenses
Vulnerability functions convert wind speed or shake intensity to damage ratios (0-100% of insured value).
3. Financial component
Applies insurance terms to physical damage:
- Deductibles
- Coverage limits
- Sub-limits (windstorm deductibles)
- Reinsurance terms
Output is a distribution of potential losses for the insured portfolio.
Key model outputs
| Metric | Definition | What It Tells You |
|---|---|---|
| Expected loss (EL) | Average annual loss as % of limit | The actuarial cost of the risk |
| Attachment return period | How often losses reach attachment point | A 100-year attachment means 1% annual PoA |
| Exhaustion return period | How often losses reach exhaustion point | Higher = more buffer before total loss |
| OEP (Occurrence Exceedance Probability) | Probability single event exceeds threshold | Per-occurrence risk assessment |
| AEP (Aggregate Exceedance Probability) | Probability annual aggregate exceeds threshold | Full-year risk assessment |
| Standard deviation | Volatility around expected loss | Risk beyond average |
| Tail value at risk (TVaR) | Average loss in worst X% of scenarios | Tail risk exposure |
Reading an exceedance probability curve
The EP curve shows probability of exceeding various loss levels:
| Return Period | Annual Probability | Meaning |
|---|---|---|
| 10-year | 10% | 1 in 10 chance of exceeding this loss |
| 50-year | 2% | 1 in 50 chance |
| 100-year | 1% | 1 in 100 chance |
| 250-year | 0.4% | 1 in 250 chance |
| 500-year | 0.2% | 1 in 500 chance |
Cat bonds typically attach at 50-200 year return periods, meaning 0.5-2% annual attachment probability.
Model uncertainty
Models are educated guesses. AIR and RMS can produce different expected loss estimates by 20-40% on the same risk.
Sources of divergence
| Component | Source of Uncertainty |
|---|---|
| Hazard | Climate change impacts, historical data gaps |
| Vulnerability | Limited post-event damage surveys |
| Exposure | Data quality, geocoding accuracy |
| Financial | Coverage interpretation, demand surge |
Why models differ
Different views on climate change:
- Impact on hurricane frequency/intensity
- Sea level rise and storm surge
- Wildfire season length and intensity
Different damage functions:
- Same wind speed, different estimated damage
- Secondary peril treatment (water intrusion, debris)
Different assumptions:
- Demand surge (construction cost inflation post-event)
- Loss amplification from multiple events
- Business interruption modeling
Model version matters
A cat bond analyzed on AIR 2019 hurricane model vs. AIR 2023 model can show 15-25% different expected loss.
What changes in model updates
| Update Type | Typical Impact |
|---|---|
| New historical events incorporated | +/- 5-15% EL |
| Climate conditioning adjustments | +10-25% EL for hurricane |
| Vulnerability function refinements | +/- 5-10% |
| New secondary peril inclusion | +5-15% |
Post-2017 model updates generally increased modeled losses for Florida hurricane. Always confirm you’re looking at current model versions.
Version verification checklist
- Model vendor and product name
- Version number (e.g., AIR Touchstone 9.0)
- View name (long-term, near-term, climate-conditioned)
- Event catalog vintage
- Date of model run
Running your own analysis
Don’t rely solely on sponsor-provided modeling.
Getting independent model access
Direct license:
- Annual subscription to AIR, RMS, or CoreLogic
- Full flexibility to run scenarios
- $100K+ annual cost for comprehensive access
- Typical for large ILS funds
Service bureau:
- Pay-per-run through modeling firm or broker
- Lower cost, less flexibility
- Suitable for occasional investors
Third-party analysis:
- Hire specialized consultant
- Independent review of sponsor output
- $10-50K per deal
What to request from sponsors
| Item | Purpose |
|---|---|
| Full model output files | Run your own scenarios |
| Exposure data (anonymized) | Verify data quality |
| Multiple model runs | Compare AIR vs. RMS |
| Sensitivity analysis | Understand parameter impacts |
| Model version documentation | Confirm currency |
Sensitivity analysis
Test how results change when assumptions vary:
| Variable | Test Range | Why It Matters |
|---|---|---|
| Attachment point | +/- $5B from modeled | Check layer stability |
| Demand surge | On/off, 10-20% variation | Major driver for large events |
| Secondary perils | Include/exclude storm surge | Often undermodeled |
| Model uncertainty | 10th/50th/90th percentile | Range of reasonable outcomes |
| Near-term vs. long-term | Compare views | Climate impact assessment |
Sensitivity example
| Scenario | Expected Loss |
|---|---|
| Base case (AIR long-term) | 2.5% |
| Climate-conditioned view | 3.0% |
| Demand surge 20% higher | 2.8% |
| Secondary perils expanded | 2.7% |
| Combined stress | 3.5% |
If spread is 550 bps, base case multiple is 2.2x. Under combined stress, multiple drops to 1.6x. Is that adequate compensation for model uncertainty?
Exposure data quality
For indemnity triggers, exposure data drives everything.
Data quality checklist
| Element | Check |
|---|---|
| Completeness | Are all policies included? |
| Geocoding | What % is resolved to lat/long vs. ZIP centroid? |
| Construction type | Coded correctly? |
| Year built | Available and accurate? |
| Coverage amounts | Current replacement cost? |
| Occupancy | Residential vs. commercial correct? |
Common data problems
Unknown construction type: Model assigns average vulnerability, may over or underestimate actual
ZIP centroid geocoding: All risk placed at center of ZIP code, missing micro-geography
Stale replacement costs: Values from original policy, not updated for inflation
Missing secondary characteristics: Roof type, number of stories, basement presence
Blending multiple models
Sophisticated investors don’t rely on a single model output.
Blending approaches
Simple average: (AIR EL + RMS EL) / 2
Weighted average: Weight by historical accuracy for peril type
Conservative blend: Use higher of two models
Stress blend: Use 75th percentile of higher model
When to override models
Consider loading additional conservatism for:
- Perils with limited loss history (wildfire, flood)
- Emerging markets with sparse exposure data
- Climate-sensitive perils where trends accelerate
- First-time sponsors with unproven exposure data
Climate-conditioned views
Models now offer “climate-conditioned” views that adjust for current climate state.
Long-term vs. near-term views
| View Type | Basis | Typical EL Impact |
|---|---|---|
| Long-term average | Historical average climate | Baseline |
| Near-term conditioned | Recent SST, ENSO state | +10-25% for US hurricane |
| Forward-looking | Projected 2030-2050 climate | Varies widely |
Near-term views argue that recent climate state predicts near-future risk better than long-term average. For US hurricane, this typically increases expected loss.
Asking the right questions
When reviewing model output:
- Which climate view was used?
- Is the sponsor showing the more favorable view?
- What’s the delta between long-term and near-term?
- How would EL change under continued warming?
Model limitations
Models are tools, not truth.
What models do poorly
| Area | Limitation |
|---|---|
| Unprecedented events | No historical analog |
| Secondary perils | Often simplified |
| Demand surge | Highly variable, hard to model |
| Climate change | Uncertain future state |
| Correlated losses | Multi-event scenarios |
| Tail behavior | Limited data in extremes |
Healthy skepticism
- Models missed Katrina’s storm surge magnitude
- Models underestimated 2017 wildfire losses
- Models didn’t anticipate freeze-thaw Texas 2021 losses
The historical record is short. The future may not resemble the past. Factor model uncertainty into your required return.
status: draft
For how climate is changing ILS risk assessment, see Climate and emerging perils. For trigger structures that use modeled loss, see Trigger types.