Why Diagnose Catastrophe Model Uncertainty?
A catastrophe loss estimate is the result of a chain of inputs and transformations. The locations and characteristics of insured properties determine what is exposed; a hazard model describes the events that may affect those properties; vulnerability functions translate hazard into physical damage; and financial terms translate damage into insured payments. A change in the final estimate can arise at any of these stages. Without diagnosing the source, it is difficult to decide whether the difference calls for better exposure data, a different hazard assumption, revised vulnerability mapping, or a correction to the financial calculation.
A model result is conditional on its data, assumptions, event set, and calculation method. It is not a guarantee of the loss that will occur. This tutorial is aligned with the CAS Exam 5 (Basic Ratemaking and Estimated Claim Liabilities) syllabus.
These categories are useful for diagnosis, but they are not necessarily independent. For example, a building’s construction class is exposure information used to select a vulnerability function. A location error can change the hazard assigned to a property and therefore the vulnerability calculation as well. A financial term may also magnify or dampen the effect of a change in modeled damage.
Random Event Variability and Uncertainty About the Model
A catastrophe model can represent many possible events, each with a frequency and a distribution of losses. The randomness of which events occur, and when they occur, is aleatory variability. It remains even if the exposure data and model assumptions are known exactly. By contrast, uncertainty about the true exposure values, hazard rates, vulnerability functions, or financial interpretation is often called epistemic uncertainty: it reflects incomplete knowledge or alternative defensible assumptions.
A modeled average annual loss (AAL) is an expected loss per year under a stated model and set of inputs. It is not the loss expected in every year. Nor does a single AAL measure all uncertainty: two models may have different AALs, and each model may also have substantial year-to-year variability around its own estimate.
A Simple Expected-Loss Calculation
For a single event type, suppose the expected number of events per year is \(\lambda\), the total exposed value is \(V\), and the expected damage fraction conditional on an event is \(q\). If the damage fraction is measured against that value, the mean loss per event is \(Vq\). The expected annual loss is therefore \(\lambda Vq\).
The reasoning uses linearity of expectation. If the annual loss is the sum of the losses from all events, its expectation is the sum of the expected event losses. With a common mean loss per event, that sum is the expected event count multiplied by the mean loss. Independence among event losses is not needed for this expected-value calculation, although dependence matters for the distribution of annual loss. The simple product is unsuitable if \(\lambda\), \(V\), or \(q\) does not represent the modeled quantity—for example, if an event’s damage fraction varies materially by location and the calculation uses an unweighted portfolio average.
Worked Example: Isolating Exposure and Vulnerability Changes
A hypothetical insurer models two groups of properties. Group A has insured value of $10 million, an expected 0.02 events per year affecting the group, and mean damage of 25% of value per event. Group B has value of $5 million, an expected 0.01 events per year, and mean damage of 40% per event. Calculate the baseline expected annual gross loss. Then calculate the effect of changing only Group A’s exposure value to $12 million, and separately the effect of using revised damage fractions of 30% for A and 35% for B. Treat event rates as unchanged.
Method. Use event frequency multiplied by mean loss per event for each group, then add the group expectations. This calculation avoids attributing a change in the result to the wrong model component.
Baseline Group A: mean loss per event is $10 million × 0.25 = $2.5 million. Expected annual loss is 0.02 × $2.5 million = $50,000.
Baseline Group B: mean loss per event is $5 million × 0.40 = $2 million. Expected annual loss is 0.01 × $2 million = $20,000.
Baseline total: $50,000 + $20,000 = $70,000.
Exposure-only change: keep Group A’s rate and damage fraction unchanged. Its mean event loss becomes $12 million × 0.25 = $3 million, giving 0.02 × $3 million = $60,000. Group B remains at $20,000, so the revised total is $80,000, an increase of $10,000.
Vulnerability-only change: return to the original values and rates. Group A gives $10 million × 0.30 × 0.02 = $60,000. Group B gives $5 million × 0.35 × 0.01 = $17,500. The revised total is $77,500, an increase of $7,500 from baseline.
Check: both isolated changes increase the estimate relative to $70,000, but by different amounts. The exposure-only result changes Group A’s value, while the vulnerability-only result changes both damage fractions; the two calculations should not be confused or combined unless both changes are intended.
Common mistake avoided: applying a new damage fraction to the entire portfolio when only one group’s vulnerability assumption changed. Each group’s value, event rate, and damage fraction must stay matched.
Diagnosing Hazard and Financial-Module Effects
Hazard uncertainty changes the modeled event distribution: for example, the annual frequency of a severe event or the intensity assigned to a location. The financial module acts after gross loss has been estimated. A change in a deductible or policy limit can alter the insured payment even when the hazard and physical damage estimates are unchanged. Separating the two effects requires holding the other inputs fixed.
Worked Example: Changing Hazard Frequency Versus Policy Terms
A hypothetical portfolio has two event categories. A moderate event has an expected annual count of 0.04 and gross loss of $2 million per event. A severe event has an expected annual count of 0.01 and gross loss of $8 million per event. For each occurrence, the policy pays the amount of gross loss above a $500,000 deductible, subject to a $5 million limit on payment above the deductible. Calculate the expected annual insured loss. Then calculate it if the severe-event rate alone rises to 0.015. Finally, return to the original rates and calculate the result under a $1 million deductible and a $4 million limit on payment above the deductible.
Method. First map each gross event loss to its policy payment, then multiply those payments by their expected annual event counts. This keeps hazard changes distinct from financial-term changes.
Under the original terms, the payment for a moderate event is the lesser of $2 million − $500,000 = $1.5 million and the $5 million payment limit. The payment is therefore $1.5 million.
For a severe event, the amount above the deductible is $8 million − $500,000 = $7.5 million. The payment limit reduces this to $5 million.
The original expected annual insured loss is 0.04 × $1.5 million + 0.01 × $5 million = $60,000 + $50,000 = $110,000.
If the severe-event rate rises to 0.015, with all losses and terms unchanged, the result is 0.04 × $1.5 million + 0.015 × $5 million = $60,000 + $75,000 = $135,000. The increase of $25,000 is the effect of the changed severe-event rate under these assumptions.
Now return to the original event rates and change the terms. For a moderate event, the payment is $2 million − $1 million = $1 million, below the $4 million payment limit. For a severe event, the amount above the deductible is $8 million − $1 million = $7 million, which is capped at $4 million. Expected annual insured loss is 0.04 × $1 million + 0.01 × $4 million = $40,000 + $40,000 = $80,000.
Check: the hazard-only change gives $135,000 using the original terms; the terms-only change gives $80,000 using the original event rates. Neither result should be described as the effect of the other component.
Common mistake avoided: applying the deductible after the payment limit, or treating the stated limit as a cap on gross loss. Here the contract definition explicitly makes the limit a cap on payment above the deductible.
Model Choice, Data Limitations, and Loss Indications
A model comparison is informative only when its scope is understood. If two modeled AALs differ, possible explanations include different hazard assumptions, event catalogs, exposure geocoding, vulnerability functions, or financial mappings. Comparing only the final numbers does not reveal which component produced the difference. Where possible, hold exposure, hazard, vulnerability, and terms fixed in turn, changing only one component for a controlled sensitivity. If components interact, also test relevant combinations rather than assuming their effects add.
A model’s scenario weights require particular care. If weights represent probabilities under a specified probability model, a weighted average can be an expected value under that model. If weights are merely management judgments assigned to alternative models, the result is a decision-weighted indication, not automatically a statistical expectation or confidence interval.
Worked Example: Interpreting a Data Correction and Model Ensemble
A hypothetical portfolio contains 100 buildings, each insured for $1 million. A common hazard view is used to estimate expected annual damage fractions. For 60 buildings with verified masonry construction, the fraction is 0.020. For the remaining 40 buildings, construction is missing and the model defaults to a wood classification with fraction 0.035. Calculate the resulting indication. Then calculate it if inspection supports using the masonry fraction for all 40 buildings with missing data. Separately, three alternative model versions estimate AALs of $1.2 million, $1.8 million, and $2.4 million; an analyst assigns them decision weights of 0.2, 0.5, and 0.3. Calculate the weighted indication and explain its meaning.
Method. First calculate the exposure-weighted loss indication for the stated damage fractions. Then make a separate, controlled data correction. Finally, calculate the decision-weighted model average without confusing it with the data correction.
For the 60 verified buildings, total value is 60 × $1 million = $60 million. Their indication is $60 million × 0.020 = $1.2 million.
For the 40 buildings with missing construction, total value is 40 × $1 million = $40 million. Under the default fraction, their indication is $40 million × 0.035 = $1.4 million. The initial total is $1.2 million + $1.4 million = $2.6 million.
If inspection supports the masonry fraction of 0.020 for the 40 buildings, their indication becomes $40 million × 0.020 = $800,000. The corrected total is $1.2 million + $800,000 = $2 million, a reduction of $600,000. This difference isolates the effect of the specified construction-class correction under the common hazard and damage-fraction assumptions.
For the model ensemble, the weighted indication is 0.2 × $1.2 million + 0.5 × $1.8 million + 0.3 × $2.4 million = $240,000 + $900,000 + $720,000 = $1.86 million. This is a weighted average under the stated decision weights. Unless those weights are justified as probabilities, it is not a probability-based expected loss.
Check: the $2.6 million and $2 million figures compare two exposure-data treatments under the stated common assumptions. The $1.86 million figure summarizes three model outputs using separate decision weights. These quantities answer different questions and should not be substituted for one another without an explicit reconciliation.
Common mistake avoided: calling a weighted model average an objective expected loss without explaining what the weights mean. The calculation is arithmetically valid, but its interpretation depends on the basis for the weights.
A Practical Diagnostic Sequence
A disciplined review makes the loss indication more useful by showing which assumptions drive it and how much the estimate responds to plausible alternatives. The sequence below is a diagnostic framework, not a claim that every component can be isolated perfectly.
State whether the comparison concerns gross damage, insured loss, or another clearly defined loss measure, and whether it is an event loss or an annual expectation.
Check completeness, location quality, insured values, occupancy, and construction. Identify defaults or missing fields rather than treating them as confirmed attributes.
Compare event frequencies, footprints, intensities, and the period or event set represented. Keep hazard changes separate from exposure changes where practicable.
Confirm that the damage function is appropriate for the exposure attributes and hazard measure. Test credible alternatives if the mapping is uncertain.
Verify the policy information and the order in which terms are applied. Confirm that modeled payments use the same contractual definitions as the intended loss measure.
Show controlled changes, material interactions, and the basis for any model weights. A single point estimate without its assumptions can conceal important uncertainty.
Common Traps and Interpretation
- Treating every difference as a hazard difference. Exposure data and vulnerability mapping can change the estimated loss even if the event rates are identical.
- Calling one annual expectation a prediction for next year. An expected value summarizes a modeled distribution; actual annual loss can be much higher or lower.
- Adding sensitivities without checking interactions. A location change can alter both assigned hazard and vulnerability. Separate effects measured under one baseline may not sum to the effect of changing both together.
- Ignoring the financial definition. A gross damage estimate and a policy payment are different measures. A deductible or limit can change the latter without changing the physical damage.
- Overinterpreting a data correction. A recalculation using a new construction class demonstrates the effect of that assumption. It does not establish that the new classification is correct unless the evidence supports it.
Check Your Understanding
For each problem, identify the relevant calculation and keep the stated assumptions fixed.
- A hypothetical property group has $8 million of insured value, an expected 0.015 events per year, and a mean damage fraction of 0.30 per event. What is its expected annual gross loss? (Answer: $36,000)
- Two hypothetical groups have values of $20 million and $10 million. Their event rates are 0.01 and 0.02, and their respective mean damage fractions are 0.20 and 0.10. What is their combined expected annual gross loss? (Answer: $60,000)
- A hypothetical event occurs at an expected rate of 0.03 per year and causes $3.2 million gross loss. The policy has a $400,000 deductible and a $2 million limit on payment above the deductible. What is expected annual insured loss for this event type? (Answer: $60,000)
- Three model versions produce hypothetical AALs of $900,000, $1.5 million, and $2.1 million, with stated decision weights 0.25, 0.50, and 0.25. What is the weighted indication, and what should it be called if the weights are not probabilities? (Answer: $1.5 million; a decision-weighted indication, not necessarily a probability-based expected loss.)
- A portfolio has $50 million of reported insured value. A data review supports adding missing properties worth 20% of the reported value. If the modeled expected annual loss per dollar of value remains unchanged, by what percentage does the loss indication increase? (Answer: 20%)