Stress Testing Model and Exposure Uncertainty

Why Stress Catastrophe Results?

A catastrophe model result depends on more than the event scenario itself. It also depends on what is insured, how the hazard is represented, how damage responds to hazard intensity, and what it costs to settle claims after an event. If any of those inputs are uncertain, a single modeled loss can give a false impression of precision.

Stress testing makes that uncertainty decision-relevant. It asks how a result changes under specified, plausible adverse assumptions. A stress is not a forecast, a probability statement, or proof that the stressed result will occur. Its purpose is to reveal whether an assumption could materially change a management choice, such as whether an exposure limit is adequate or whether a portfolio action should be reconsidered. This approach is aligned with the CAS Exam 9 (Risk Management for Actuaries) syllabus.

The earlier material on catastrophe model uncertainty and sensitivity analysis distinguishes changes in assumptions from changes in the model output. Here the emphasis is on constructing coherent stresses and interpreting their implications. The figures below are hypothetical. They use simple factor adjustments for transparency; actual catastrophe models may require rerunning event footprints, vulnerability functions, coverage terms, or financial calculations rather than multiplying an existing result.

Build a Stress From the Loss Mechanism

A useful first step is to identify how each uncertain input enters the calculation. For a simplified portfolio with total insured value \(E\), a fixed event loss ratio \(r\), and no policy-level complications, modeled loss is \(L=Er\). The loss ratio is treated as fixed only for this simplified exposure calculation. If the exposure changes location, construction, occupancy, or coverage, the loss ratio may change too.

For a fixed portfolio and event, hazard determines the intensity delivered to each location, while vulnerability maps intensity and building characteristics to damage. A more severe hazard does not necessarily produce a proportional change in loss: vulnerability curves can be nonlinear, and a changed footprint may affect a different mix of locations. When the model provides a revised event loss under the stressed hazard or vulnerability, that revised result is preferable to applying an unsupported percentage factor.

Demand surge and claims inflation also need separate definitions. Demand surge is event-related scarcity and cost escalation, often concentrated in affected areas. Claims inflation can reflect broader price trends or claim-cost changes over the period being assessed. A scenario may include both, but only if each factor represents a distinct effect and the baseline has not already incorporated it.

Exposure Errors Under Fixed Loss Ratios

When the event, locations, coverage, and loss ratio are held fixed, exposure changes scale the result directly. If reported exposure is \(E\), the assumed loss ratio is \(r\), and exposure is found to be higher by a proportion \(e\), then the stressed exposure is \(E(1+e)\). The stressed loss is \(E(1+e)r\), so the increase from baseline \(Er\) is \(Ere\).

This result follows by substituting the stressed exposure into the assumed loss equation: \(L_{\text{stress}}=E(1+e)r\). Distributing \(r\) gives \(Er+Ere\); because \(Er=L_{\text{base}}\), the incremental loss is \(L_{\text{base}}e\). The derivation relies on a constant loss ratio. It does not establish that every exposure correction changes loss proportionally: a correction to location, occupancy, construction, limits, or deductibles may change the loss ratio or modeled event response.

Combining Distinct Stress Factors

For a transparent first-pass scenario, distinct assumptions can sometimes be represented by multiplicative factors. Let the baseline loss be \(L_0\); let \(f_E\), \(f_H\), \(f_V\), \(f_D\), and \(f_I\) denote stipulated factors for exposure, hazard, vulnerability, demand surge, and claims inflation. If the scenario explicitly assumes each factor applies to the same loss base in sequence, the adjusted result is the product of those factors and the baseline.

The expression follows by applying the adjustments successively: after the exposure adjustment the result is \(L_0f_E\); applying the next adjustment gives \(L_0f_Ef_H\), and continuing yields the product. Multiplication is appropriate only when each factor has been defined as a relative adjustment to the result produced by the preceding steps. If the model supplies a direct stressed loss, or if effects interact in a way not captured by these factors, use the model output or an explicitly specified interaction instead.

Factor order does not change the numerical product, but the order of the underlying model operations may matter when factors are not simple multipliers. For example, changing hazard intensity can move a location to a different part of a nonlinear vulnerability curve. In that case, multiplying a baseline loss by a hazard percentage may not represent the model's response.

Which Uncertainty Matters to the Decision?

A large stress result is not automatically the most useful one. Management should compare stresses on a common baseline and ask whether each changes a decision. Useful comparisons include the dollar change, the percentage change, and the distance to a defined threshold. If separate stresses are run one at a time, their effects can be ranked without assuming they all occur together. A combined scenario is a separate exercise and should have a stated rationale.

A one-at-a-time comparison also helps identify where additional information may be most valuable. For example, a large hazard sensitivity might motivate review of the event assumptions, while a large exposure sensitivity might motivate data validation. The stress ranking alone does not show which uncertainty is most likely, nor does it establish the value or cost of a particular intervention.

Interpretation and Common Traps

Stress results are most useful when the calculation is traceable from assumption to decision. Report the baseline, each changed assumption, what was held fixed, the stressed result, and the decision threshold or action being considered. If an assumption cannot be supported by the model, present it as an illustrative sensitivity rather than as a model-derived estimate.

• Using a percentage without identifying its base. State whether a percentage applies to exposure, the modeled loss, a loss ratio, or a cost component.
• Assuming linearity where the model is nonlinear. Exposure scaling under fixed loss ratios is not equivalent to changing hazard intensity on a vulnerability curve.
• Double counting cost changes. Check whether demand surge or inflation is already embedded in repair costs, trend adjustments, or the baseline model output.
• Adding percentages when adjustments compound. For sequential multiplicative factors, calculate their product; for model-derived stresses, use the model's recalculated result.
• Treating a stress as a probability. A specified adverse scenario says what happens under its assumptions, not how likely those assumptions are.
• Ranking only by percentage. A percentage increase should be interpreted alongside the baseline dollar amount and the decision threshold.