Reverse Stress Tests for Catastrophe Portfolios

Introduction

A forward stress test asks what happens to a portfolio under a specified catastrophe scenario. A reverse stress test starts with the outcome management cannot accept, such as exhausting a loss capacity or falling below an earnings floor, and works backward to find the catastrophe conditions that would produce it. The result is not a forecast of which event will occur. It is a structured description of how a breach could happen and which assumptions or exposures matter.

For catastrophe portfolios, the reverse calculation can reveal whether a breach requires one extreme event, a combination of perils, or an event combined with adverse exposure or model assumptions. It is useful only when the threshold, loss measure, scenario assumptions, and limits on feasible event conditions are stated clearly. The methods below are aligned with the CAS Exam 9 (Risk Management for Actuaries) syllabus.

Definition: A reverse stress test specifies a material adverse outcome first, then identifies the scenario conditions under which that outcome is reached or exceeded. A catastrophe reverse stress test defines a portfolio loss model $X(\boldsymbol{s},\boldsymbol{\theta})$, a threshold $K$, and the set of feasible event conditions $\boldsymbol{s}$ and assumptions $\boldsymbol{\theta}$ for which $X(\boldsymbol{s},\boldsymbol{\theta}) \ge K$. The event conditions may describe one peril or several perils; $\boldsymbol{\theta}$ may represent uncertainty in exposure, vulnerability, or model parameters.

Set the Threshold Before Solving

The threshold must correspond to a decision-relevant outcome. A loss threshold might be the largest catastrophe loss the organization can absorb. An earnings threshold can be converted into an allowable catastrophe loss by subtracting the minimum acceptable earnings from earnings before catastrophe loss. A capital threshold can be framed as the loss at which available capital would be exhausted or fall below a defined floor. Keep the units and scope consistent: a gross loss should not be compared with a threshold defined for net loss.

A simple scenario model may approximate portfolio loss as a baseline amount plus contributions from separate event drivers. The coefficients translate scenario indices into dollars; they are not universal physical laws. A reverse result is meaningful only over the range where the model is a reasonable representation of the portfolio. If the model is nonlinear, has event interactions, or has step changes, use that model directly rather than extending a linear approximation beyond its valid range.

Define the breach.
State the loss, earnings, or capital measure and whether equality with the threshold counts as a breach.

Specify the scenario model.
Identify baseline losses, event drivers, exposure assumptions, and the dollar units of every term.

Solve the boundary.
Find the combinations that reach the threshold, then check which combinations are feasible within the stated scenario ranges.

Evaluate responses.
Recalculate the boundary or losses after a proposed exposure control, operational change, or capital adjustment, and test whether the response remains effective under relevant uncertainty.

Why the Boundary Equation Works

Suppose, for a specified scenario model, the portfolio loss is $X=B+a s+b t$, where $B$ is a baseline loss, $s$ and $t$ are event indices, and $a$ and $b$ are dollar losses per index unit. If a breach occurs when $X\ge K$, substituting the model into the breach condition gives

B+a s+b t \ge K.

Subtracting the baseline loss from both sides gives $a s+b t\ge K-B$. This shows why the event-driven loss must use the remaining headroom, $K-B$, rather than the full threshold $K$. For a fixed value of $s$, the boundary value of $t$ is $(K-B-a s)/b$, provided $b>0$. Values above that boundary breach the threshold, while values below it do not. This interpretation applies only within the model’s feasible event ranges.

Worked Example: Finding Combined Wind and Wildfire Conditions

A hypothetical property portfolio has a catastrophe loss capacity of $90 million. Its modeled baseline catastrophe loss is $15 million. For a local scenario approximation, wind contributes $5 million per wind index unit $x$, and wildfire contributes $3 million per wildfire index unit $y$. The model is $X=15+5x+3y$, in millions of dollars. The feasible ranges are $0\le x\le10$ and $0\le y\le12$. Find the combinations that exhaust the $90 million capacity, and determine whether a scenario with $x=8$ can breach it.

Method. Set the loss equal to the capacity to find the boundary. A loss at or above $90 million exhausts capacity, so the breach condition is $15+5x+3y\ge90$. The event-driven loss must therefore be at least $90-15=75$ million:

5x+3y\ge75.

For $x=8$, the wind contribution is $5(8)=40$ million. The wildfire contribution must be at least $75-40=35$ million, so $3y\ge35$, or $y\ge35/3=11.6667$. This is feasible because the stated maximum is $y=12$. At the boundary, the total loss is $15+40+35=\$90$ million.

For comparison, at the maximum wildfire index $y=12$, wildfire contributes $3(12)=36$ million. Wind must then contribute at least $75-36=39$ million, giving $x\ge39/5=7.8$. Thus the feasible boundary includes points such as $(x,y)=(8,11.6667)$ and $(7.8,12)$. The maximum allowed indices produce $15+50+36=\$101$ million, confirming that a breach is possible in the stated model.

Common mistake avoided. Using the full $90 million threshold as the required event-driven loss would give $5x+3y\ge90$, incorrectly omitting the $15 million baseline. With $x=8$, that error would require $y\ge16.6667$, which appears infeasible. The correct remaining headroom is $75 million, and the correct boundary is feasible.

Reverse Tests for Earnings and Uncertainty

An earnings threshold can be converted into a loss threshold before analyzing event conditions. If earnings before catastrophe loss are $E_0$, catastrophe loss is $X$, and the minimum acceptable earnings are $E_{\min}$, then post-catastrophe earnings are $E_0-X$. Falling below the floor occurs when $E_0-X<E_{\min}$, or equivalently $X>E_0-E_{\min}$. If equality with the floor is itself unacceptable, use $X\ge E_0-E_{\min}$. State this convention explicitly.

Uncertainty can be represented by a multiplier or by changing model parameters, but it should not be confused with an event index. A multiplier $q$, for example, may represent a specified adverse adjustment to modeled losses. It does not automatically represent the probability of that event or a confidence level. The assumed range of $q$ must be stated, and the resulting combination must be checked against that range.

Worked Example: Solving for an Earnings-Floor Breach Under Uncertainty

A hypothetical insurer has $32 million of earnings before catastrophe loss and requires at least $5 million of earnings after the event. A wind-zone exposure of $400 million has a modeled damage fraction of $0.04s$, and a second-zone exposure of $250 million has a modeled damage fraction of $0.06t$. Thus, the modeled loss before uncertainty adjustment is $16s+15t$ million. A common uncertainty multiplier $q$ scales both losses, with $1.00\le q\le1.10$. Find the breach boundary at the maximum multiplier. Then test $s=0.6,\ t=0.8$.

Method. The maximum catastrophe loss consistent with the earnings floor is $32-5=27$ million. The post-event earnings are $32-q(16s+15t)$, so the earnings floor is breached when the loss exceeds $27 million. At $q=1.10$, the boundary for reaching the floor is

1.10(16s+15t)=27, \qquad 16s+15t=\frac{27}{1.10}=24.5455.

For $s=0.6$, the first zone contributes $16(0.6)=9.6$ million before the multiplier. The second-zone contribution required to reach the boundary is $24.5455-9.6=14.9455$ million. Therefore $15t=14.9455$, so $t=0.9964$, to four decimal places. At this boundary the adjusted loss is $1.10(24.5455)=\$27.00$ million, leaving exactly $5.00 million in earnings. A strict breach requires a value just beyond the boundary.

For the specified test point, the unadjusted loss is $16(0.6)+15(0.8)=9.6+12=21.6$ million. The adjusted loss is $1.10(21.6)=\$23.76$ million, and post-event earnings are $32-23.76=\$8.24$ million. This point does not breach the $5 million floor. The check is consistent with the boundary result: at $s=0.6$, $t=0.8$ is below the boundary value $0.9964$.

Common mistake avoided. Applying the uncertainty multiplier to just one zone would not match the stated model. Here $q$ scales the combined modeled loss, so the correct expression is $q(16s+15t)$, not $16s+q(15t)$.

Testing Whether a Response Changes the Result

A reverse test should lead to a decision question, not just a boundary. A proposed response may reduce exposure in one region, alter the vulnerability represented by a loss coefficient, or increase the capital available to absorb losses. These actions change different parts of the problem: exposure or vulnerability changes the modeled loss, while additional capital changes the threshold. Report both the residual loss and the revised threshold where applicable.

A response is effective for a specified scenario if the resulting loss is below the applicable threshold. It is not necessarily effective for all scenarios. The event conditions used for comparison should be fixed when isolating a response’s effect; if the event assumptions also change, distinguish the two effects. In particular, a response that avoids a breach in one tested scenario may still leave a breach under a more severe or differently combined event.

Worked Example: Comparing Exposure Controls and Added Capital

A hypothetical portfolio breaches its loss capacity when catastrophe loss exceeds $62 million. For the scenario under review, loss is $X=12+6x+4y$ million, where $x=5$ and $y=6$. Management considers three separate actions: reduce the coefficient on $x$ from 6 to 4.5 at a cost of $3 million; reduce the coefficient on $y$ from 4 to 3 at a cost of $2 million; or add $6 million of loss capacity. Compare the resulting losses or capacity and determine which actions avoid a breach in this scenario.

Baseline. The modeled loss is $12+6(5)+4(6)=12+30+24=\$66$ million. Since $66 million exceeds the $62 million capacity, the scenario breaches the threshold by $4 million.

Action 1: reduce the $x$-coefficient. The revised loss is $12+4.5(5)+4(6)=12+22.5+24=\$58.5$ million. This is $3.5 million below the original capacity, so the action avoids a breach in this scenario.

Action 2: reduce the $y$-coefficient. The revised loss is $12+6(5)+3(6)=12+30+18=\$60$ million. This is $2 million below the original capacity, so this action also avoids a breach.

Action 3: add loss capacity. The loss remains $66 million, but the revised capacity is $62+6=\$68$ million. The scenario is $2 million below the revised capacity and does not breach it.

The three actions all avoid a breach for the specified event conditions, but they work through different mechanisms and have different costs. The cost figures alone do not establish which action is preferable; the decision also depends on how reliably each action changes the modeled exposure or available capacity and how it performs under other scenarios.

Common mistake avoided. A response should not be credited with reducing modeled loss when it only increases capacity. Under Action 3, the loss remains $66 million; it is the threshold that changes to $68 million.

Interpretation and Common Traps

A reverse stress test identifies a set of conditions, not necessarily a unique catastrophe. If several combinations lie on the breach boundary, each is a different route to the same adverse outcome. Whether a combination is credible requires a separate assessment of the scenario ranges, event dependence, and model limitations. Do not infer event likelihood from the fact that the algebraic solution is feasible.

Omitting the baseline. Subtract any loss already included in the model from the threshold before solving for additional event loss.
Ignoring feasibility. A mathematical solution outside the stated event range does not describe a feasible scenario under that model.
Mixing measures or units. Keep gross versus net scope consistent, and convert all losses and thresholds to the same currency units.
Misstating equality. Decide whether reaching the threshold counts as a breach. Use $X\ge K$ for a breach at equality and $X>K$ for a breach only above the threshold.
Over-interpreting an approximation. A linear scenario model may be useful locally; it should not be treated as valid outside the range for which it was specified.
Confusing severity with uncertainty. An event index describes the scenario; an uncertainty adjustment describes a change to the modeled loss. Keep their roles distinct.

Key takeaway: Define the adverse outcome first, translate it into a consistent loss threshold, solve for feasible event combinations, and then recalculate the loss or threshold under each proposed response. The result is a decision-focused description of breach conditions, not a prediction of catastrophe frequency.

Check Your Understanding

For each problem, use only the stated scenario model and treat reaching the threshold as the boundary.

A portfolio has a $75 million loss threshold, a $9 million baseline loss, and event loss $4a+5b$ million. If $a=6$, find the boundary value of $b$. (Answer: $b=8.4$)
Earnings before catastrophe loss are $18 million, and the minimum acceptable earnings are $2 million. Loss is $8u+6v$ million. At $v=1.5$, find the boundary value of $u$. (Answer: $u=0.875$)
A scenario’s unadjusted loss is $20 million and its uncertainty multiplier can be as high as 1.08. The loss threshold is $22 million. Does the maximum multiplier cause a breach, and what multiplier would reach the threshold? (Answer: No; maximum adjusted loss is $21.6 million, and the boundary multiplier is 1.10.)
Loss is $10+5r+3s$ million, with a breach only when loss exceeds $40 million. At $r=4,\ s=5$, compare the original loss with the loss after reducing the coefficient on $3s$ from 3 to 2. (Answer: Original loss is $45 million and breaches; revised loss is $40 million and does not breach.)
A model has event loss $2a+3b$ million and a $30 million threshold, with $0\le a\le8$ and $0\le b\le8$. For $a=5$, find the minimum boundary value of $b$, and state whether the boundary is feasible. If instead $b=4$, can any feasible $a$ reach the threshold? (Answer: $b=20/3$, approximately 6.6667, which is feasible; with $b=4$, $a$ would need to be 9, so no feasible $a$ reaches the threshold.)

What You'll Learn