Catastrophe Risk Loads from AAL and Tail Metrics

From Expected Loss to a Catastrophe Risk Load

A catastrophe premium must account for more than the average modeled loss if the insurer is also charging for the financial strain of unusually severe years. The average annual loss (AAL) measures expected annual loss, while tail measures describe losses in the upper part of the annual loss distribution. A risk-load calculation uses the difference between a chosen tail measure and AAL as a capital proxy, then applies an explicitly stated charge rate.

The measure used and the target return period matter. VaR identifies a threshold loss at a selected percentile; TVaR averages losses in the tail beyond that percentile. For the same loss distribution and percentile, TVaR is at least as large as VaR, but the resulting loads can differ substantially. These are mechanical indications under stated assumptions, not universal prices.

Definition: Let $X$ be a nonnegative random variable representing annual aggregate catastrophe loss for a specified portfolio, coverage basis, and period. Its average annual loss is $E[X]$. For percentile $q$, define $\operatorname{VaR}_q(X)=\inf\{x:F_X(x)\ge q\}$. Define $\operatorname{TVaR}_q(X)$ as the average of the worst $1-q$ fraction of outcomes, including only the fraction of probability mass needed at a discrete boundary. Equivalently, $\operatorname{TVaR}_q(X)=\frac{1}{1-q}\int_q^1 \operatorname{VaR}_u(X)\,du$.

For a nonnegative loss, AAL can also be written as the area under its annual exceedance probability (AEP) curve:

\operatorname{AAL}=E[X]=\int_0^\infty P(X>x)\,dx.

This identity follows by expressing each realized loss $X$ as the area under its indicator function: $X=\int_0^\infty \mathbf{1}_{\{X>x\}}\,dx$. Taking expectations gives $E[X]=\int_0^\infty E[\mathbf{1}_{\{X>x\}}]\,dx=\int_0^\infty P(X>x)\,dx$. In a discrete annual loss table with outcomes $x_j$ and probabilities $p_j$, the equivalent calculation is $E[X]=\sum_j p_jx_j$.

Return Periods and Tail Measures

A return-period target $T$ years conventionally corresponds to an annual exceedance probability $1/T$, and therefore to percentile $q=1-1/T$. For example, a 10-year target uses $q=0.90$, and a 20-year target uses $q=0.95$. A return period describes an annual probability, not a schedule: a 10-year loss can occur in consecutive years.

In a continuous distribution, the loss at percentile $q$ is the VaR threshold, and TVaR is the conditional mean above that threshold. Discrete distributions require care. A quantile can have positive probability at the boundary, so the upper $1-q$ tail may include only part of that probability mass. Averaging every outcome at or above VaR without adjusting the boundary generally uses too much probability and gives the wrong TVaR.

Risk-load indication: For a chosen risk measure $R$, such as $\operatorname{VaR}_q(X)$ or $\operatorname{TVaR}_q(X)$, define the indicated capital proxy as $K_R=R-E[X]$. If an annual capital-charge rate $c$ is specified, the indicated risk load is $L_R=cK_R$, and the loss-plus-load indication is $E[X]+L_R$. This formulation assumes $R\ge E[X]$ and that the stated rate applies to the whole excess. It is not a complete premium calculation: expenses, profit provisions, reinsurance costs, and other adjustments are outside this formula.

The charge rate $c$ must be dimensionless and applied consistently. If losses are in dollars, $K_R$ and $L_R$ are in dollars; if losses are in thousands of dollars, those results remain in thousands until converted. The formula does not say that the entire VaR or TVaR is an additional load. The expected loss is already included, so the capital proxy is the tail measure minus AAL.

Worked Calculations

Worked Example: A One-in-Ten Tail With a Boundary Atom

For a hypothetical portfolio, annual aggregate loss $X$, in thousands of dollars, is 0 with probability 0.80, 200 with probability 0.15, and 1,000 with probability 0.05. Calculate AAL, VaR and TVaR at a one-in-ten target. Then calculate both indicated loads using an annual capital-charge rate of 8%.

Method and AAL. Use the discrete expectation formula. The contributions are $0.80(0)=0$, $0.15(200)=30$, and $0.05(1{,}000)=50$, all in thousands. Thus $E[X]=0+30+50=80$ thousand dollars, or $80,000.

VaR. A one-in-ten target has $q=1-1/10=0.90$. The cumulative probabilities are 0.80 at zero and 0.95 at 200 thousand dollars. Therefore $\operatorname{VaR}_{0.90}(X)=200$ thousand dollars.

TVaR. The upper tail has probability 0.10. It contains all 0.05 probability at 1,000 and only 0.05 of the 0.15 probability at 200. Its tail loss total is $0.05(1{,}000)+0.05(200)=50+10=60$ thousand dollars. Divide by 0.10: $\operatorname{TVaR}_{0.90}(X)=60/0.10=600$ thousand dollars.

Loads. The VaR-based capital proxy is $200-80=120$ thousand dollars, so its load is $0.08(120)=9.6$ thousand dollars, or $9,600. The TVaR-based proxy is $600-80=520$ thousand dollars, so its load is $0.08(520)=41.6$ thousand dollars, or $41,600.

Check and trap avoided. The selected TVaR tail has exactly 0.10 probability, and its average is 600 thousand dollars, between 200 and 1,000 as expected. Using only the 1,000 outcome would incorrectly report TVaR as 1,000 thousand dollars; it omits the required 0.05 probability mass at 200.

Worked Example: Comparing VaR and TVaR Loads at Two Targets

A hypothetical annual aggregate loss distribution, in thousands of dollars, is 0 with probability 0.70, 100 with probability 0.20, 400 with probability 0.08, and 1,000 with probability 0.02. Calculate AAL and the VaR- and TVaR-based loads at 10-year and 20-year targets, using a 5% annual capital-charge rate.

Method and AAL. The expected loss is $0.70(0)+0.20(100)+0.08(400)+0.02(1{,}000)=0+20+32+20=72$ thousand dollars.

10-year target. Here $q=0.90$. The cumulative probability reaches 0.90 at 100, so VaR is 100 thousand dollars. The upper 0.10 probability consists of 0.08 at 400 and 0.02 at 1,000. Thus TVaR is $[0.08(400)+0.02(1{,}000)]/0.10=(32+20)/0.10=520$ thousand dollars. The VaR-based load is $0.05(100-72)=1.4$ thousand dollars; the TVaR-based load is $0.05(520-72)=22.4$ thousand dollars.

20-year target. Here $q=0.95$. The cumulative probability at 100 is 0.90, while at 400 it is 0.98, so VaR is 400 thousand dollars. The upper 0.05 probability consists of all 0.02 at 1,000 and 0.03 of the probability at 400. TVaR is $[0.02(1{,}000)+0.03(400)]/0.05=(20+12)/0.05=640$ thousand dollars. The VaR-based load is $0.05(400-72)=16.4$ thousand dollars; the TVaR-based load is $0.05(640-72)=28.4$ thousand dollars.

Check and trap avoided. Both tail averages use exactly the probability specified by their target. Raising the target from 10 to 20 years increases both indications, but VaR jumps from 100 to 400 thousand dollars because the percentile crosses a discrete probability boundary. Treating a return period as the probability of loss being exactly VaR would not produce these calculations.

Worked Example: Turning a Tail Measure Into a Load

A hypothetical catastrophe model reports an AAL of $240,000 and a 1-in-25-year TVaR of $1,040,000 for annual aggregate loss. A stated annual capital-charge rate is 6%. Calculate the capital proxy, indicated risk load, and loss-plus-load indication. Also find the load as a percentage of AAL.

Method. The 1-in-25 target corresponds to $q=1-1/25=0.96$; the supplied TVaR is used at that target. The capital proxy is $K=1{,}040{,}000-240{,}000=\$800{,}000$.

The indicated load is $L=0.06(\$800{,}000)=\$48{,}000$. Adding the expected loss gives $\$240{,}000+\$48{,}000=\$288{,}000$. As a percentage of AAL, the load is $\$48{,}000/\$240{,}000=0.20=20\%$.

Check and trap avoided. The charge applies to the $800,000 excess, not to the entire $1,040,000 tail metric. Charging 6% on TVaR would give $62,400, which is not the stipulated excess-capital method. The 20% ratio to AAL is a result of these inputs, not the capital-charge rate itself.

Interpreting and Comparing Indications

AAL and tail-based loads answer different questions. AAL is the expected annual loss and is insensitive to how losses are distributed around that mean, provided the mean stays fixed. VaR is a percentile threshold and can change abruptly in a discrete or heavily rounded model. TVaR incorporates the severity of outcomes throughout the selected upper tail, so it distinguishes distributions that share the same VaR but have different extreme losses.

For a common distribution, percentile, and charge rate, TVaR is no smaller than VaR, and therefore the TVaR-based excess-capital load is no smaller than the VaR-based load when both excesses are nonnegative. Increasing the percentile generally increases VaR and TVaR, but the size of the change depends on the loss distribution. Consequently, a return-period target alone does not determine a load: the loss model, tail measure, AAL, and charge rate are all needed.

Comparisons are meaningful only when the measures use the same annual loss basis. In particular, an occurrence exceedance measure and an annual aggregate exceedance measure do not describe the same random variable. The calculations here use annual aggregate loss throughout. If the modeled loss definition, policy terms, portfolio, or time basis changes, the resulting AAL and tail metrics may no longer be comparable.

A useful derivation for the excess-capital form is direct. If $R$ is the selected tail measure, then $R=E[X]+(R-E[X])$. The first term is the expected loss funded in the base indication; the second is the excess associated with the chosen tail measure. Charging rate $c$ on that excess gives $L=c(R-E[X])$, hence a loss-plus-load indication of $E[X]+c(R-E[X])$. This decomposition explains why one subtracts AAL before applying the charge and why the result depends on the chosen risk measure.

Key takeaway: State the annual loss basis, target percentile, risk measure, AAL, and capital-charge rate before calculating a catastrophe risk load. For discrete losses, construct the tail average using exactly $1-q$ probability, including a fractional boundary mass where required.

Check Your Understanding

For each problem, use annual aggregate losses and apply the stated capital-charge rate only to the selected risk measure in excess of AAL.

A hypothetical annual loss is 0 with probability 0.60, $50,000 with probability 0.30, and $250,000 with probability 0.10. Find AAL, VaR at a 10-year target, and TVaR at that target. (Answer: AAL = $40,000; VaR = $50,000; TVaR = $250,000.)
A hypothetical annual loss is 0 with probability 0.75, $100,000 with probability 0.20, and $600,000 with probability 0.05. Find AAL and TVaR at a 10-year target, showing how much probability is taken at $100,000. (Answer: AAL = $50,000; the upper 10% uses all 5% at $600,000 and 5% at $100,000; TVaR = $350,000.)
A model gives AAL of $120,000, 10-year VaR of $300,000, and 10-year TVaR of $500,000. At a 6% annual capital-charge rate, calculate both loads and both loss-plus-load indications. (Answer: VaR load = $10,800 and indication = $130,800; TVaR load = $22,800 and indication = $142,800.)
A hypothetical annual loss, in thousands of dollars, is 0 with probability 0.50, 80 with probability 0.30, 300 with probability 0.15, and 900 with probability 0.05. Calculate AAL, VaR and TVaR at a 10-year target, then the TVaR-based load at a 4% charge rate. (Answer: AAL = $114,000; VaR = $300,000; TVaR = $600,000; load = $19,440.)
For a hypothetical portfolio, AAL is $180,000 and 20-year TVaR is $780,000. At a 5% annual capital-charge rate, calculate the load and identify the percentile represented by the target. (Answer: $q=0.95$; load = $30,000.)

What You'll Learn