Tail Risk Measures from Exceedance Curves

Reading Tail Measures from an Exceedance Curve

A catastrophe exceedance curve summarizes how often annual losses exceed different thresholds. It can therefore answer two related questions: what loss corresponds to a specified annual exceedance probability, and what is the average loss in the upper portion of the annual loss distribution? These quantities help describe tail severity beyond the average annual loss.

First identify which loss definition the curve represents. An occurrence exceedance probability (OEP) curve concerns the maximum event loss in a year; an aggregate exceedance probability (AEP) curve concerns the total loss in a year. The same calculations apply to either curve, but the resulting loss measure and its interpretation differ. Use the curve that matches the decision being evaluated.

Definition: Let $L$ be a nonnegative random variable representing the annual loss on the selected basis, and define its annual exceedance curve by $q(x)=P(L>x)$. For confidence level $p$, the value at risk is $v_p=\inf\{x:F_L(x)\ge p\}$, where $F_L(x)=P(L\le x)$. The return period associated with annual exceedance probability $q$ is $T=1/q$ years. Tail value at risk at confidence level $p$, also called expected shortfall, is $TVaR_p=\frac{1}{1-p}\int_p^1 v_r\,dr$, where $v_r$ is the loss quantile at level $r$.

A return period is the reciprocal of an annual probability, not a promise that an event occurs on a regular schedule. If annual losses are stationary and independent from year to year, the probability of at least one exceedance of a threshold in $n$ years is $1-(1-q)^n$. The assumptions matter: changing risk over time or dependence across years means this simple multi-year calculation need not apply.

When a curve is tabulated, finding a return-period loss usually requires inversion. If the required probability lies between two tabulated points, an interpolation rule is needed. Linear interpolation in exceedance probability versus loss is a practical approximation, not a universal property of catastrophe losses. For a TVaR calculation, the curve must also describe the tail far enough to integrate the exceedance probabilities; a truncated curve requires an explicit tail assumption.

From VaR to Tail Value at Risk

VaR identifies a threshold, while TVaR summarizes losses in the upper tail. For a continuous distribution, there is no probability mass exactly at $v_p$, so $P(L>v_p)=1-p$, and TVaR can be read as the conditional mean $E[L\mid L>v_p]$. For a discrete distribution, the probability above $v_p$ may be less than $1-p$. In that case, TVaR includes only the fraction of probability mass at $v_p$ needed to fill the upper $1-p$ tail.

TVaR_p=v_p+\frac{1}{1-p}\int_{v_p}^{\infty}q(x)\,dx

This exceedance-area formula works for continuous and discrete loss distributions when $v_p$ is defined by the quantile convention above. It also shows why TVaR exceeds or equals VaR: the area is nonnegative. If the curve is piecewise linear, calculate each segment’s area as a trapezoid. If the curve is available only at discrete loss outcomes, use the probability masses or an equivalent exact exceedance-area calculation.

Choose the loss basis.
Confirm whether $L$ is annual aggregate loss or annual maximum event loss, and whether it is gross or net of the relevant terms.

Find the quantile.
For a return period $T$, use target annual exceedance probability $q=1/T$, corresponding to confidence level $p=1-q$. Invert the curve, interpolating only under a stated rule.

Integrate the tail.
For $TVaR_p$, find the area under $q(x)$ from $v_p$ onward, divide by $1-p$, then add $v_p$.

Check interpretation and completeness.
Verify units and tail coverage, and account for any probability mass at the quantile rather than treating a discrete tail as continuous.

Worked Calculations

Worked Example: Interpolating a Return-Period Loss

A hypothetical catastrophe model reports an annual aggregate loss exceedance curve with losses in thousands of dollars. The curve has $q(40{,}000)=0.0200$ and $q(60{,}000)=0.0050$. Estimate the 100-year loss using linear interpolation in exceedance probability versus loss. Also find the probability of at least one exceedance of that estimated loss in 30 years. Assume the curve is stationary, annual losses are independent across years, and there is no discounting or mortality assumption.

Method. A 100-year return period corresponds to annual exceedance probability $q=1/100=0.0100$. This target is between 0.0200 and 0.0050. The fraction of the probability interval from 0.0200 down to 0.0100 is

\frac{0.0200-0.0100}{0.0200-0.0050} =\frac{0.0100}{0.0150} =\frac{2}{3}.

The loss is therefore two-thirds of the way from 40,000 to 60,000:

40{,}000+\frac{2}{3}(60{,}000-40{,}000) =40{,}000+13{,}333.33 =53{,}333.33.

Rounded at the final step to the nearest thousand dollars, the estimated 100-year loss is $53,333 thousand (approximately $53.333 million). Under the stated independence and stationarity assumptions, the 30-year probability of at least one annual exceedance is

1-(1-0.0100)^{30}=1-0.99^{30}=0.2603.

The probability is 0.2603, or 26.03%. A 100-year loss does not mean the loss occurs exactly once every 100 years. Common mistake avoided: using $1/p=1/0.99$ as the return period. The return period uses the exceedance probability $1-p=0.01$, not the confidence level.

Worked Example: Integrating a Piecewise-Linear Tail

A hypothetical insurer’s annual aggregate loss curve is represented, in thousands of dollars, by the points $(20{,}000,0.1000)$, $(30{,}000,0.0600)$, $(40{,}000,0.0300)$, $(50{,}000,0.0100)$, and $(60{,}000,0)$. Assume linear interpolation between successive points, no probability mass at the 90th-percentile loss, and no losses above 60,000 under this specified curve. Calculate $TVaR_{0.90}$. No discounting or mortality assumption applies.

Step 1: Find VaR. At $p=0.90$, the tail probability is $1-p=0.10$. The curve has $q(20{,}000)=0.10$, so $v_{0.90}=20{,}000$.

Step 2: Calculate the tail area. Each area is a trapezoid with width 10,000. In units of thousands of dollars, multiplied by probability, the four areas are

\begin{aligned} A_1&=10{,}000\left(\frac{0.1000+0.0600}{2}\right)=800,\\ A_2&=10{,}000\left(\frac{0.0600+0.0300}{2}\right)=450,\\ A_3&=10{,}000\left(\frac{0.0300+0.0100}{2}\right)=200,\\ A_4&=10{,}000\left(\frac{0.0100+0}{2}\right)=50. \end{aligned}

Thus the total area beyond VaR is $800+450+200+50=1{,}500$ in these units. Divide by the tail probability and add VaR:

TVaR_{0.90}=20{,}000+\frac{1{,}500}{0.10} =20{,}000+15{,}000 =35{,}000.

The result is $35,000 thousand (or $35 million). The area has units of loss times probability, so dividing by a probability returns a loss. Common mistake avoided: dividing the area by 0.90. TVaR averages the upper 10% of the distribution, so the divisor is $1-p=0.10$.

Worked Example: TVaR When Loss Has a Probability Mass at VaR

A hypothetical annual aggregate loss distribution, in thousands of dollars, is $P(L=0)=0.70$, $P(L=10{,}000)=0.15$, $P(L=30{,}000)=0.10$, and $P(L=80{,}000)=0.05$. Calculate $VaR_{0.90}$ and $TVaR_{0.90}$. Treat the stated probabilities as the complete annual loss distribution; there is no discounting or mortality assumption.

Step 1: Locate the quantile. The cumulative probabilities are $F_L(0)=0.70$, $F_L(10{,}000)=0.85$, and $F_L(30{,}000)=0.95$. The first loss at which the cumulative probability reaches 0.90 is 30,000. Therefore $VaR_{0.90}=30{,}000$.

Step 2: Fill the upper 10% tail. The 80,000 outcome has probability 0.05, so it supplies half of the required upper-tail probability. The remaining 0.05 comes from half of the 0.10 probability mass at 30,000. The tail average is

TVaR_{0.90} =\frac{(0.05)(80{,}000)+(0.05)(30{,}000)}{0.10} =\frac{4{,}000+1{,}500}{0.10} =55{,}000.

As a check using the exceedance-area formula, for $30{,}000\le x<80{,}000$, $q(x)=0.05$, and above 80,000 it is zero. The area is $50{,}000(0.05)=2{,}500$, giving $30{,}000+2{,}500/0.10=55{,}000$. Thus $TVaR_{0.90}$ is $55,000 thousand.

Common mistake avoided: averaging only outcomes strictly above VaR gives 80,000, but those outcomes occupy only 5% of probability, not the required upper 10%. The correct tail includes the necessary fraction of the mass at VaR. This example also shows why the strict exceedance probability $P(L>30{,}000)=0.05$ need not equal 0.10 at a discrete quantile.

Why the Tail-Area Formula Holds

For any nonnegative random variable $Z$, its expectation can be expressed as the area under its survival function: $E[Z]=\int_0^\infty P(Z>y)\,dy$. This follows by writing each realized value $Z$ as the length of the interval from zero to $Z$, then taking the expectation of that length. Apply the identity to the positive part of annual loss above VaR, $(L-v_p)^+=\max(L-v_p,0)$:

\begin{aligned} E[(L-v_p)^+] &=\int_0^\infty P((L-v_p)^+>y)\,dy\\ &=\int_0^\infty P(L>v_p+y)\,dy\\ &=\int_{v_p}^\infty P(L>x)\,dx\\ &=\int_{v_p}^\infty q(x)\,dx. \end{aligned}

The upper $1-p$ portion of the loss distribution has average $v_p+E[(L-v_p)^+]/(1-p)$. Substituting the area identity gives the stated formula. For a discrete distribution this is the quantile-integral definition of TVaR, which effectively takes only as much probability at $v_p$ as is needed to form a tail of size $1-p$.

Interpretation and Common Traps

Return-period loss and TVaR answer different questions. A return-period loss is a quantile: it identifies a threshold associated with a specified annual exceedance probability. TVaR is an average over a specified upper tail and therefore reflects the severity of losses beyond a threshold as well as the threshold itself. Two portfolios can have the same return-period loss but different TVaR if their more extreme tails differ.

Mixing annual loss bases: An OEP threshold describes the annual maximum event; an AEP threshold describes annual total loss. Do not label one as the other.
Reading a return period as a schedule: A 100-year return period means an annual exceedance probability of 0.01 under the stated model, not one occurrence every century.
Using the wrong tail probability: At confidence level $p$, the upper-tail probability is $1-p$. That is the divisor in the TVaR formula.
Ignoring the far tail: A curve ending at a positive exceedance probability does not establish that the remaining tail area is zero. The omitted tail needs a justified extension or an explicit limitation on the result.
Overlooking interpolation and atoms: A tabulated curve may require an approximation between points, while a discrete loss distribution can have a jump at VaR. State the method and handle probability mass accordingly.

Key takeaway: Convert a return period $T$ to annual exceedance probability $1/T$, invert the correctly defined annual loss curve to obtain the quantile, and calculate TVaR from the area beyond that quantile divided by the upper-tail probability. Check loss basis, units, interpolation, and tail completeness before interpreting either measure.

These calculations are aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus. In applied work, tail measures are model-based summaries: their precision depends on the annual loss definition, simulated or fitted tail, and assumptions used to represent losses beyond the observed curve.

Check Your Understanding

Use the stated loss units and interpolation assumptions. Answers are given at the end of each problem.

A hypothetical annual aggregate curve has exceedance probability 0.0400 at $25,000 thousand. What return period does that point represent? (Answer: 25 years.)
A curve is linear in exceedance probability versus loss between ($15,000 thousand, 0.0800) and ($25,000 thousand, 0.0300). Estimate the loss at a 25-year return period. (Answer: $23,000 thousand.)
A piecewise-linear annual loss curve has points ($10,000 thousand, 0.0500), ($20,000 thousand, 0.0300), ($30,000 thousand, 0.0100), and ($40,000 thousand, 0). Find $TVaR_{0.95}$. (Answer: $23,000 thousand.)
An annual loss distribution has probabilities 0.60 at 0, 0.20 at $5,000 thousand, 0.15 at $20,000 thousand, and 0.05 at $50,000 thousand. Calculate $VaR_{0.80}$ and $TVaR_{0.80}$. (Answer: VaR = $5,000 thousand; TVaR = $27,500 thousand.)
Assume independent, stationary years and annual exceedance probability 0.0100 for a selected loss threshold. What is the probability of at least one exceedance in 20 years? (Answer: $1-0.99^{20}=0.1821$, or 18.21%.)

What You'll Learn