Exam ASTAM · Simulation Precision and Design · Tutorial 667 of 979

Tail Value-at-Risk Estimation and Standard Errors

Learn to estimate TVaR from simulated losses and calculate a standard error that reflects both the excess losses and random tail membership.

Intermediate 12 min read

What You'll Learn

Define upper-tail TVaR using the loss quantile and expected excess above it
Estimate TVaR from simulated losses using an empirical quantile
Derive the asymptotic variance when the quantile is estimated
Calculate a plug-in standard error using all simulated observations
Distinguish estimated-threshold uncertainty from a fixed-threshold tail-mean calculation
Recognize when the standard error approximation is unreliable

Why TVaR Estimation Needs More Than a Tail Average

A simulated value-at-risk identifies a loss threshold at a chosen confidence level. Tail value-at-risk (TVaR) adds information about the losses beyond that threshold. Because both the threshold and the losses above it are estimated from a finite simulation, a useful standard error must reflect more than the spread of the observed tail losses alone.

The central issue is that an estimated quantile determines which simulation outcomes count as tail observations. If the threshold moves, the membership and size of the tail can change. The standard error developed here incorporates that effect for an empirical TVaR estimator.

Definition: Let $X$ be a loss with distribution function $F_X$. For confidence level $\alpha$, define the upper quantile $q_\alpha=\inf\{x:F_X(x)\geq\alpha\}$ and let $p=1-\alpha$. The upper-tail TVaR is $$ \operatorname{TVaR}_\alpha(X)=q_\alpha+\frac{E[(X-q_\alpha)_+]}{p}, $$ where $(x)_+=\max(x,0)$. If $X$ has a continuous distribution at $q_\alpha$, this equals $E[X\mid X>q_\alpha]$. If there is probability mass at the quantile, the displayed definition allocates only the necessary fraction of that mass to the upper tail; it need not equal $E[X\mid X>q_\alpha]$.

In a simulation, let $X_1,\ldots,X_n$ be independent losses from the same distribution. Write $X_{(j)}$ for the $j$th ordered loss, and use the empirical quantile $\hat q_\alpha=X_{(\lceil n\alpha\rceil)}$ (with the index capped at $n$, if needed). A convenient empirical estimator is $$ \widehat{\operatorname{TVaR}}_\alpha =\hat q_\alpha+\frac{1}{np}\sum_{i=1}^{n}(X_i-\hat q_\alpha)_+. $$

The excess term is zero for every observation at or below the estimated threshold and is the amount above the threshold otherwise. The denominator is $np$, not the observed number of exceedances. This form gives the empirical upper-tail average with any needed fractional weight at the boundary. When there are no ties and $np$ is an integer, it equals the average of the largest $np$ simulated losses.

Worked TVaR Estimation

Worked Example: Estimating an Upper-Tail Average

A hypothetical simulation produces 12 losses, measured in thousands of dollars, in increasing order: $1,2,3,4,5,6,7,8,9,11,14,20$. Estimate TVaR at $\alpha=0.75$ using the empirical quantile convention above.

Method. The tail probability is $p=1-0.75=0.25$, so the nominal number of tail observations is $np=12(0.25)=3$. The empirical quantile index is $\lceil12(0.75)\rceil=9$, giving $\hat q_{0.75}=9$.

The three losses strictly above 9 are 11, 14, and 20. Their excesses above the threshold are $2,5,11$, respectively. Thus the sum of positive excesses is $2+5+11=18$, and $$ \widehat{\operatorname{TVaR}}_{0.75} =9+\frac{18}{12(0.25)} =9+\frac{18}{3} =15. $$

The estimate is 15 thousand dollars. As a check, the average of the three largest losses is $(11+14+20)/3=45/3=15$, which agrees. This method avoids the mistake of averaging the excesses alone: 6 is the average excess, not the average loss in the tail.

Why the Standard Error Uses All Simulation Outcomes

Assume the simulated losses are independent and identically distributed, the distribution is continuous with positive density at $q_\alpha$, and the positive excess $(X-q_\alpha)_+$ has finite variance. For fixed $\alpha$, these conditions allow an asymptotic standard error as the number of simulations grows. Let $Z=(X-q_\alpha)_+$. Then the asymptotic variance of the empirical TVaR estimator is $$ \operatorname{Var}\!\left(\widehat{\operatorname{TVaR}}_\alpha\right) \ \mathrel{\approx}\ \frac{\operatorname{Var}(Z)}{np^2}. $$

A plug-in estimate replaces the unknown quantile by $\hat q_\alpha$ and calculates $Z_i=(X_i-\hat q_\alpha)_+$ for every simulated loss. Let $\bar Z=n^{-1}\sum_i Z_i$ and $s_Z^2=(n-1)^{-1}\sum_i(Z_i-\bar Z)^2$. The estimated standard error is $$ \widehat{\operatorname{SE}}\!\left(\widehat{\operatorname{TVaR}}_\alpha\right) =\frac{s_Z}{p\sqrt{n}}. $$

This calculation includes zeros for losses at or below the estimated quantile. Those zeros are important: the sample variance is calculated across all $n$ simulations, not just the observations classified as tail losses. The formula estimates variability in the excess contribution across the full simulation and therefore reflects random tail membership as well as variation in tail excesses.

Write TVaR as a quantile plus an excess.
$\operatorname{TVaR}_\alpha=q_\alpha+E[(X-q_\alpha)_+]/p$, where $p=1-\alpha$.

Track the effect of estimating the quantile.
For a continuous distribution with density $f_X(q_\alpha)>0$, a small perturbation in the distribution changes the quantile. In the excess expectation, changing the threshold contributes $-p$ times that quantile change.

Observe the cancellation.
The quantile change in the first term is canceled by the threshold-change contribution from the excess term. The resulting first-order influence of one observation is $(Z-E[Z])/p$.

Obtain the variance.
The variance of that influence is $\operatorname{Var}(Z)/p^2$. Averaging $n$ independent contributions divides the variance by $n$, giving $\operatorname{Var}(Z)/(np^2)$.

This derivation explains an important point: uncertainty in the estimated quantile has not been ignored. Its first-order effect cancels with the corresponding movement in the excess term for this TVaR functional. The remaining influence still depends on $q_\alpha$, and the empirical standard error uses the estimated threshold in every $Z_i$.

Worked Example: Calculating the Plug-In Standard Error

Suppose 10 hypothetical simulated losses, in thousands of dollars, are $2,3,4,5,6,7,8,9,12,16$. Estimate TVaR at $\alpha=0.80$ and its asymptotic standard error.

Method. Here $p=0.20$, $np=2$, and the quantile index is $\lceil10(0.80)\rceil=8$. Therefore $\hat q_{0.80}=9$. The two losses above 9 have excesses 3 and 7. The estimated TVaR is $$ 9+\frac{3+7}{10(0.20)}=9+\frac{10}{2}=14. $$

For the standard error, the ten estimated excesses are $0,0,0,0,0,0,0,0,3,7$. Their sum is 10, so $\bar Z=10/10=1$. Their sum of squares is $3^2+7^2=58$. The sum of squared deviations from the mean is $$ \sum_{i=1}^{10}(Z_i-\bar Z)^2 =\sum_{i=1}^{10}Z_i^2-10\bar Z^2 =58-10(1)^2=48. $$

Thus $s_Z^2=48/9=16/3$, and $$ \widehat{\operatorname{SE}} =\frac{\sqrt{16/3}}{0.20\sqrt{10}} =\frac{20}{\sqrt{30}} \approx 3.651. $$

The estimated TVaR is 14 thousand dollars and its estimated standard error is 3.651 thousand dollars. The calculation uses all 10 excesses, including eight zeros. Using only the two positive excesses to calculate $s_Z$ would estimate a different quantity and would not implement this standard-error formula.

Random Tail Counts and a Common Exam Trap

Let $K=\sum_{i=1}^n I(X_i>\hat q_\alpha)$ be the number of simulated losses strictly above the estimated quantile. In a sample without ties, $K$ is usually close to $np$, but the cutoff itself comes from the sample. A standard error based only on the sample variance of the $K$ tail losses divided by $K$ treats the cutoff as if it were fixed. It does not generally measure the uncertainty of the empirical TVaR estimator.

For comparison, if a threshold $q$ were fixed in advance and the target were the conditional mean $E[X\mid X>q]$, a sample mean among exceedances would have a different sampling structure. That is not the same procedure as estimating a quantile from the same simulation and then estimating TVaR. Use a standard error matched to the estimator and its threshold-selection method.

Worked Example: Avoiding a Tail-Only Standard Error

A hypothetical simulation gives 20 losses, in thousands of dollars, in increasing order: $1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,20,24,30,40$. Estimate TVaR and its standard error at $\alpha=0.80$, then compare with the tail-only calculation.

The tail probability is $p=0.20$, so $np=4$. The empirical quantile index is $\lceil20(0.80)\rceil=16$, giving $\hat q_{0.80}=16$. The four losses above it are 20, 24, 30, and 40, with excesses 4, 8, 14, and 24. Their sum is 50, so $$ \widehat{\operatorname{TVaR}}_{0.80} =16+\frac{50}{20(0.20)} =16+\frac{50}{4} =28.5. $$

Across all 20 observations, the excesses consist of 16 zeros and $4,8,14,24$. Their sum is 50, their mean is $\bar Z=50/20=2.5$, and their sum of squares is $$ 4^2+8^2+14^2+24^2=16+64+196+576=852. $$ The sum of squared deviations is $852-20(2.5)^2=852-125=727$, so $s_Z^2=727/19$. The standard error is $$ \frac{\sqrt{727/19}}{0.20\sqrt{20}}\approx 6.916. $$

The tail-only sample has mean 28.5. Its squared deviations from that mean sum to $(-8.5)^2+(-4.5)^2+(1.5)^2+(11.5)^2=227$, giving sample variance $227/3$. Dividing this variance by four and taking a square root gives $\sqrt{(227/3)/4}\approx4.350$. That is not the standard error of the estimated-quantile TVaR procedure. The difference illustrates the trap: a tail-only calculation ignores the effect of selecting the cutoff from the same simulation.

Practical Checks and Limitations

Before reporting a TVaR estimate and standard error, verify that the loss orientation is correct: this tutorial uses the upper tail of large losses. Check that $p=1-\alpha$, that the quantile convention is stated, and that the standard error uses the same units as the simulated losses. A standard error is an estimate of Monte Carlo variability, not a guarantee that the estimate is close to the true TVaR in a particular run.

The approximation relies on independent, identically distributed simulations and a sufficiently regular distribution near the quantile. Strong dependence, very few tail observations, a discontinuity at the quantile, or an infinite second moment can make the approximation unreliable. In particular, when $np$ is small, the tail estimate may be dominated by only a handful of simulated losses. The numerical standard error should then be interpreted cautiously rather than as evidence of stable tail estimation.

Key takeaway: For an empirical upper-tail TVaR estimate, calculate the positive excess above the estimated quantile for every simulation, including zeros. Estimate the standard error as $s_Z/(p\sqrt n)$, under the stated regularity assumptions; do not substitute a tail-only sample standard deviation without checking that it matches the estimator.

Check Your Understanding

Use the empirical quantile convention and plug-in standard error from this tutorial unless a question specifies otherwise. Loss amounts are in thousands of dollars where stated.

A simulation gives ordered losses $2,3,4,5,6,8,10,18$. Find the estimated TVaR at $\alpha=0.75$. (Answer: 14)
At $\alpha=0.75$, the eight estimated excesses above the empirical threshold are $0,0,0,0,0,0,2,10$. Calculate the plug-in standard error. (Answer: 4.957)
For $n=10$, $\alpha=0.80$, and estimated excesses consisting of eight zeros, 2, and 6, calculate the plug-in standard error. (Answer: 3.055)
Why is the sample variance of only the losses above an estimated quantile not generally the variance estimate used for empirical TVaR? (Answer: It treats the cutoff as fixed and omits the uncertainty associated with selecting the threshold and tail membership.)
A simulation gives ordered losses $1,2,3,5,6,8,10,13,17,25$. Estimate TVaR at $\alpha=0.70$. (Answer: 18.333)

Back to Tutorials