Monte Carlo Estimation Of TVaR

Why Estimate TVaR by Simulation?

A fitted loss distribution can sometimes be used to calculate TVaR directly. But direct calculation may be difficult when a model has a complicated form or when losses are generated by a multi-step process. Monte Carlo estimation offers another approach: simulate losses from the model, locate the upper tail in the simulated sample, and average the appropriate tail observations.

The important detail is that the tail of a finite simulation need not contain a whole number of observations. For example, if 12 losses are simulated and the target tail probability is 20%, the tail has an effective size of 2.4 observations. Averaging the two largest losses alone would not use the full 20% tail. The empirical quantile definition provides a precise way to include the fractional part.

Definition: Let $X$ be a loss with a finite mean, and let $p$ be a confidence level between 0 and 1. TVaR at level $p$ is the average of the upper quantiles over levels from $p$ to 1: $$ \operatorname{TVaR}_p(X)=\frac{1}{1-p}\int_p^1 F_X^{-1}(u)\,du. $$ For a simulation of $n$ losses, the empirical TVaR estimate is obtained by applying the same quantile-average definition to the empirical distribution, which assigns probability $1/n$ to each simulated loss.

The Empirical Tail-Average Formula

Suppose the simulated losses, arranged in increasing order, are $X_{(1)}\leq X_{(2)}\leq\cdots\leq X_{(n)}$. At confidence level $p$, define the effective number of tail observations as $k=n(1-p)$. Write $k=h+r$, where $h=\lfloor k\rfloor$ and $0\leq r<1$. The empirical TVaR estimate uses the largest $h$ observations in full and a fraction $r$ of the next-largest observation.

\widehat{\operatorname{TVaR}}_p = \frac{ \displaystyle\sum_{j=0}^{h-1}X_{(n-j)}+rX_{(n-h)} }{h+r}, \qquad h=\lfloor n(1-p)\rfloor,\quad r=n(1-p)-h.

When $r=0$, the formula reduces to the average of the largest $k$ losses. When $k$ is not an integer, the fractional weight makes the total tail weight equal to $k$. This formulation uses the empirical quantile-integral definition and is useful even when simulated losses are discrete or tied.

Sort the simulated losses.
Write the observations in increasing order so the upper tail is easy to identify.

Calculate the tail size.
For $n$ simulated losses and confidence level $p$, calculate $k=n(1-p)$.

Locate the threshold and tail observations.
Use the largest whole number of observations in full, then include the required fraction of the next observation.

Divide by the tail size.
Divide the weighted tail total by $k$, not by the number of observations that received a positive weight.

For a sample without ties, the empirical $p$-quantile is $X_{(\lceil np\rceil)}$ under the usual inverse empirical distribution convention. When $k$ is not an integer, the fractional observation in the formula is the observation at that threshold. With ties, several observations can have the threshold value; the weighted quantile-integral calculation still gives the empirical TVaR.

Why the Fractional-Observation Formula Works

The empirical distribution has a quantile function that is constant over intervals of width $1/n$: it equals $X_{(j)}$ on the interval from $(j-1)/n$ to $j/n$, apart from endpoint conventions that do not affect the integral. The interval from $p$ to 1 therefore covers $k=n(1-p)$ such observation-widths. It includes $h$ complete upper intervals and a remaining fraction $r$ of the next interval.

\int_p^1 \widehat{F}^{-1}(u)\,du = \frac{1}{n} \left( \sum_{j=0}^{h-1}X_{(n-j)}+rX_{(n-h)} \right).

Dividing this integral by $1-p=k/n=(h+r)/n$ gives the weighted tail average in the formula. Thus the fractional weight is not an arbitrary adjustment: it is the part of the empirical quantile step that lies above $p$.

Worked Example: A Fractional Tail from an Exponential Model

A hypothetical annual loss model is exponential with mean $25, so $X\sim\operatorname{Exponential}(\text{mean }25)$. Twelve independent losses are simulated, producing the ordered values $10, $12, $14, $15, $17, $19, $20, $23, $26, $31, $38, and $50. Estimate TVaR at $p=0.80$. Use the empirical quantile-integral method and round only the final estimate to two decimal places.

Method and threshold. Here $n=12$, so the effective tail size is $k=12(1-0.80)=2.4$. Thus $h=2$ and $r=0.4$. The empirical threshold is the 10th ordered value because $\lceil np\rceil=\lceil9.6\rceil=10$, which is $31. The two losses above the threshold are $38 and $50; the calculation includes 0.4 of the threshold loss.

Weighted tail calculation. The weighted tail total is $50+38+0.4(31)=50+38+12.4=100.4$. Dividing by the tail size gives $100.4/2.4=41.8333\ldots$. The estimated TVaR is $41.83.

Check and common mistake. The weights total $1+1+0.4=2.4$, as required. Averaging only the two largest losses would give $(50+38)/2=$44.00$; that is the wrong calculation because it omits part of the empirical tail mass.

Using Repeated Runs to Assess Simulation Variability

A single Monte Carlo estimate depends on the particular random losses generated. To assess this run-to-run variability, repeat the simulation independently using the same model, sample size, and confidence level. Calculate TVaR separately for each run. The sample standard deviation of those estimates describes how much a run’s estimate typically varies across repetitions. The standard error of their average estimates how much the average of the repetitions varies.

If $R$ independent runs produce estimates $T_1,\ldots,T_R$, their mean and sample standard deviation are

\bar{T}=\frac{1}{R}\sum_{b=1}^R T_b, \qquad s_T=\sqrt{\frac{\sum_{b=1}^R(T_b-\bar{T})^2}{R-1}}.

The estimated standard error of the average of the $R$ run estimates is $s_T/\sqrt{R}$. This is not the standard deviation of one run: the latter is estimated by $s_T$. Repeated runs assess Monte Carlo variability conditional on the chosen model and parameters. They do not measure error from choosing an inappropriate model or estimating its parameters inaccurately.

Worked Example: Comparing Independent Simulation Runs

A hypothetical annual loss model is exponential with mean $40. Four independent runs each simulate $n=10$ losses at confidence level $p=0.80$. In each run, the third-largest loss (the empirical threshold) and the two largest losses are reported: Run 1: $25, $42, $50; Run 2: $28, $38, $54; Run 3: $30, $45, $61; Run 4: $27, $40, $57. Estimate TVaR in each run, then calculate the mean, sample standard deviation, and standard error of the four estimates. Round final dollar amounts to two decimal places.

Method. Each run has tail size $k=10(1-0.80)=2$, so its estimate is the average of its two largest losses. The third-largest loss identifies the threshold; it is not included because the tail size is exactly two.

Run 1: $(42+50)/2=46$. Run 2: $(38+54)/2=46$. Run 3: $(45+61)/2=53$. Run 4: $(40+57)/2=48.5$. Thus the four estimates are $46.00, $46.00, $53.00, and $48.50.

Their mean is $(46+46+53+48.5)/4=193.5/4=$48.375$, which rounds to $48.38. The squared deviations from $48.375 are $5.640625$, $5.640625$, $21.390625$, and $0.015625$; their sum is $32.6875$. Therefore $s_T=\sqrt{32.6875/3}=\sqrt{10.895833\ldots}=3.3009\ldots$, or $3.30. The estimated standard error of the mean across the four runs is $3.3009\ldots/\sqrt{4}=$1.6505\ldots$, or $1.65.

Check and common mistake. The run-to-run standard deviation is about $3.30, while the standard error of the four-run average is about $1.65. Dividing by $\sqrt{4}$ estimates the variability of the average of four runs; it does not describe the variability of a single run.

Ties, Interpretation, and Practical Limitations

Consider a hypothetical discrete annual loss model that assigns probability 0.2 to each of $0, $5, $10, $20, and $40. A simulation of ten losses yields the ordered sample $0, $0, $5, $5, $10, $10, $20, $20, $40, and $40. At $p=0.75$, the tail size is $10(1-0.75)=2.5$. The empirical threshold is the eighth observation, $20, and the quantile-integral calculation uses the two $40 losses and half of one $20 loss: $(40+40+0.5(20))/2.5=$36.00$.

A common alternative is to average all observations at or above the threshold. That would give $(20+20+40+40)/4=$30.00$, which is not the empirical quantile-integral TVaR here. The two calculations answer different questions because the first includes a fractional amount of threshold mass while the second includes all observations tied at the threshold. Use the definition specified for the task, and do not silently replace it with a strict-exceedance or inclusive-exceedance average.

Monte Carlo estimation has practical limits. At a high confidence level, only a small fraction of simulated losses contribute to the estimate. Increasing the number of simulated losses generally gives more tail observations, but a large sample does not correct an unsuitable model. Repeated runs help reveal simulation variability; differences between models or parameter choices are separate sources of uncertainty.

Check the tail size. Calculate $n(1-p)$ before selecting observations.
Check the weights. Their sum must equal $n(1-p)$ before dividing.
Check the threshold convention. A threshold tie does not mean every tied observation is automatically included in a quantile-integral tail average.
Separate sources of error. Repeated simulation runs assess random-number variation conditional on the assumed model, not model or parameter uncertainty.

Key takeaway: Estimate empirical TVaR by averaging the upper $n(1-p)$ simulated observations, using a fractional weight when that tail size is not an integer. Independent runs quantify run-to-run Monte Carlo variability; they do not establish that the assumed loss model is correct.

Check Your Understanding

For each problem, use the empirical quantile-integral convention unless the question asks for a simulation-variability calculation.

A simulation produces eight ordered losses: $1, $3, $4, $6, $8, $10, $14, and $22. Estimate TVaR at $p=0.75$. (Answer: $18.00)
Ten simulated losses, in increasing order, are $0, $2, $3, $4, $5, $7, $9, $12, $18, and $30. Estimate TVaR at $p=0.85$, showing how the fractional tail is handled. (Answer: $26.00)
Four independent runs give TVaR estimates of $90, $100, $95, and $105. Calculate their mean and the standard error of their mean using the sample standard deviation. Round to two decimal places. (Answer: Mean $97.50; standard error $3.23)
Eight ordered simulated losses are $0, $0, $5, $10, $10, $20, $20, and $50. At $p=0.625$, calculate the empirical quantile-integral TVaR, and compare it with the average of all losses at or above the empirical threshold. (Answer: Quantile-integral TVaR $30.00; average at or above the threshold $22.00)
Repeated runs from one fitted model show little variation in estimated TVaR. Does this establish that the fitted model is appropriate for the true loss process? (Answer: No; it indicates limited run-to-run Monte Carlo variability, not model appropriateness.)

What You'll Learn