Monte Carlo Precision for Tail-Risk Estimates

Why Monte Carlo Precision Matters

A simulation produces estimates, not exact risk measures. Two independent runs with the same model and the same number of simulated losses can give different estimates of value-at-risk (VaR) and tail value-at-risk (TVaR). That variation is Monte Carlo error. A reported estimate without an assessment of this error can make an unstable result look definitive.

Precision is particularly important in the tail. At confidence level $p$, only about $n(1-p)$ of $n$ simulated losses are expected to exceed the VaR. If $p=0.99$, for example, a run of 2,000 simulations has only about 20 observations in the upper one-percent tail. A large total simulation count does not necessarily mean that the tail estimate is precise.

Definition: For a loss $X$, the VaR at confidence level $p$ is the $p$-quantile $q_p$, defined as $q_p=\inf\{x:F_X(x)\ge p\}$. For a continuous loss distribution, TVaR at level $p$ is $E[X\mid X>q_p]$. Monte Carlo precision describes the sampling uncertainty in estimates of these quantities when losses are generated by simulation.

Precision of a Simulated VaR

Suppose $X_1,\ldots,X_n$ are independent, identically distributed simulated losses from a continuous distribution with density $f_X$. Let $q_p$ be the true $p$-quantile and let $\widehat q_p$ be the corresponding empirical quantile. When $f_X(q_p)>0$, the large-sample standard error is approximately

\operatorname{SE}(\widehat q_p) \approx \sqrt{\frac{p(1-p)}{n f_X(q_p)^2}}.

The estimate is more precise when $n$ is larger and when the density near the quantile is higher. A thin density at the quantile means that a small change in probability corresponds to a large change in loss amount, making the estimated VaR more variable. In practice, $f_X(q_p)$ may be estimated from the model or from a density estimate based on the simulated losses. The formula is not reliable if the distribution has an atom at the quantile or if the density is zero or poorly behaved there.

An approximate two-sided confidence interval with confidence level $1-\gamma$ is

\widehat q_p \ \pm\ z_{1-\gamma/2}\operatorname{SE}(\widehat q_p),

where $z_{1-\gamma/2}$ is the corresponding standard normal quantile. This is a large-sample approximation, not a guarantee of exact coverage. It can be weak when the expected number $n(1-p)$ of tail observations is small.

Justification: The empirical fraction of losses at or below $q_p$ has approximate variance $p(1-p)/n$, because each simulated loss falls at or below $q_p$ with probability $p$. Near $q_p$, a small change $\Delta x$ in the loss threshold changes the cumulative probability by approximately $f_X(q_p)\Delta x$. Converting the variability in the empirical fraction to variability in the estimated quantile therefore divides its standard deviation by $f_X(q_p)$, giving $\sqrt{p(1-p)/n}/f_X(q_p)$.

Worked Example: A VaR Estimate and Its Approximate Interval

A hypothetical insurer simulates $n=10{,}000$ independent annual losses from a Uniform distribution on $[0,1000]$ dollars. The estimated 95% VaR from one run is $\widehat q_{0.95}=\$952.40$. Estimate its standard error and an approximate 95% confidence interval. Assume the stated simulation model is correct.

Method. Use the quantile standard-error formula with the model’s density at the true 95% quantile. For a Uniform$[0,1000]$ loss, the density is $1/1000=0.001$ per dollar throughout the interval. The true 95% quantile is $0.95(1000)=\$950$, so the density at that quantile is $0.001$.

The standard error is

\operatorname{SE}(\widehat q_{0.95}) \approx \sqrt{\frac{0.95(0.05)}{10{,}000(0.001)^2}} = \sqrt{\frac{0.0475}{0.01}} = \sqrt{4.75} \approx \$2.18.

For a 95% interval, use $z_{0.975}=1.96$. The margin of error is $1.96(2.179)\approx \$4.27$, so the interval is

\$952.40\pm \$4.27 = [\$948.13,\ \$956.67].

The interval is centered on the simulated estimate, not the true model quantile. The true value of \$950 lies inside it, which is a useful check but does not prove that the approximation has exact 95% coverage. The method avoids treating the simulated estimate as exact.

A distribution-free alternative for a continuous loss uses order statistics. Let $X_{(1)}\le\cdots\le X_{(n)}$ be the sorted simulated losses, and let $B$ be the number of losses at or below the true quantile $q_p$. Then $B\sim\operatorname{Binomial}(n,p)$. If integers $r<s$ are chosen so that $P(r\le B\le s-1)$ is at least the desired confidence level, then $[X_{(r)},X_{(s)}]$ is a confidence interval for $q_p$. This approach does not require estimating the density at the quantile, though the interval can be wide and its coverage is determined by the chosen binomial probabilities.

Tail Counts and TVaR Precision

For $n$ simulations, the number of losses above the true $p$-quantile has a binomial distribution with mean $n(1-p)$ and variance $np(1-p)$. This expected tail count is a basic precision check. It does not by itself determine the accuracy of a TVaR estimate: tail losses can vary greatly in size, and a few exceptionally large observations can have a substantial effect.

For a continuous loss distribution, write $\alpha=1-p$. An equivalent expression for TVaR is

\operatorname{TVaR}_p(X) = q_p+\frac{E[(X-q_p)_+]}{\alpha}, \qquad (y)_+=\max(y,0).

A natural simulation estimate replaces $q_p$ by $\widehat q_p$ and the expectation by the sample average of the positive excesses. Under regularity conditions, the first-order effect of estimating the quantile cancels in this expression: changing the threshold changes both the threshold term and the expected excess term. This gives the approximate standard error

\widehat{\operatorname{TVaR}}_p = \widehat q_p+ \frac{1}{n\alpha} \sum_{j=1}^{n}(X_j-\widehat q_p)_+, \qquad \widehat{\operatorname{SE}}(\widehat{\operatorname{TVaR}}_p) \approx \frac{s_Z}{\alpha\sqrt n},

where $Z_j=(X_j-\widehat q_p)_+$ and $s_Z$ is the sample standard deviation of $Z_1,\ldots,Z_n$. This approximation includes the zero excesses as well as the positive ones. It therefore reflects that only a fraction of simulations contribute excess losses. It requires a sufficiently large simulation and a finite variance for the excess loss; a very heavy tail can make the normal approximation unreliable.

For a continuous distribution, if $n\alpha$ is an integer and the empirical quantile convention agrees with the tail boundary, the empirical TVaR can also be computed as the average of the largest $n\alpha$ simulated losses. If $n\alpha$ is not an integer, the quantile-integral definition requires appropriate fractional weighting at the boundary; simply rounding the number of tail observations changes the estimate.

Worked Example: Tail Count and TVaR Uncertainty

A hypothetical simulation produces 100 sorted losses in thousands of dollars. The first 90 losses are $1,2,\ldots,90$, and the largest 10 are $100,110,\ldots,190$. Estimate the 90% VaR and TVaR, then calculate an approximate standard error for TVaR using the excess-loss formula.

Method. Here $p=0.90$, $\alpha=0.10$, and $n\alpha=10$, so the empirical upper tail contains exactly 10 observations. The empirical 90% quantile is the 90th ordered loss, $qhat=90$ thousand dollars. The 10 largest losses have mean

\widehat{\operatorname{TVaR}}_{0.90} = \frac{100+110+\cdots+190}{10} = \frac{1{,}450}{10} = 145 \text{ thousand dollars}.

For the standard-error calculation, the positive excesses above 90 are $10,20,\ldots,100$, and the other 90 excesses are zero. Thus

\sum_{j=1}^{100} Z_j=10+20+\cdots+100=550, \qquad \sum_{j=1}^{100} Z_j^2=10^2+20^2+\cdots+100^2=38{,}500.

The sample variance of the 100 excesses, in squared thousands of dollars, is

s_Z^2 = \frac{38{,}500-550^2/100}{99} = \frac{35{,}475}{99} \approx 358.33, \qquad s_Z\approx 18.93.

Therefore,

\widehat{\operatorname{SE}}(\widehat{\operatorname{TVaR}}_{0.90}) \approx \frac{18.93}{0.10\sqrt{100}} = 18.93 \text{ thousand dollars}.

A rough 95% interval is $145\pm1.96(18.93)$, or approximately $[107.90,182.10]$ thousand dollars. This interval is only a normal approximation; with 10 tail observations, its reliability may be limited. The method avoids reporting the apparent precision of the tail average without accounting for the variability of the simulated excesses.

Planning a Simulation Size

The VaR standard-error formula can be rearranged to plan a simulation. If the target approximate confidence-interval half-width is $h$, then a sufficient size under the same large-sample approximation is

n\ \ge\ \frac{z_{1-\gamma/2}^2p(1-p)} {h^2 f_X(q_p)^2}.

This calculation addresses VaR precision only. A simulation designed to estimate TVaR should also have a substantial number of tail observations and should account for the variability of excess losses. There is no universal tail-count cutoff that guarantees precision: the appropriate count depends on the required accuracy and on how variable the tail is.

Worked Example: Choosing a Simulation Size for VaR

A hypothetical risk analyst uses a Uniform$[0,1000]$-dollar loss model and wants the approximate 95% VaR confidence-interval half-width at level $p=0.99$ to be no more than \$5. Find the minimum simulation size under the quantile standard-error approximation, and assess the expected tail count.

Method. The true 99% quantile is \$990 and the density there is $f_X(q_{0.99})=0.001$ per dollar. Set the approximate half-width $1.96\operatorname{SE}$ equal to \$5 and solve for $n$:

n \ge \frac{1.96^2(0.99)(0.01)} {5^2(0.001)^2} = \frac{3.8416(0.0099)}{25(0.000001)} = \frac{0.03803184}{0.000025} = 1521.2736.

Since the number of simulations must be an integer, the minimum is $n=1{,}522$. The expected number of losses above the 99% quantile is only $1{,}522(0.01)=15.22$. Thus the VaR target is met approximately, but a TVaR estimate from this run would be based on very few tail observations and could remain unstable. For comparison, 100,000 simulations would give an expected tail count of 1,000. The calculation avoids assuming that satisfying a VaR precision target automatically makes the TVaR estimate precise.

Common Mistakes and Interpretation

Using the total simulation count as the tail count. At level $p$, the expected number of exceedances is $n(1-p)$, not $n$.
Ignoring the density in VaR precision. The standard error depends on $f_X(q_p)$; the binomial variability alone is not a loss-scale standard error.
Using the standard deviation of only the exceedances in the excess-loss formula. The stated influence-function standard error uses all $n$ excess variables, including zeros, and divides by $\alpha\sqrt n$. A conditional tail-mean calculation is a different approximation and should not be substituted without stating its assumptions.
Rounding the tail count without checking the definition. If $n(1-p)$ is not an integer, an empirical tail average may need fractional boundary weighting to match the quantile-integral definition of TVaR.
Interpreting an approximate interval as a guarantee. Normal confidence intervals rely on large-sample conditions. Heavy tails, few tail observations, or irregular distributions can undermine those conditions.

Precision assessment is aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus. In an applied analysis, report the estimate together with its simulation size, the expected or observed tail count, the standard-error method, and any assumptions behind the interval.

Check Your Understanding

Work each problem using the stated approximation and keep the units attached to your answers.

A simulated loss has $p=0.95$, $n=2{,}500$, and estimated density $f_X(q_p)=0.002$ per dollar. Calculate the approximate VaR standard error and 95% confidence-interval half-width. (Answer: standard error $\$2.18$; half-width $\$4.27$.)
A simulation uses $n=4{,}000$ losses to estimate 98% VaR and TVaR. What is the expected number of losses above the true VaR? (Answer: $4{,}000(0.02)=80$ losses.)
An estimated 95% TVaR is $\$620$, $n=1{,}000$, and the sample variance of $Z_j=(X_j-\widehat q_{0.95})_+$ is $\$90{,}000^2$. Use the excess-loss formula for the approximate standard error and 95% interval. (Answer: standard error $\$189.74$; interval approximately $[\$248.11,\$991.89]$.)
For a Uniform$[0,2000]$-dollar loss model, estimate the standard error of a simulated 99% VaR from $n=10{,}000$ simulations, using the model density at the quantile. Also state the expected tail count. (Answer: standard error approximately $\$1.99$; expected tail count $100$.)
For continuous simulated losses, explain how an order-statistic confidence interval for $q_p$ can be chosen without estimating a density. (Answer: choose $r<s$ so that $P(r\le B\le s-1)$ meets the target confidence level for $B\sim\operatorname{Binomial}(n,p)$, and use $[X_{(r)},X_{(s)}]$.)

What You'll Learn