Selecting a VaR Estimator in a Worked Case

Why the VaR Estimator Matters

Value at Risk (VaR) summarizes a loss distribution at a chosen confidence level, but the number depends on how that distribution is represented. An empirical estimate uses observed losses; a parametric estimate uses a specified probability model; and a Monte Carlo estimate uses simulated losses generated from a model. These methods can produce different answers because they rely on different evidence and assumptions.

Estimator selection is therefore a modeling decision, not simply a choice of calculation technique. A risk manager should ask whether the data are representative, whether a proposed distribution is credible, and whether the risk structure makes simulation necessary. A precise-looking quantile is not reliable if its data or model are unsuitable.

Definition: For a loss $X$ with cumulative distribution function $F_X$, the VaR at confidence level $\alpha$, where $0<\alpha<1$, is the lower quantile $$ \operatorname{VaR}_{\alpha}(X)=\inf\{x:F_X(x)\geq \alpha\}. $$ The corresponding exceedance probability is $1-\alpha$. For a finite sample of losses, this tutorial uses the nearest-rank empirical convention: sort the $n$ observations and select observation $k=\lceil n\alpha\rceil$. Other empirical quantile conventions exist, so a calculation should state which convention it uses.

Three Estimators and Their Evidence

The three methods differ in what they treat as the source of the loss distribution. The historical empirical method directly summarizes observed losses. It avoids choosing a parametric distribution, but it cannot show losses more extreme than those represented in the data, and its tail resolution is limited by sample size. A parametric method specifies a distribution and estimates or is given its parameters; it can estimate quantiles beyond the observations, but its result depends on the model being appropriate. A Monte Carlo method simulates losses from a specified model, then estimates the quantile from the simulated sample. It is useful for complex risks, but simulation does not remove errors in the model inputs or structure.

Estimator	What it uses	Useful when	Main limitation
Empirical	Observed historical losses	Data are relevant and a distributional assumption is not justified	Tail detail is limited by the sample and its representativeness
Parametric	A specified loss distribution and its parameters	A defensible model describes the risk and a direct quantile is available	Misspecification can distort the estimated quantile
Monte Carlo	Losses generated from a specified stochastic model	Portfolio features or dependencies make direct quantile calculation difficult	Model risk and simulation error both affect the estimate

A practical decision should begin with the target: define the loss measure, time horizon, and confidence level. Then examine the evidence. For historical data, consider whether the observation period reflects current exposures and conditions. For a parametric model, consider whether the chosen distribution and parameter estimates are defensible for the tail of interest. For simulation, check whether the risk drivers, dependencies, and loss mapping are adequately specified and whether enough simulations can be run.

Fix the quantity being estimated.
Specify the loss definition, horizon, and confidence level. Estimates based on different horizons or confidence levels are not directly comparable.

Assess the evidence.
Use empirical data when observations are sufficiently relevant for the target. Use a parametric distribution only when its form and parameters are supportable.

Check computational structure.
If the loss mapping or portfolio structure prevents a tractable direct calculation, simulation may be appropriate, provided the underlying model can be specified and validated.

Report the estimator’s uncertainty.
State its key assumptions and limitations. Empirical tail resolution, parametric model risk, and simulation error are different concerns and should not be treated as interchangeable.

Worked Cases

Worked Example: Empirical VaR with a Short Relevant Loss History

A hypothetical insurer is estimating the 95% VaR of one-year net losses, in thousands of dollars, for a small specialty portfolio. It has 20 annual observations judged relevant to the current coverage and mix of business. No parametric distribution has been supported by the available evidence. The ordered losses are $18,20,21,23,24,26,27,29,30,32,34,35,37,39,41,44,47,52,68,110$. Select and calculate an estimator using the stated nearest-rank convention.

Method and selection. Use empirical VaR. The case provides relevant observations but no defensible parametric model. Historical simulation is therefore a transparent estimate from the available data, although the short sample gives limited detail about the upper tail.

Step 1: Determine the rank. Here $n=20$ and $\alpha=0.95$, so $$ k=\lceil n\alpha\rceil=\lceil 20(0.95)\rceil=\lceil 19\rceil=19. $$

Step 2: Select the ordered loss. The 19th observation is 68 (thousand dollars). Thus the empirical estimate is $\$68{,}000$.

Check. At this value, 19 of 20 observations are at or below $\$68{,}000$, giving an empirical cumulative proportion of $19/20=0.95$. The estimate is consistent with the specified quantile convention. The largest observation, $\$110{,}000$, is above it.

Common mistake avoided. Selecting the 20th observation because 95% is “near the top” would give $\$110{,}000$, the sample maximum. Under the stated convention, the rank is $\lceil 19\rceil=19$, not 20. The estimate should also be described as sample-based rather than as a guaranteed bound on future losses.

Worked Example: Parametric VaR from a Normal Loss Model

A hypothetical risk unit has one-year losses $X$ modeled as normal with mean $\mu=\$2.4$ million and standard deviation $\sigma=\$0.6$ million. Assume this model and its parameters are appropriate for the risk being measured. Calculate the 97.5% VaR and explain why a parametric estimator is suitable for this case.

Method and selection. Use the normal-model quantile. The case explicitly supplies a distribution and parameters, and the normal quantile is available directly. This calculation is conditional on the stated normal model; the arithmetic alone does not establish that the model fits real losses.

Step 1: Standardize the loss. If $X\sim N(\mu,\sigma^2)$, then $Z=(X-\mu)/\sigma\sim N(0,1)$. Let $z_{\alpha}$ be the standard normal quantile satisfying $P(Z\leq z_{\alpha})=\alpha$. Then $$ P\left(X\leq \mu+\sigma z_{\alpha}\right) =P\left(\frac{X-\mu}{\sigma}\leq z_{\alpha}\right) =\alpha. $$ Consequently, $\operatorname{VaR}_{\alpha}(X)=\mu+\sigma z_{\alpha}$ for this continuous normal model.

Step 2: Substitute the values. At $\alpha=0.975$, $z_{0.975}\approx1.96$. Therefore, $$ \operatorname{VaR}_{0.975}(X) =\$2.4\text{ million}+1.96(\$0.6\text{ million}) =\$2.4\text{ million}+\$1.176\text{ million} =\$3.576\text{ million}. $$

Check. Standardizing the result gives $(3.576-2.4)/0.6=1.96$, the stated 97.5th standard normal quantile. The estimate is $\$3{,}576{,}000$.

Common mistake avoided. The notation $N(\mu,\sigma^2)$ uses variance as its second parameter. Here the standard deviation is $\$0.6$ million, not $(\$0.6\text{ million})^2$. Multiplying the variance by the normal quantile would use the wrong scale and produce the wrong units.

Worked Example: Monte Carlo VaR for a Nonlinear Loss

A hypothetical portfolio has one-year net loss, in millions of dollars, $$ X=\max\left(0,\ 2.0+0.9Z_1+0.6Z_2+0.2Z_1^2\right), $$ where $Z_1$ and $Z_2$ are jointly normal with means zero, variances one, and correlation $0.40$. A direct quantile calculation for this nonlinear loss is not available in the case. A validated simulation produces 10,000 losses; in the ascending sorted simulation, observations 9,899, 9,900, and 9,901 are respectively $\$5.89$ million, $\$5.92$ million, and $\$5.94$ million. Estimate the 99% VaR using the nearest-rank convention.

Method and selection. Use Monte Carlo VaR. The model specifies correlated risk factors and a nonlinear loss function, so simulation can propagate the joint risk-factor outcomes through the loss calculation. The result remains conditional on the stated model and its inputs.

Step 1: Find the required rank. For $n=10{,}000$ simulations and $\alpha=0.99$, $$ k=\lceil n\alpha\rceil=\lceil 10{,}000(0.99)\rceil=9{,}900. $$

Step 2: Read the simulated order statistic. The 9,900th simulated loss is $\$5.92$ million. The Monte Carlo estimate of $\operatorname{VaR}_{0.99}(X)$ is therefore $\$5{,}920{,}000$.

Check. The given adjacent order statistics satisfy $5.89<5.92<5.94$, so the selected 9,900th value is correctly located in the sorted simulation. Under the empirical convention, 9,900 of the 10,000 simulated losses are at or below the estimate.

Common mistake avoided. Taking the 99th observation would confuse a 99% confidence level with a rank of 99. The rank depends on both the confidence level and the number of simulated losses. Also, the 10,000 simulations are not 10,000 new historical observations: they are generated from the specified model.

Justifying the Choice and Interpreting the Result

The normal-model derivation illustrates a general point about estimator choice: each calculation is justified by the representation of the loss distribution. The empirical method estimates the distribution using the observed sample’s relative frequencies. The parametric method obtains a quantile from the assumed distribution. The Monte Carlo method approximates the model distribution by generating many outcomes and taking a sample quantile. Simulation can address computational complexity, but it does not turn an unsupported model into a reliable one.

Several constraints can affect the decision. A historical sample may not reflect current exposures or conditions; even a relevant sample may contain too few observations to resolve a high quantile. A parametric model can estimate beyond the observed range, but the tail result may be sensitive to the distributional form and parameter estimates. Monte Carlo can represent nonlinear loss functions and dependencies, but requires a sound specification of risk factors and a sufficiently precise simulation. The appropriate choice is the method whose assumptions best match the case and can be explained and supported.

Key takeaway: Choose the estimator by evaluating the data, distributional assumptions, and computational requirements together. State the method and quantile convention, calculate the corresponding rank or model quantile, and make the estimate’s limitations clear.

Common Selection Traps

Treating more calculation as more evidence. A large simulation can reduce simulation error, but it cannot correct an inappropriate risk model.
Assuming historical data are automatically representative. Observed losses are useful only to the extent that the observation period and portfolio are relevant to the target risk.
Using a parametric tail without checking the assumption. A convenient distribution is not, by itself, evidence that the distribution describes the relevant tail.
Ignoring the quantile convention. Different sample-quantile conventions can select or interpolate different values, especially in small samples. State the convention used.
Confusing confidence level with sample rank. For a finite sample, the rank is based on both $n$ and $\alpha$, not on $\alpha$ alone.

Check Your Understanding

For each case, identify the suitable estimator or apply the stated convention. Answers use the data and assumptions given in each question.

A portfolio has eight relevant ordered annual losses, in thousands of dollars: $2,3,3.5,4,4.5,6,8,12$. Using the nearest-rank empirical convention, calculate the 87.5% VaR. (Answer: $\$8{,}000$)
A hypothetical one-year loss is modeled as normal with mean $\$800{,}000$, standard deviation $\$150{,}000$, and 95% confidence level. Use $z_{0.95}=1.645$ to calculate VaR. (Answer: $\$1{,}046{,}750$)
A portfolio’s losses are generated by several correlated risk factors and a nonlinear payoff, and no closed-form loss distribution is available. A validated model and adequate computing capacity are available. Which estimator is most suitable, and what key limitation should be disclosed? (Answer: Monte Carlo VaR; disclose model risk as well as simulation error.)
A simulation has 5,000 losses. Under the nearest-rank convention, what rank estimates the 98% VaR? (Answer: rank 4,900)
A 99% empirical VaR is requested from 25 observations, but the observations cover an old portfolio whose exposures differ materially from the current one. Is the empirical estimator automatically suitable because it avoids a distributional assumption? (Answer: No; the observations may not represent current risk, so estimator choice requires relevant data or a defensible alternative model.)

What You'll Learn