Parameter Uncertainty in VaR and TVaR

Why Risk-Measure Estimates Have Uncertainty

A fitted model produces estimates of risk measures such as Value-at-Risk and Tail Value-at-Risk. Those estimates are uncertain because the model parameters were estimated from a finite data set. This is distinct from the variability of a future loss: even if the model parameters were known exactly, future losses would still vary.

This tutorial focuses on uncertainty conditional on a specified parametric model. It does not quantify uncertainty about whether that model is appropriate. The delta method provides an approximation based on parameter-estimator variability and local sensitivity. A parametric bootstrap estimates variability by repeatedly generating data from the fitted model and refitting it.

Definition: For a continuous loss $X$ with distribution function $F_X$, density $f_X$, and finite tail mean, define the $p$-level Value-at-Risk by $q_p=F_X^{-1}(p)$. Define the $p$-level Tail Value-at-Risk by $\operatorname{TVaR}_p=E[X\mid X>q_p]$. A fitted estimate replaces the unknown model parameter $\theta$ with an estimate $\hat{\theta}$. Parameter uncertainty is the sampling variability of the resulting estimate across repeated data sets; it is not the variability of $X$ itself. These definitions use a continuous distribution; for a distribution with atoms, the tail-mean convention must be specified.

The Delta Method for VaR

Suppose a model has parameter vector $\theta$, estimated by $\hat{\theta}$, with approximate covariance matrix $\Sigma$. Under standard large-sample conditions, $\hat{\theta}$ is approximately multivariate normal around $\theta$. If a risk measure is a differentiable function $g(\theta)$, its estimated variance is approximated by the gradient of $g$ applied to the parameter covariance matrix.

\widehat{\operatorname{Var}}\{g(\hat{\theta})\} \approx \nabla g(\hat{\theta})^{\mathsf T}\hat{\Sigma}\nabla g(\hat{\theta}).

For a quantile, the gradient follows by differentiating the defining equation. Assume $F_X(q_p;\theta)=p$, that $f_X(q_p;\theta)>0$, and that the distribution is sufficiently smooth in its parameters. Differentiating with respect to $\theta$ gives

\nabla_\theta F_X(q_p;\theta) + f_X(q_p;\theta)\nabla_\theta q_p=0, \qquad \nabla_\theta q_p = -\frac{\nabla_\theta F_X(q_p;\theta)}{f_X(q_p;\theta)}.

The density in the denominator matters: where the density near the quantile is small, a modest change in the fitted distribution can move the quantile substantially. The approximation can be unreliable when the density is near zero, estimates lie at parameter boundaries, or the estimator’s sampling distribution is strongly non-normal.

Worked Example: VaR and TVaR Under an Exponential Model

A hypothetical portfolio is modeled with independent losses from an exponential distribution with scale parameter $\theta$, so $F_X(x)=1-e^{-x/\theta}$ for $x\geq 0$. From $n=100$ observations, the maximum likelihood estimate is $\hat{\theta}=100$ dollars. Estimate 99% VaR and TVaR, and their standard errors using the delta method. Use a 95% Wald interval for each estimate.

Method and assumptions. For an exponential model with known form and estimated scale, the maximum likelihood estimator is the sample mean, with approximate variance $\theta^2/n$. Substitute $\hat{\theta}$ to estimate that variance. The model is continuous and has a positive density at the 99th percentile, so the quantile delta method applies.

VaR. Solving $1-e^{-q_p/\theta}=p$ gives $q_p=-\theta\ln(1-p)$. At $p=0.99$, $-\ln(0.01)=4.60517$, so

\widehat{q}_{0.99}=100(4.60517)=\$460.52.

The estimated standard error of $\hat{\theta}$ is $100/\sqrt{100}=\$10$. Since the quantile is $4.60517\theta$, its standard error is $4.60517(10)=\$46.05$. The 95% Wald interval is $460.517\pm1.96(46.052)$, or $[\$370.26,\$550.78]$.

TVaR. By the memoryless property of the exponential distribution, the expected loss above a threshold $q$ is $q+\theta$. Thus $\operatorname{TVaR}_{0.99}=\theta(4.60517+1)$. The estimate is $100(5.60517)=\$560.52$, and its standard error is $5.60517(10)=\$56.05$. Its 95% Wald interval is $560.517\pm1.96(56.052)$, or $[\$450.66,\$670.38]$.

Check. The estimated TVaR exceeds the estimated VaR by exactly $\hat{\theta}=\$100$, as expected for an exponential tail. The calculation avoids treating the loss standard deviation, $\$100$, as the standard error of either fitted risk measure.

Propagating Uncertainty into TVaR

For a continuous loss with finite mean, the tail expectation can also be written in terms of the survival function $S_X(x)=1-F_X(x)$:

\operatorname{TVaR}_p = q_p+\frac{\int_{q_p}^{\infty}S_X(x)\,dx}{1-p}.

This expression follows because $E[(X-q_p)_+]=\int_{q_p}^{\infty}S_X(x)\,dx$, while $P(X>q_p)=1-p$. Conditional on exceeding $q_p$, the expected excess is therefore the integral divided by $1-p$; adding $q_p$ gives the conditional tail mean. When the model parameters affect both the quantile and the survival probabilities above it, both effects contribute to TVaR uncertainty. If an explicit differentiable formula for $\operatorname{TVaR}_p(\theta)$ is available, apply the same gradient-covariance formula as for any other smooth function.

Worked Example: Two Estimated Parameters and a Normal-Model VaR

For a hypothetical aggregate-loss approximation, suppose $X\sim N(\mu,\sigma^2)$, with estimates $\hat{\mu}=100$ and $\hat{\sigma}=20$. The estimated covariance matrix of $(\hat{\mu},\hat{\sigma})$ is diagonal, with variances 4 and 1. Find the 95% VaR and its delta-method standard error. This calculation illustrates propagation with two fitted parameters; a normal model should not be used as a severity model without checking whether its negative-loss support is acceptable.

Method. Write $q_p=\mu+z_p\sigma$, where $z_p$ is the $p$-quantile of the standard normal distribution. At $p=0.95$, $z_p=1.64485$. The gradient with respect to $(\mu,\sigma)$ is $(1,z_p)$. The estimated covariance matrix is $\begin{pmatrix}4&0\\0&1\end{pmatrix}$, so

\widehat{q}_{0.95}=100+1.64485(20)=132.897, \qquad \widehat{\operatorname{Var}}(\hat{q}_{0.95}) = (1,1.64485) \begin{pmatrix}4&0\\0&1\end{pmatrix} \begin{pmatrix}1\\1.64485\end{pmatrix} = 4+1.64485^2=6.7055.

Therefore the estimated standard error is $\sqrt{6.7055}=2.590$. The result combines uncertainty in the estimated mean and uncertainty in the estimated standard deviation; it does not use the future-loss standard deviation of 20 as the quantile’s standard error.

For comparison, the normal-model TVaR has the form $\mu+\sigma\phi(z_p)/(1-p)$, where $\phi$ is the standard normal density. At $z_{0.95}=1.64485$, $\phi(z_{0.95})\approx0.10314$, so the multiplier is approximately $0.10314/0.05=2.0627$. The estimated TVaR is $100+20(2.0627)=141.25$. Its gradient is $(1,2.0627)$, giving variance $4+2.0627^2\approx8.2548$ and standard error approximately $2.873$. The calculation uses the same parameter covariance matrix as the VaR calculation.

Check. Both calculations use the full covariance matrix. If the parameter estimates were correlated, the covariance term would also contribute; omitting it can overstate or understate the estimated risk-measure variance.

Parametric Bootstrap

The parametric bootstrap provides another way to estimate parameter uncertainty, especially when a risk measure is difficult to differentiate or its delta-method approximation is questionable. It is conditional on the chosen model: it simulates data from the fitted distribution, not from an unknown true distribution.

Fit the model.
Estimate the parameters from the observed data and calculate the fitted VaR or TVaR.

Generate bootstrap samples.
For each replicate, simulate a data set of the same size from the fitted model, following the original sampling design.

Refit and recalculate.
Refit the parameters in each simulated data set, then calculate the target risk measure using that replicate’s estimates.

Summarize the replicates.
The standard deviation of the replicate risk measures estimates their standard error. Their empirical quantiles can be used for a bootstrap interval.

Worked Example: Bootstrap Standard Error for Exponential VaR

A fitted exponential model is used for a hypothetical loss sample. For illustration, five parametric bootstrap refits produce scale estimates $\$80,\$90,\$100,\$110,\$120$. Calculate the bootstrap standard error of the fitted 99% VaR. The small number of replicates is for showing the arithmetic only; a practical analysis would use many more.

Method. For each refit, calculate $q_{0.99}^{*}=\hat{\theta}^{*}[-\ln(0.01)]$, using the multiplier $4.60517$. The five bootstrap VaR estimates are $\$368.41,\$414.47,\$460.52,\$506.57,\$552.62$.

Their mean is $4.60517(100)=\$460.52$. The sample variance of the five scale estimates is

\frac{(80-100)^2+(90-100)^2+(100-100)^2+(110-100)^2+(120-100)^2}{5-1} = \frac{1000}{4}=250.

Because each bootstrap VaR is the scale estimate multiplied by $4.60517$, the sample standard deviation of the VaR estimates is $4.60517\sqrt{250}=\$72.81$. This is the bootstrap estimate of the VaR standard error from these five replicates. It need not equal the delta-method estimate: bootstrap variability itself is estimated from a finite number of replicates, and this deliberately small set is especially unstable.

Check. The bootstrap reports the variability of a refitted risk estimate across simulated data sets. It does not report the standard deviation of future losses, which under the fitted exponential model is $\$100$.

What the Uncertainty Measure Does—and Does Not—Say

A delta-method standard error or bootstrap interval describes sampling uncertainty in a fitted risk measure, under the assumed model and fitting procedure. It does not describe the range of individual losses, guarantee coverage if the model is misspecified, or automatically incorporate uncertainty from choosing among different model families. If the data are censored, truncated, weighted, or dependent, the simulation and refitting steps must preserve those features; otherwise the bootstrap estimates the wrong sampling process.

A useful practical distinction is:

Quantity	What varies?	Typical calculation
Future-loss variability	The loss $X$, with parameters held fixed	$\operatorname{Var}(X)$ under the fitted model
VaR or TVaR estimation uncertainty	The fitted parameter estimates, across repeated data sets	Delta-method standard error or bootstrap standard deviation

The distinction is consequential in risk reporting. A high VaR may reflect a genuinely heavy modeled tail, substantial parameter uncertainty, or both. Reporting only the point estimate conceals the precision of the fitted risk measure. Conversely, a small standard error does not establish that the selected tail model is appropriate.

Common Exam Trap

A frequent error is to substitute the standard deviation of future losses for the standard error of estimated VaR. In the exponential example, the fitted loss standard deviation is $\hat{\theta}=\$100$. It would be wrong to report this as the standard error of the 99% VaR. The fitted quantile is $4.60517\hat{\theta}$, and the estimated standard error of $\hat{\theta}$ is $\$10$, so the delta-method VaR standard error is $4.60517(\$10)=\$46.05$. These answer different questions: loss dispersion versus estimation precision.

Another trap is to ignore covariance when several parameters are estimated. For $g(\theta_1,\theta_2)$, the variance approximation includes twice the product of the two gradient components and their covariance. A zero covariance can be used only when it is given or justified, not merely because it makes the arithmetic simpler.

Key takeaway: First identify what is random. For parameter uncertainty, quantify how the fitted risk measure changes as the estimated parameters vary. Use the delta method when the risk measure is smooth and its gradient and parameter covariance are available; use a correctly specified parametric bootstrap when repeated refitting is more practical. Neither measure is the same as future-loss variability.

Check Your Understanding

For each problem, treat the stated model and parameter covariance information as given. Distinguish the estimated risk measure from its estimation standard error.

A hypothetical exponential loss model has scale estimate $\$50$ from $n=25$ observations. Using the large-sample variance $\hat{\theta}^2/n$, calculate the fitted 95% VaR and its delta-method standard error. (Answer: VaR $\$149.79$; standard error $\$29.96$.)
For a normal model, $\hat{\mu}=40$, $\hat{\sigma}=5$, and the estimated covariance matrix of $(\hat{\mu},\hat{\sigma})$ is $\begin{pmatrix}9&-1\\-1&1\end{pmatrix}$. At the 90th percentile, use $z_{0.90}=1.2816$ to calculate VaR and its delta-method standard error. (Answer: VaR $46.41$; standard error $2.84$.)
For an exponential model, four bootstrap refits give scale estimates $\$40,\$50,\$60,\$70$. Using the sample standard deviation across refits and the 90% VaR multiplier $-\ln(0.10)=2.3026$, calculate the bootstrap standard error of VaR. (Answer: $\$29.74$.)
An exponential model has scale estimate $\$80$ from $n=64$ observations. Calculate fitted 95% TVaR and its delta-method standard error, using $-\ln(0.05)=2.9957$ and $\operatorname{TVaR}_p=\theta[-\ln(1-p)+1]$. (Answer: TVaR $\$319.66$; standard error $\$39.96$.)

Quantity	What varies?	Typical calculation
Future-loss variability	The loss \(X\), with parameters held fixed	\(\operatorname{Var}(X)\) under the fitted model
VaR or TVaR estimation uncertainty	The fitted parameter estimates, across repeated data sets	Delta-method standard error or bootstrap standard deviation

What You'll Learn