Value-at-Risk for Continuous Loss Distributions

Introduction

Value-at-risk (VaR) gives a loss threshold associated with a chosen probability level. For a specified confidence level, it answers: what loss amount will the modeled loss not exceed with that probability? This makes VaR a direct way to summarize a loss distribution when a single threshold is needed for risk assessment or capital analysis.

For a continuous loss distribution, the calculation is based on the inverse cumulative distribution function. The essential exam skills are to interpret the confidence level correctly, choose the right tail probability, and use the distribution’s parameters consistently. This tutorial develops those skills for uniform, exponential, and normal loss models, aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus.

Definition: Let $X$ be a loss with cumulative distribution function $F_X(x)=P(X\leq x)$. For $0<p<1$, the value-at-risk at confidence level $p$ is the $p$-quantile $$ \operatorname{VaR}_p(X)=F_X^{-1}(p). $$ For a continuous distribution that is strictly increasing at the quantile, this is the unique value $q$ satisfying $F_X(q)=p$. Thus $P(X\leq q)=p$ and $P(X>q)=1-p$.

Confidence Level and Tail Probability

A confidence level $p$ describes the probability to the left of the VaR threshold. Its complement, $1-p$, is the probability in the upper tail. For example, a 99% confidence level corresponds to a 1% upper-tail probability. These are two descriptions of the same quantile, not two different VaR calculations.

The quantile convention is important. In this tutorial, the subscript $p$ in $\operatorname{VaR}_p(X)$ denotes the cumulative probability, or confidence level. For continuous losses there is no probability mass concentrated at a single value, so the strict and non-strict upper-tail probabilities agree at the threshold.

p=P(X\leq \operatorname{VaR}_p(X)), \qquad 1-p=P(X>\operatorname{VaR}_p(X)).

If a problem instead gives an upper-tail probability $\varepsilon$, convert it first: the matching confidence level is $p=1-\varepsilon$. Then evaluate the $p$-quantile. Use the loss distribution specified in the problem; do not confuse a quantile of $X$ with a quantile of a different variable, such as a payment after a coverage modification.

Why Inverting the CDF Gives the VaR

The inverse-CDF formula follows directly from the meaning of the threshold. Suppose $X$ has a continuous distribution whose CDF is strictly increasing near the desired quantile. A threshold $q$ captures a proportion $p$ of losses at or below it precisely when $F_X(q)=P(X\leq q)=p$. Since $F_X$ is strictly increasing there, the equation has a unique solution, namely $q=F_X^{-1}(p)$. The probability above that threshold is the remaining probability $1-p$.

F_X(q)=p \quad\Longrightarrow\quad q=F_X^{-1}(p)=\operatorname{VaR}_p(X), \qquad P(X>q)=1-F_X(q)=1-p.

For a continuous distribution whose CDF is not strictly increasing everywhere, the generalized inverse $F_X^{-1}(p)=\inf\{x:F_X(x)\geq p\}$ defines the quantile. The calculations below use distributions for which the relevant quantiles are unique. Discrete losses and distributions with atoms require care at jumps in the CDF and are not treated here.

Uniform Losses

Suppose a hypothetical loss is uniformly distributed on $[a,b]$, where $a<b$. Its CDF on the support is $F_X(x)=(x-a)/(b-a)$. Setting this equal to $p$ and solving gives a simple linear quantile formula.

\operatorname{VaR}_p(X)=a+p(b-a), \qquad 0<p<1.

Worked Example: Uniform Loss and a 97.5% Confidence Level

A hypothetical equipment warranty portfolio is modeled with annual loss $X\sim\operatorname{Uniform}(0,40{,}000)$, with amounts in dollars. Calculate the VaR at 97.5% confidence and state the corresponding upper-tail probability. Assume losses are measured at one valuation date without discounting; no interest rate or mortality assumption applies.

Method. Use the uniform quantile formula. The model parameters are $a=0$ and $b=40{,}000$, and the confidence level is $p=0.975$.

Substitute into the formula:

\operatorname{VaR}_{0.975}(X) =0+0.975(40{,}000-0) =0.975(40{,}000) =39{,}000.

Therefore, the 97.5% VaR is $39,000.00. The upper-tail probability is $1-0.975=0.025$, or 2.5%. As a check, the uniform CDF at $39,000 is $39{,}000/40{,}000=0.975$, as required.

Common mistake avoided. Using the upper-tail probability, 0.025, as the CDF probability would give $0.025(40{,}000)=\$1{,}000$. That is the 2.5th percentile, not the 97.5th percentile. The requested confidence level determines the probability to the left of the VaR.

Exponential Losses

Use the rate parameterization for an exponential loss: $X$ has density $f_X(x)=\lambda e^{-\lambda x}$ for $x\geq0$, where $\lambda>0$. Its CDF is $F_X(x)=1-e^{-\lambda x}$. The mean is $1/\lambda$; if the model is specified by its mean $\theta$, then $\lambda=1/\theta$.

To find the $p$-quantile, set the CDF equal to $p$ and solve for $x$. This gives a logarithm, and the negative sign is necessary because $\ln(1-p)$ is negative for $0<p<1$.

p=1-e^{-\lambda x} \quad\Longrightarrow\quad e^{-\lambda x}=1-p \quad\Longrightarrow\quad \operatorname{VaR}_p(X)=-\frac{\ln(1-p)}{\lambda}.

Worked Example: Exponential Loss Specified by Its Mean

A hypothetical small-business liability portfolio has annual loss $X$ modeled as exponential with mean $\theta=\$8{,}000$. Calculate the 99% VaR and verify the upper-tail probability. Losses are evaluated at one valuation date without discounting; no interest rate or mortality assumption applies.

Method. Convert the mean to the rate parameter, then use the exponential quantile formula. Since $E[X]=1/\lambda=\theta$,

\lambda=\frac{1}{8{,}000}=0.000125 \quad\text{per dollar}.

Here $p=0.99$, so $1-p=0.01$. Substitution gives

\operatorname{VaR}_{0.99}(X) =-\frac{\ln(0.01)}{0.000125} =\frac{4.605170186}{0.000125} =36{,}841.3615\ldots.

Rounded to cents, the 99% VaR is $36,841.36. The model’s survival function is $P(X>x)=e^{-\lambda x}$. At the calculated threshold it is $e^{-0.000125(36{,}841.3615)}=e^{-4.605170186}=0.01$, confirming a 1% upper-tail probability.

Common mistake avoided. The mean $8,000 is not the exponential rate. Substituting 8,000 for $\lambda$ would give a dimensionally and numerically incorrect result. Use $\lambda=1/\theta$ when the mean, rather than the rate, is given.

Normal Loss Models

For a normal model $X\sim N(\mu,\sigma^2)$, $\mu$ is the mean and $\sigma$ is the standard deviation; the second parameter in this notation is the variance. Standardize with $Z=(X-\mu)/\sigma$, where $Z\sim N(0,1)$. If $\Phi$ is the standard normal CDF and $z_p=\Phi^{-1}(p)$, solving $P(X\leq x)=p$ gives

\operatorname{VaR}_p(X)=\mu+\sigma z_p.

A normal model has support over all real numbers, so it can assign positive probability to negative losses. The formula remains mathematically valid, but whether a normal model is suitable for a particular nonnegative loss depends on the context and the model assumptions. For the calculation, use the standard deviation, not the variance, as the multiplier of $z_p$.

Worked Example: Normal Loss at 99% Confidence

A hypothetical annual loss model is $X\sim N(120{,}000,30{,}000^2)$, in dollars. Calculate the 99% VaR and identify the probability above that threshold. The loss is measured at one valuation date without discounting; no interest rate or mortality assumption applies.

Method. Use the normal quantile formula. The mean is $\mu=\$120{,}000$, the standard deviation is $\sigma=\$30{,}000$, and the standard normal 99th percentile is $z_{0.99}\approx2.3263479$.

\begin{aligned} \operatorname{VaR}_{0.99}(X) &=120{,}000+30{,}000(2.3263479)\\ &=120{,}000+69{,}790.437\\ &=189{,}790.437. \end{aligned}

Rounded to cents, the 99% VaR is $189,790.44. By the definition of the 99th percentile, $P(X\leq189{,}790.44)$ is approximately 0.99 and the probability above the unrounded quantile is $1-0.99=0.01$. The small difference that may result from evaluating the rounded dollar amount is due only to rounding.

Common mistake avoided. The variance is $30{,}000^2$, but the formula requires the standard deviation $30{,}000$. Multiplying $z_{0.99}$ by $30{,}000^2$ does not produce a dollar amount and is not a valid standardization.

Choosing and Checking the Calculation

The calculation depends on both the distributional assumption and the stated probability convention. Before substituting values, check that the model’s parameters use the same definitions as the formula and that the requested level is a cumulative confidence level or an upper-tail probability. A quick probability check is often enough to detect a reversed quantile.

Identify the loss model.
Record its distribution and parameters, including whether an exponential parameter is a rate or a mean and whether a normal parameter is a variance or standard deviation.

Set the probability convention.
For confidence level $p$, use the $p$-quantile. If given upper-tail probability $\varepsilon$, convert using $p=1-\varepsilon$.

Invert the model CDF.
Apply the distribution-specific quantile formula or solve $F_X(q)=p$.

Check interpretation and units.
Confirm that the probability at or below the threshold is $p$, that the probability above it is $1-p$, and that the result has loss units.

A useful general check is monotonicity: for the same loss distribution, a higher confidence level cannot have a lower quantile. Thus a 99% VaR should be at least as large as a 95% VaR. This check does not replace calculation, but it can reveal that the confidence level and tail probability have been interchanged.

Key takeaway: For a continuous loss with a unique $p$-quantile, $\operatorname{VaR}_p(X)$ is the value $q$ such that $F_X(q)=p$. The probability above the threshold is $1-p$; keep that distinction clear when selecting or inverting the CDF.

Check Your Understanding

For each problem, identify the relevant probability level before calculating. Answers are rounded to cents where money is requested.

A hypothetical repair-cost loss is $X\sim\operatorname{Uniform}(0,12{,}000)$. Find its 90% VaR and the probability that the loss exceeds the threshold. (Answer: VaR = $10,800.00; exceedance probability = 0.10.)
A hypothetical claim loss is exponential with mean $2,500. Find the 96% VaR. (Answer: $8,047.19.)
A hypothetical annual loss is $X\sim N(50{,}000,10{,}000^2)$. Using $z_{0.95}=1.644854$, find the 95% VaR. (Answer: $66,448.54.)
A hypothetical exponential loss has rate $\lambda=0.0002$ per dollar. Find the threshold with an upper-tail probability of 0.005. (Answer: $26,491.59.)
A continuous uniform loss is modeled on $[200,800]$ dollars. Find the 80% VaR and state the probability above it. (Answer: VaR = $680.00; exceedance probability = 0.20.)

What You'll Learn