Interpreting VaR Confidence Levels

What a VaR Confidence Level Says

A VaR statement combines two distinct pieces of information: a loss threshold and a probability level. For example, “the 99% VaR is $4 million” identifies $4 million as the threshold and 99% as the confidence level. Interpreting the statement correctly means identifying what probability lies at or below that threshold, and what probability lies above it.

Throughout, let $X$ denote a nonnegative loss over one specified measurement period, and let $F_X(x)=P(X\le x)$ be its cumulative distribution function. The confidence level is denoted by $\alpha$, where $0<\alpha<1$. The corresponding tail probability is $1-\alpha$. A VaR statement does not, by itself, describe the size of losses beyond the threshold.

Definition: At confidence level $\alpha$, the Value at Risk is the lower $\alpha$-quantile of the loss distribution: $$ \operatorname{VaR}_{\alpha}(X)=\inf\{x:F_X(x)\ge\alpha\}. $$ The associated tail probability is $1-\alpha$. The quantile definition implies $P(X\le \operatorname{VaR}_{\alpha}(X))\ge\alpha$ and $P(X>\operatorname{VaR}_{\alpha}(X))\le 1-\alpha$. For a continuous distribution with a strictly increasing CDF at the quantile, these probabilities are exactly $\alpha$ and $1-\alpha$, respectively.

The inequality matters. For a distribution with probability concentrated at particular loss amounts, the CDF may jump over $\alpha$. In that case the probability at or below VaR can be greater than the stated confidence level, and the probability strictly above VaR can be less than the stated tail probability. The probability of a loss equal to VaR is part of $P(X\le\operatorname{VaR}_{\alpha}(X))$, not part of the strict exceedance probability $P(X>\operatorname{VaR}_{\alpha}(X))$.

Worked Example: Reading a VaR From a Discrete Loss Distribution

A hypothetical insurer models one-year losses $X$, in thousands of dollars, with the following distribution: $P(X=0)=0.80$, $P(X=100)=0.12$, $P(X=300)=0.06$, and $P(X=900)=0.02$. Find the 95% VaR, then state the probability of a loss at or below that threshold and the probability of a strict exceedance.

Method. Accumulate the probabilities in increasing loss order and find the first loss amount whose cumulative probability is at least 0.95. The probabilities sum to $0.80+0.12+0.06+0.02=1.00$, so the table defines a complete distribution.

At $X=0$, the cumulative probability is $0.80$. At $X=100$, it is $0.80+0.12=0.92$. At $X=300$, it is $0.92+0.06=0.98$. Since $0.92<0.95\le0.98$, the first loss amount reaching 0.95 is 300 thousand dollars.

\operatorname{VaR}_{0.95}(X)=\$300{,}000.

The probability at or below the threshold is $P(X\le300)=0.98$, or 98.00%. The strict exceedance probability is $P(X>300)=P(X=900)=0.02$, or 2.00%. The probability of a loss equal to the threshold is $P(X=300)=0.06$, or 6.00%; therefore $P(X\ge300)=0.06+0.02=0.08$, or 8.00%.

Check. The cumulative probability at $300,000 is at least 95%, while the cumulative probability at the next lower possible loss, $100,000, is only 92%. Thus $300,000 is the 95% quantile. The common mistake to avoid is treating the 5% tail probability as $P(X\ge300)$; the VaR tail claim concerns $P(X>300)$, which here is only 2%.

Why the Tail Probability Is the Complement

The probability statements follow directly from the quantile definition and the complement rule. Let $q=\operatorname{VaR}_{\alpha}(X)$. By the definition of the lower quantile, the cumulative probability at $q$ has reached or passed $\alpha$, so $F_X(q)\ge\alpha$. Since $F_X(q)=P(X\le q)$, the probability of a strict exceedance is:

P(X>q)=1-P(X\le q)=1-F_X(q)\le1-\alpha.

This establishes the precise general claim: at least an $\alpha$ proportion of the modeled probability is at or below VaR, and no more than $1-\alpha$ is strictly above it. If the CDF is continuous at $q$ and $F_X(q)=\alpha$, the inequalities become equalities. If the CDF jumps at $q$, the probability at or below the threshold can exceed $\alpha$, as in the first example.

The quantile is also the smallest threshold at which the cumulative probability reaches the selected confidence level. That is why a discrete distribution can have a larger-than-$\alpha$ probability at or below VaR: moving to the next lower possible loss would leave the cumulative probability below $\alpha$, while moving to the quantile includes the entire probability mass at that loss.

Worked Example: Translating a Continuous-Model VaR Into Probabilities

A hypothetical portfolio’s one-year loss $X$ is modeled as uniform on the interval from $0 to $500,000. Find its 99% VaR and state both the probability of exceeding that amount and the probability of exceeding a separate $450,000 threshold.

Method. Under this model, $X\sim\operatorname{Uniform}(0,500{,}000)$, so for $0\le x\le500{,}000$, $F_X(x)=x/500{,}000$. The model is continuous, and its CDF increases throughout this interval.

At the 99% quantile, set the CDF equal to 0.99 and solve:

\frac{q}{500{,}000}=0.99 \qquad\Longrightarrow\qquad q=0.99(500{,}000)=\$495{,}000.

Thus $\operatorname{VaR}_{0.99}(X)=\$495{,}000$. Because the distribution is continuous, $P(X\le495{,}000)=0.9900$, and $P(X>495{,}000)=1-0.9900=0.0100$, or 1.00%.

For the separate threshold of $450,000, the cumulative probability is $450{,}000/500{,}000=0.9000$. Therefore $P(X>450{,}000)=1-0.9000=0.1000$, or 10.00%. This threshold is the 90% quantile under the stated model, not the 99% VaR.

Check. The computed quantile is 99% of the interval’s upper endpoint, and the remaining 1% of the uniform interval lies above it. The separate threshold is 90% of that endpoint, leaving 10% above. The common mistake to avoid is attaching the 1% tail probability to every threshold discussed; the tail probability corresponds to the particular VaR confidence level, not to an arbitrary loss amount.

Interpreting Thresholds When There Is a Probability Mass

For discrete losses, keep three events separate: $X<q$, $X=q$, and $X>q$. The confidence claim uses $X\le q$, while the usual exceedance claim uses $X>q$. If a question instead asks for the probability of reaching or exceeding the threshold, calculate $P(X\ge q)$ and include the probability mass at $q$. These quantities need not be close when a substantial probability is concentrated exactly at the VaR amount.

A probability statement about an individual threshold can also be expressed in reverse. If a model gives $P(X\le c)=0.96$, then $c$ is a 96th-percentile threshold when it is the lower 96% quantile, and $P(X>c)=0.04$. In a discrete model, merely knowing $P(X\le c)=0.96$ does not always establish that $c$ is the 96% VaR: one must also check that no smaller loss threshold already has cumulative probability of at least 0.96.

Worked Example: Interpreting Two Confidence Levels for a Hypothetical Cyber-Loss Model

A hypothetical organization models its one-year cyber loss $X$ as taking values $0, $1.2 million, $6 million, or $20 million, with probabilities 0.92, 0.05, 0.02, and 0.01, respectively. Find the 95% and 99% VaR amounts and interpret the strict exceedance probabilities at each amount.

Method. Check that the probabilities sum to one: $0.92+0.05+0.02+0.01=1.00$. Then calculate the cumulative probabilities in increasing loss order. At $0, the CDF is 0.92. At $1.2 million, it is $0.92+0.05=0.97$. At $6 million, it is $0.97+0.02=0.99$. At $20 million, it is 1.00.

For the 95% VaR, the cumulative probability at $0 is below 0.95, while the cumulative probability at $1.2 million is 0.97. Hence $\operatorname{VaR}_{0.95}(X)=\$1.2$ million. The probability at or below this amount is 0.97, or 97.00%, and the strict exceedance probability is $0.02+0.01=0.03$, or 3.00%.

For the 99% VaR, the cumulative probability at $1.2 million is only 0.97, while the cumulative probability at $6 million is 0.99. Thus $\operatorname{VaR}_{0.99}(X)=\$6$ million. The probability at or below this amount is 0.99, or 99.00%, and the strict exceedance probability is 0.01, or 1.00%.

Check. Each VaR is the smallest listed loss amount whose cumulative probability reaches its stated confidence level. At the 95% VaR, the probability of a loss at least as large as $1.2 million is $0.05+0.02+0.01=0.08$, or 8.00%, because this includes equality. At the 99% VaR, the probability of a loss at least as large as $6 million is $0.02+0.01=0.03$, or 3.00%. These are not the strict exceedance probabilities. The common mistake to avoid is reading “95% VaR” as implying exactly a 5% chance of a loss at least as large as the threshold.

Common Interpretation Traps

Reversing confidence and tail probability. A 99% confidence level has a 1% complementary tail probability, not a 99% tail probability.
Calling the tail probability exact in every model. In general, $P(X>\operatorname{VaR}_{\alpha}(X))\le1-\alpha$. Equality is not guaranteed when the loss distribution has jumps.
Using “at least” when the question asks for strict exceedance. $P(X\ge q)$ includes losses equal to $q$; $P(X>q)$ does not.
Interpreting VaR as a maximum possible loss. Losses can exceed VaR. The probability statement describes how much probability lies above or at the threshold, not an upper bound on severity.
Turning a one-period probability into a multi-period claim without assumptions. A one-year exceedance probability does not alone determine the chance of one or more exceedances over several years. Such a calculation would require a specified multi-period model.
Assuming VaR describes the size of tail losses. A VaR statement gives a quantile threshold and a probability claim; it does not specify how large losses are when they exceed that threshold.

Key takeaway: For $q=\operatorname{VaR}_{\alpha}(X)$, at least $\alpha$ probability lies at or below $q$, and at most $1-\alpha$ lies strictly above $q$. Use equality only when the distributional conditions support it, and distinguish strict exceedance from meeting or exceeding the threshold. This interpretation is aligned with the CAS Exam 9 (Risk Management for Actuaries) syllabus.

Check Your Understanding

For each problem, identify the confidence level or threshold probability and be careful about whether equality at the threshold is included.

A hypothetical annual loss is uniform from $0 to $2,000,000. Find the 97.5% VaR and its strict exceedance probability. (Answer: VaR = $1,950,000; strict exceedance probability = 0.0250.)
A discrete loss $X$, in thousands of dollars, equals 0 with probability 0.85, 50 with probability 0.10, and 200 with probability 0.05. Find the 90% VaR, $P(X>\operatorname{VaR}_{0.90}(X))$, and $P(X\ge\operatorname{VaR}_{0.90}(X))$. (Answer: VaR = $50,000; strict exceedance probability = 0.0500; probability of meeting or exceeding VaR = 0.1500.)
For a continuous modeled loss, a threshold $c=\$4$ million satisfies $P(X\le c)=0.9600$. State the confidence level represented by $c$ and the probability of a strict exceedance. (Answer: 96% confidence; strict exceedance probability = 0.0400.)
A discrete annual loss equals $0 with probability 0.90, $3 million with probability 0.08, and $12 million with probability 0.02. Find the 95% VaR, then distinguish the probabilities of a strict exceedance and of a loss at least as large as VaR. (Answer: VaR = $3 million; strict exceedance probability = 0.0200; probability of a loss at least as large as VaR = 0.1000.)
A report states that a hypothetical model’s 99.5% VaR is $8 million. Give the general probability bounds implied by this statement, without assuming a continuous loss distribution. (Answer: $P(X\le \$8\text{ million})\ge0.9950$ and $P(X>\$8\text{ million})\le0.0050$.)

What You'll Learn