Tutorials › CAS Exam 9 - Risk Management for Actuaries › Common Interpretation Errors In TVaR Calculations

Home  /  Tutorials  /  CAS Exam 9 / Tail Value at Risk

CAS Exam 9 · Tail Value at Risk · Tutorial 273 of 995

Common Interpretation Errors In TVaR Calculations

Distinguish the TVaR loss convention, the correct tail, the observations that belong in the average, and the units of the result.

Intermediate 10 min read

What You'll Learn

  • Define TVaR using a loss variable and a stated confidence level.
  • Convert a gain or profit convention into the corresponding loss convention.
  • Distinguish the confidence level from the probability in the upper tail.
  • Average the correct tail, including the appropriate fraction of probability mass at a cutoff.
  • Explain why TVaR is a loss amount, not a probability.
  • Check interpretations using units, tail direction, and the loss distribution.

Why TVaR Interpretation Requires Care

A TVaR calculation can be numerically correct and still be interpreted incorrectly. Common causes include treating a profit as a loss, using the tail probability as the confidence level, averaging the wrong observations, or reporting a probability where the result should be a loss amount. These mistakes matter because TVaR is used to describe the severity of losses in a specified part of a distribution; the sign convention and the selected tail determine what that statement means.

Throughout this tutorial, \(X\) denotes a loss, so larger values of \(X\) mean worse outcomes. The confidence level is \(\alpha\), with \(0<\alpha<1\), and the upper-tail probability is \(1-\alpha\). For a distribution that may be discrete or have atoms, define the quantile by \( \operatorname{VaR}_p(X)=\inf\{x:F_X(x)\ge p\}\). The quantile-integral definition of TVaR handles both continuous and discrete distributions.

Definition: For an integrable loss \(X\), the upper-tail TVaR at confidence level \(\alpha\) is $$ \operatorname{TVaR}_{\alpha}(X)=\frac{1}{1-\alpha}\int_{\alpha}^{1}\operatorname{VaR}_{p}(X)\,dp. $$ It is expressed in the same units as \(X\), such as dollars. The probability \(1-\alpha\) specifies the size of the tail being summarized; it is not the value of TVaR.

For a continuous loss distribution, this definition agrees with the conditional mean of losses above the quantile: \(E[X\mid X>\operatorname{VaR}_{\alpha}(X)]\). For a discrete distribution or one with an atom at the quantile, that conditional mean need not agree. The quantile integral includes just the amount of probability mass needed at the cutoff to fill the upper tail of size \(1-\alpha\).

Deriving the Continuous-Tail Interpretation

The conditional-mean interpretation for a continuous distribution follows from a probability transformation. Suppose \(X\) has a continuous distribution and finite expectation, and let \(U=F_X(X)\). Then \(U\) is uniform on \((0,1)\), and the loss associated with percentile \(p\) is \(\operatorname{VaR}_p(X)\). The event \(U>\alpha\) corresponds to \(X>\operatorname{VaR}_{\alpha}(X)\), apart from events of probability zero. Therefore,

$$ \int_{\alpha}^{1}\operatorname{VaR}_{p}(X)\,dp =E[X\,\mathbf{1}_{\{X>\operatorname{VaR}_{\alpha}(X)\}}] =(1-\alpha)E[X\mid X>\operatorname{VaR}_{\alpha}(X)]. $$

Dividing by \(1-\alpha\) gives the conditional-mean formula. Continuity matters: it ensures that the probability strictly above the quantile is \(1-\alpha\). With an atom at the quantile, the probability strictly above it may be smaller than \(1-\alpha\); averaging only those strictly larger losses then omits part of the required tail.

Worked Example: Converting Gains to Losses

Worked Example: Converting Gains to Losses

A hypothetical portfolio has an annual net result \(G\), measured in millions of dollars, of \(+8\) with probability \(0.95\) and \(-12\) with probability \(0.05\). Positive \(G\) is a gain and negative \(G\) is a loss. Find the TVaR at \(\alpha=0.90\), using \(X=-G\) as the loss variable. State the method and avoid confusing the best outcomes with the worst losses.

Method: define loss as the negative of the net result, then integrate the upper 10% of the loss distribution. The loss outcomes are \(X=-8\) with probability \(0.95\) and \(X=12\) with probability \(0.05\). The relevant tail probability is \(1-\alpha=0.10\).

The upper 10% consists of all the 5% outcome at loss \(12\), plus 5% of the probability mass at loss \(-8\). Equivalently, from the quantile definition, \(\operatorname{VaR}_p(X)=-8\) for \(0.90<p\le 0.95\), and it is \(12\) for \(0.95<p\le1\). Thus,

$$ \operatorname{TVaR}_{0.90}(X) =\frac{0.05(-8)+0.05(12)}{0.10} =\frac{-0.40+0.60}{0.10} =2 $$

The answer is a TVaR of \(2\) million dollars. The tail includes both a gain outcome, represented by a negative loss, and the severe loss outcome. A common mistake is to average the upper 10% of \(G\), which selects the \(+8\) gain outcomes and answers a different question. The correction is to convert to \(X=-G\) and select the upper tail of \(X\). As a check, the weighted tail total is \(0.20\) million probability-weighted dollars; dividing by tail probability \(0.10\) gives \(2\) million dollars.

Worked Example: Using the Confidence Level, Not the Tail Probability

Worked Example: Using the Confidence Level, Not the Tail Probability

A hypothetical annual loss \(X\), measured in thousands of dollars, is uniformly distributed from \(20\) to \(80\). Find its 90% TVaR. Then identify the result of mistakenly using the 10% tail probability as the confidence level. State the method and the error it avoids.

Method: find the quantile at the stated confidence level and average the loss values from that quantile to the maximum. For a uniform distribution on \([20,80]\), the cumulative distribution function is \(F_X(x)=(x-20)/60\) on that interval. At \(p=0.90\),

$$ \operatorname{VaR}_{0.90}(X)=20+0.90(80-20)=20+54=74. $$

The upper 10% runs from \(74\) to \(80\). A uniform distribution has a constant density, so its average loss over this interval is its midpoint:

$$ \operatorname{TVaR}_{0.90}(X)=\frac{74+80}{2}=77 $$

The correct result is \(77\) thousand dollars. If \(0.10\) is mistakenly used as the confidence level, the resulting lower-tail calculation is \(\operatorname{VaR}_{0.10}=20+0.10(60)=26\), followed by an average from \(26\) to \(80\) of \(53\) thousand dollars. That is TVaR at 10% confidence, not the requested 90% TVaR. The check is that a 90% upper-tail measure must use the highest 10% of losses, not the highest 90%.

Worked Example: Averaging the Correct Mass at a Cutoff

Worked Example: Averaging the Correct Mass at a Cutoff

A hypothetical set of ten equally likely annual losses, in thousands of dollars, is \(1,2,3,4,5,6,7,9,9,20\). Calculate empirical TVaR at \(\alpha=0.75\) using the quantile-integral definition. Explain why averaging only losses strictly above the VaR is wrong. State the method and the tie-related trap.

Method: represent the empirical distribution with probability \(0.10\) at each observation, then take the upper \(25\%\) of probability mass. Its 75th percentile is \(9\): the cumulative probability is \(0.70\) through loss \(7\), and \(0.90\) through loss \(9\). Thus, the quantile is \(9\) over the percentile interval from \(0.75\) to \(0.90\), of width \(0.15\), and is \(20\) from \(0.90\) to \(1\), of width \(0.10\). The total tail width is \(0.25\).

$$ \operatorname{TVaR}_{0.75}(X) =\frac{0.15(9)+0.10(20)}{0.25} =\frac{1.35+2.00}{0.25} =13.4 $$

The result is \(13.4\) thousand dollars. The top 25% of probability mass consists of all the mass at \(20\) and only \(0.15\) of the \(0.20\) mass tied at \(9\). Averaging only losses strictly above the quantile would give \(20\), using just 10% of probability mass and omitting the required portion at \(9\). Averaging all losses at or above the quantile would give \((9+9+20)/3=12.6667\), using too much mass at \(9\). The quantile integral avoids both errors by taking exactly the required tail mass. As a check, the weighted tail sum is \(3.35\), and \(3.35/0.25=13.4\).

Worked Example: Distinguishing a Probability from a Loss Amount

Worked Example: Distinguishing a Probability from a Loss Amount

A hypothetical annual loss \(X\), in thousands of dollars, is uniformly distributed from \(0\) to \(500\). At confidence level \(\alpha=0.98\), calculate VaR and TVaR, and report the probability of exceeding VaR. State the units of each answer and avoid labeling the tail probability as TVaR.

Method: calculate the 98th-percentile loss, average the upper 2% of the uniform distribution, and calculate the exceedance probability separately. The VaR is

$$ \operatorname{VaR}_{0.98}(X)=0+0.98(500)=490 $$

The upper 2% covers losses from \(490\) to \(500\), whose average is

$$ \operatorname{TVaR}_{0.98}(X)=\frac{490+500}{2}=495. $$

The probability of exceeding \(490\) is \(1-0.98=0.02\). Therefore, VaR is \(490\) thousand dollars, TVaR is \(495\) thousand dollars, and the exceedance probability is \(0.0200\), or 2%. TVaR is not “2%”: it is a loss amount. The probability answers how often losses exceed VaR, while TVaR describes the average loss in the specified upper tail. As a check, the interval from \(490\) to \(500\) is exactly 2% of the uniform distribution and has midpoint \(495\).

Common Interpretation Traps

A reliable reading of a TVaR result starts by stating the loss variable, confidence level, and units. Then check the tail and, for a discrete distribution, how much probability mass is included at its cutoff. In particular:

  • Reversing gains and losses: if the modeled variable is profit, the upper tail is usually the best outcomes. For a loss-based TVaR, define loss consistently, often as the negative of profit.
  • Swapping confidence and tail probability: at confidence level \(\alpha\), the tail size is \(1-\alpha\). A 95% TVaR concerns the upper 5%, not the upper 95%.
  • Averaging the wrong observations: sort by loss, not by record order or probability of occurrence. With a cutoff atom, include only enough mass at the cutoff to make the tail probability \(1-\alpha\).
  • Reporting the tail probability as the measure: \(1-\alpha\) is a probability. TVaR has the units of loss, such as dollars or thousands of dollars.
Key takeaway: TVaR at level \(\alpha\) is an upper-tail loss measure based on a tail of probability \(1-\alpha\). Confirm the sign convention, use the correct tail, average exactly the required tail mass, and report the result in loss units.

Check Your Understanding

For each problem, use the stated loss convention and confidence level. Give the requested amount and, where applicable, the probability as a separate quantity.

  1. A hypothetical loss is uniform on \([0,120]\) thousand dollars. Find \(\operatorname{VaR}_{0.95}\) and \(\operatorname{TVaR}_{0.95}\). (Answer: VaR \(=114\) thousand dollars; TVaR \(=117\) thousand dollars.)
  2. A hypothetical empirical distribution has four equally likely losses of \(2,4,9,15\) thousand dollars. Find TVaR at \(\alpha=0.75\), and state how many observations' probability mass is in the upper tail. (Answer: \(15\) thousand dollars; one observation's mass.)
  3. A portfolio's net result \(G\), in thousands of dollars, is \(+6\) with probability \(0.90\) and \(-24\) with probability \(0.10\). Define loss by \(X=-G\). Find \(\operatorname{TVaR}_{0.90}(X)\). (Answer: \(24\) thousand dollars.)
  4. Eight equally likely losses, in thousands of dollars, are \(3,5,6,8,8,12,20,40\). Find empirical TVaR at \(\alpha=0.80\), using the quantile-integral definition and fractional probability mass at the cutoff if needed. (Answer: \(32.5\) thousand dollars.)
  5. A hypothetical loss is uniform from \(100\) to \(300\) dollars. At \(\alpha=0.95\), find TVaR and the probability of exceeding VaR. State which answer is a probability. (Answer: TVaR \(=295\) dollars; exceedance probability \(=0.0500\), or 5%.)

Previous

Next

Back to Tutorials
Clinical Biostats