Integrated Risk-Measure Selection Case Workshop

Choosing a Measure Is a Decision, Not a Formula Hunt

A risk measure is useful only in relation to the decision it supports. A portfolio limit based on a percentile, a review triggered by severe tail losses, and a summary of average expected cost do not necessarily call for the same statistic. The choice becomes especially consequential when losses have discrete outcomes, dependence is uncertain, or the data provide little evidence about the tail.

This workshop uses hypothetical portfolios to compare the mean, Value at Risk (VaR), and Tail Value at Risk (TVaR). The aim is not to identify one measure that is best in every setting. It is to state what each measure captures, test its properties, and explain what decision its result would support. This approach is aligned with the CAS Exam 9 (Risk Management for Actuaries) syllabus.

Definition: Let $X$ denote a portfolio loss, with larger values representing worse outcomes. At confidence level $\alpha$, where $0<\alpha<1$, define $$ \operatorname{VaR}_{\alpha}(X)=\inf\{x:F_X(x)\geq \alpha\}. $$ This is the $\alpha$-quantile under the lower-quantile convention. Define quantile-integral TVaR by $$ \operatorname{TVaR}_{\alpha}(X)=\frac{1}{1-\alpha}\int_{\alpha}^{1}\operatorname{VaR}_{u}(X)\,du. $$ For discrete losses, the integral definition includes only the appropriate fraction of probability at the boundary loss; it is not generally the same as $E[X\mid X\geq \operatorname{VaR}_{\alpha}(X)]$.

The mean $E[X]$ summarizes average loss, not severity beyond a selected percentile. VaR identifies a threshold but does not report how large losses are once that threshold is exceeded. TVaR averages the quantile function over the upper tail, so it reflects the severity of outcomes above the VaR boundary. For discrete distributions, correctly handling probability mass at that boundary is essential.

Decision Criteria Before Calculation

Risk measures are commonly evaluated against four properties for loss variables. A measure $\rho$ is monotone if $X\leq Y$ in every outcome implies $\rho(X)\leq\rho(Y)$; translation invariant if adding a certain loss $c$ adds $c$ to the measure; positively homogeneous if scaling losses by $c\geq0$ scales the measure by $c$; and subadditive if $\rho(X+Y)\leq\rho(X)+\rho(Y)$. Subadditivity expresses that combining risks should not be assessed as requiring more risk than the sum of their separate assessments.

A measure satisfying all four properties is called coherent. For integrable losses, quantile-integral TVaR is coherent. VaR satisfies monotonicity, translation invariance, and positive homogeneity, but it is not subadditive for all loss distributions. The mean is coherent but may be insufficient when the decision specifically concerns severe losses. These are general properties, not guarantees that a particular estimate is reliable: data quality and model assumptions still matter.

Measure	What it emphasizes	Decision question it may help address	Important limitation
Mean loss	Average outcome	What loss level should be expected on average?	Can conceal a severe but infrequent tail
VaR at $\alpha$	Loss threshold at a percentile	What loss level is not exceeded in roughly an $\alpha$ fraction of outcomes?	Does not measure severity beyond the threshold; may fail subadditivity
TVaR at $\alpha$	Average of quantiles in the upper tail	How severe are losses represented by the upper $1-\alpha$ tail?	Can be sensitive to sparse, uncertain tail data

Discrete Tails: Include the Boundary Mass Correctly

Suppose $X$ has a finite or discrete distribution and $q=\operatorname{VaR}_{\alpha}(X)$. Then $F_X(q-)\leq\alpha\leq F_X(q)$, where $F_X(q-)=P(X<q)$. The probability strictly above $q$ is $1-F_X(q)$. To form a tail with total probability $1-\alpha$, TVaR uses all probability above $q$ and only $F_X(q)-\alpha$ probability at $q$.

\operatorname{TVaR}_{\alpha}(X) = \frac{E[X\,\mathbf{1}_{\{X>q\}}]+q\bigl(F_X(q)-\alpha\bigr)} {1-\alpha}.

This expression follows directly by splitting the quantile integral. For $u$ between $\alpha$ and $F_X(q)$, the quantile is $q$; for $u>F_X(q)$, the quantiles cover the distribution strictly above $q$. Thus the boundary contributes $q(F_X(q)-\alpha)$, while the remaining upper-tail contribution is $E[X\,\mathbf{1}_{\{X>q\}}]$. Their probabilities sum to $(F_X(q)-\alpha)+(1-F_X(q))=1-\alpha$, as required.

Worked Example: A Discrete Tail With a Material Boundary Mass

A hypothetical specialty insurer models annual portfolio loss $X$, in thousands of dollars, as 0 with probability 0.88, 40 with probability 0.08, and 200 with probability 0.04. Calculate the mean, 90% VaR, and 90% TVaR. A risk committee is deciding whether a percentile threshold alone is adequate for a severe-loss review.

Method and assumptions. Treat the stated probabilities as the full annual loss distribution, with no discounting or exposure adjustment. Use the lower-quantile definition and the quantile-integral definition of TVaR at $\alpha=0.90$.

Mean. Weight each loss by its probability:

$$ E[X]=0(0.88)+40(0.08)+200(0.04)=0+3.2+8=11.2. $$

The mean loss is 11.2 thousand dollars. For VaR, $F_X(0)=0.88<0.90$, while $F_X(40)=0.88+0.08=0.96\geq0.90$. Therefore $\operatorname{VaR}_{0.90}(X)=40$ thousand dollars.

For TVaR, $q=40$, $F_X(q)=0.96$, and the loss strictly above $q$ contributes $200(0.04)=8$ thousand dollars of probability-weighted loss. The boundary contribution is $40(0.96-0.90)=2.4$. Hence

\operatorname{TVaR}_{0.90}(X) =\frac{8+2.4}{1-0.90} =\frac{10.4}{0.10} =104 \quad\text{thousand dollars}.

Decision reading. VaR reports a 40-thousand-dollar threshold, whereas TVaR reports an average of 104 thousand dollars over the upper 10% tail. A review triggered only by VaR could fail to communicate how consequential the modeled 200-thousand-dollar outcome is.

Common mistake avoided. Conditioning on $X\geq40$ gives $(40(0.08)+200(0.04))/(0.08+0.04)=11.2/0.12$, or approximately 93.333 thousand dollars. That is not 90% TVaR: it includes all 8% probability at 40, although only 6% at the boundary belongs in a 10% tail. The correct TVaR is 104 thousand dollars. The check is that its tail contains 4% at 200 and 6% at 40, totaling 10%.

Dependence Can Change the Decision

Portfolio loss is often the sum $X+Y$ of component losses. Component distributions alone do not determine the distribution of that sum; the joint distribution, or dependence assumption, matters. The following hypothetical comparison holds both marginal distributions fixed and changes only the dependence. It illustrates why a risk-measure choice should be accompanied by a statement of the aggregation assumption.

Worked Example: Comparing Independent and Coincident Losses

Two hypothetical units each produce a loss of 100 thousand dollars with probability 0.06 and zero otherwise. Compare aggregate 95% VaR and TVaR if the units’ losses are independent versus perfectly coincident. State what changes for a portfolio limit based on the 95th percentile.

Method and assumptions. Losses are nonnegative and take only the stated values. In the independent case, multiply component probabilities. In the perfectly coincident case, both units have the same outcome in every scenario. Use $\alpha=0.95$.

For either unit, the loss is zero with probability 0.94. Under independence, the aggregate is zero with probability $0.94^2=0.8836$, 100 with probability $2(0.06)(0.94)=0.1128$, and 200 with probability $0.06^2=0.0036$. These probabilities sum to $0.8836+0.1128+0.0036=1.0000$.

The aggregate CDF at 100 is $0.8836+0.1128=0.9964$, while at zero it is $0.8836<0.95$. Thus independent aggregate VaR is 100 thousand dollars. For TVaR, the probability-weighted loss strictly above $q=100$ is $200(0.0036)=0.72$. The boundary contribution is $100(0.9964-0.95)=4.64$. Therefore

\operatorname{TVaR}_{0.95}(X+Y) =\frac{0.72+4.64}{0.05} =107.2 \quad\text{thousand dollars}.

Under perfect coincidence, the aggregate is zero with probability 0.94 and 200 with probability 0.06. Its 95% VaR is 200 thousand dollars, because the CDF at zero is below 0.95. The TVaR is also 200 thousand dollars: the worst 5% of outcomes all have loss 200.

Decision reading. The 95% VaR limit would be 100 thousand dollars under independence but 200 thousand dollars under perfect coincidence. The difference comes from dependence, not a change in either unit’s marginal loss distribution. The independent TVaR is 107.2 thousand dollars, also below the 200 thousand dollars under coincidence.

Common mistake avoided. Adding the two individual VaRs gives 200 thousand dollars because each unit’s 95% VaR is 100. That sum is not automatically the aggregate VaR: under independence the aggregate 95% VaR is 100. Conversely, using the independent result without testing dependence would understate the result in the perfectly coincident case. The calculation is conditional on each stated joint model.

Sparse Data: Separate an Estimate From a Conclusion

A tail statistic can be computed exactly from an empirical distribution and still be highly uncertain as an estimate of future risk. A small number of observations means that one severe loss can materially affect TVaR, while a percentile estimate can jump when a small number of observations move across its boundary. Equal exposure, a stable process, and representative observations are substantive assumptions, not consequences of performing the arithmetic.

Worked Example: Estimating Tail Measures From a Small Sample

A hypothetical risk team has 20 equally exposed annual observations, in thousands of dollars: 17 losses of 0, two losses of 50, and one loss of 300. Treat the empirical distribution as assigning probability $1/20$ to each year. Calculate its mean, 90% VaR, and 90% TVaR, then state one limitation relevant to a decision.

Method and assumptions. Each observation receives equal weight, so the empirical probabilities are 0.85 at zero, 0.10 at 50, and 0.05 at 300. This calculation describes the empirical distribution; it does not assume that 20 observations establish the true future tail.

The empirical mean is

E[X]=\frac{17(0)+2(50)+1(300)}{20} =\frac{400}{20}=20 \quad\text{thousand dollars}.

The cumulative probability is 0.85 at zero and 0.95 at 50, so $\operatorname{VaR}_{0.90}(X)=50$ thousand dollars. For TVaR, 0.05 probability lies strictly above 50, contributing $300(0.05)=15$. The boundary contribution is $50(0.95-0.90)=2.5$. Thus

\operatorname{TVaR}_{0.90}(X) =\frac{15+2.5}{0.10} =175 \quad\text{thousand dollars}.

Decision reading. The empirical TVaR is much larger than the mean because the top 10% tail consists of half an observation’s probability at 300 and half at 50. But the result relies heavily on one observed 300-thousand-dollar loss. A decision-maker should treat 175 thousand dollars as an estimate based on a short sample, not as a precise, stable future-tail value.

Common mistake avoided. Averaging only the observed losses at or above 50 gives $(2(50)+300)/3=133.333$ thousand dollars. This conditions on the full 15% of observations at or above 50, not the worst 10%. The quantile-integral calculation uses only half of the empirical mass at 50, giving the correct value of 175 thousand dollars. The selected tail probability is $0.05+0.05=0.10$.

Defending a Measure Choice

A defensible recommendation identifies both the objective and the limits of the evidence. For a percentile-based operating threshold, VaR may communicate a clear loss level, provided the percentile and aggregation model are explicit. For a decision that must account for severity beyond a threshold, TVaR provides more tail information and has the coherence advantage of subadditivity. For planning around average cost, the mean may be relevant, but it should not be presented as a substitute for a tail measure when severe outcomes drive the decision.

When dependence is uncertain, calculate results under plausible joint assumptions or stress cases rather than treating a single correlation estimate as definitive. When data are sparse, report the empirical result together with the observations and assumptions that drive it. A risk committee might then set a monitoring trigger, request additional analysis, or compare alternative portfolio actions; the selected statistic informs those choices but does not make them automatically.

A common exam trap is to select a measure solely because it is coherent, or solely because it produces the largest number. Coherence is valuable for consistent aggregation, but it does not resolve uncertainty in a fitted tail. Likewise, a larger estimate is not necessarily more appropriate if the decision concerns a percentile threshold rather than average tail severity. State what is being measured, the relevant assumptions, and the consequence of choosing that measure.

Key takeaway: Choose a risk measure by matching its mathematical properties and tail interpretation to the decision. For discrete losses, calculate boundary mass correctly; for portfolios, disclose dependence assumptions; and for limited data, distinguish the computed estimate from the strength of the evidence.

Check Your Understanding

For each hypothetical problem, use the stated lower-quantile and quantile-integral conventions. Amounts are in thousands of dollars unless noted.

A loss is 0 with probability 0.91, 25 with probability 0.06, and 125 with probability 0.03. Find 90% VaR and TVaR. (Answer: VaR = 0; TVaR = 52.5)
Two independent units each lose 80 with probability 0.04 and zero otherwise. Find aggregate 95% VaR and TVaR. (Answer: VaR = 80; TVaR = 82.56)
An empirical distribution has 10 equally weighted observations: six losses of 0, three of 40, and one of 200. Find 80% VaR and TVaR. (Answer: VaR = 40; TVaR = 120)
Explain why a portfolio’s 95% VaR cannot be inferred from its two component 95% VaRs alone, even if each component’s marginal distribution is known. (Answer: Aggregate VaR depends on the joint distribution or dependence; the same marginals can yield different aggregate percentiles.)
A short sample produces a high TVaR because one observation is exceptionally large. Give one reason to retain the result and one qualification to report before using it for a decision. (Answer: Retain it because TVaR reflects severe tail outcomes; qualify it because the sparse sample makes the tail estimate sensitive and uncertain.)

What You'll Learn