Selecting Measures with Sparse or Unreliable Data

Why Tail-Measure Estimates Need a Data Check

A calculated risk measure can look precise even when the data behind it are not. For example, an estimated 99% VaR reported to the nearest dollar may be based on only a handful of observations near the tail—or on a fitted distribution whose tail behavior is hard to determine from the sample. The arithmetic can be exact while the estimate remains uncertain.

This tutorial focuses on how the amount and quality of available data affect the choice and interpretation of VaR and TVaR. It does not make a single measure universally preferable. Instead, it provides checks for whether an estimate is supported by the observations and assumptions used to produce it. These checks are aligned with the CAS Exam 9 (Risk Management for Actuaries) syllabus.

Definition: For a loss random variable $X$ with distribution function $F_X$, the VaR at confidence level $\alpha$ is $ \operatorname{VaR}_\alpha(X)=\inf\{x:F_X(x)\geq\alpha\} $. The quantile-integral definition of TVaR is $ \operatorname{TVaR}_\alpha(X)=\frac{1}{1-\alpha}\int_\alpha^1\operatorname{VaR}_u(X)\,du $. Here $0<\alpha<1$. For a continuous loss distribution, TVaR can also be written as the expected loss conditional on exceeding the VaR threshold. For distributions with atoms, use the quantile-integral definition so that the probability mass at the boundary is handled correctly.

There are two related but distinct sources of uncertainty. Sampling uncertainty arises because a different sample from the same underlying process would produce different estimates. Model or parameter uncertainty arises because the available data may not determine the form of the tail or its parameters reliably. A larger sample can reduce sampling uncertainty, but it does not automatically resolve a misspecified model, incomplete reporting, or changes in the underlying risk.

What the Empirical Tail Actually Contains

Suppose $n$ observed losses are sorted as $X_{(1)}\leq X_{(2)}\leq\cdots\leq X_{(n)}$. The empirical distribution assigns probability $1/n$ to each observation. Under this convention, empirical VaR is the smallest observed loss for which the cumulative empirical probability is at least $\alpha$. The empirical TVaR is obtained by averaging the upper $1-\alpha$ fraction of the empirical distribution, including only the necessary fraction of a boundary observation when $n(1-\alpha)$ is not an integer.

This calculation is a description of the sample, not a guarantee that the sample represents the true tail. At a high confidence level, only a small number of observations may contribute to TVaR. At 99%, for example, 200 equally weighted observations contain only two observations in the upper 1% of the empirical distribution. The effective tail sample is therefore more informative than the overall sample size alone.

Worked Example: Counting the Observations Behind Empirical TVaR

A hypothetical insurer has 20 fully developed, comparable annual loss observations. The 18th, 19th, and 20th ordered losses are $ \$80{,}000$, $ \$120{,}000$, and $ \$300{,}000$, respectively. Using the empirical distribution, calculate VaR and TVaR at $\alpha=0.90$, and state how many observations contribute to the upper tail.

Method and assumptions. Treat the 20 losses as equally weighted observations from a stable population, with no adjustment for inflation or exposure differences. Use the empirical inverse-distribution quantile for VaR and the quantile-integral definition for TVaR.

Step 1: Find empirical VaR. The empirical cumulative probability at rank 18 is $18/20=0.90$. Therefore the smallest loss whose cumulative probability is at least 0.90 is $X_{(18)}$:

\widehat{\operatorname{VaR}}_{0.90}=X_{(18)}=\$80{,}000.

Step 2: Determine the tail weight. The upper-tail probability is $1-0.90=0.10$. In a sample of 20, this represents $20(0.10)=2$ equally weighted observations: the two largest losses.

Step 3: Average those losses.

\widehat{\operatorname{TVaR}}_{0.90} =\frac{\$120{,}000+\$300{,}000}{2} =\$210{,}000.

The sample TVaR is based on only two observations, even though the data set contains 20 losses. A common mistake is to interpret the $ \$210{,}000$ result as a stable estimate simply because it is calculated exactly. The correction is to report the empirical result together with its limited tail support and to assess whether the observations are sufficiently comparable and representative. As a check, the average of the two upper-tail observations is $ \$210{,}000$, between $ \$120{,}000$ and $ \$300{,}000$, as expected.

How Much Can the Sample Maximum Tell You?

A useful diagnostic is the chance that a sample even reaches a particular population quantile. Assume independent observations from a continuous distribution, and let $q_\alpha=\operatorname{VaR}_\alpha(X)$. Each observation has probability $\alpha$ of being below $q_\alpha$. Thus the probability that all $n$ observations are below $q_\alpha$ is $\alpha^n$. The complementary probability is the chance that the sample maximum reaches or exceeds the population quantile.

P\!\left(X_{(n)}\geq q_\alpha\right)=1-\alpha^n.

This is a statement about sample coverage, not a confidence interval for the quantile and not a direct estimate of its dollar value. It also relies on independent observations from the same continuous distribution. Dependence, changing exposure, or a changing loss process can make the calculation inappropriate.

Worked Example: Checking Support for a 99% Quantile

A hypothetical portfolio has 50 independent annual observations from a continuous, stable loss distribution. Calculate the probability that the sample maximum reaches or exceeds the population 99th-percentile loss. Then find the minimum sample size needed for that probability to be at least 95%.

Method and assumptions. Use the sample-maximum coverage formula, assuming independent observations and a fixed continuous distribution. The calculation measures whether the quantile is reached, not how accurately its dollar amount is estimated.

Step 1: Calculate coverage for 50 observations.

P\!\left(X_{(50)}\geq q_{0.99}\right) =1-(0.99)^{50} =1-0.6050 =0.3950.

Thus, even with 50 observations, the probability that the maximum reaches the population 99th percentile is only 0.3950.

Step 2: Solve for the required sample size. Require $1-(0.99)^n\geq0.95$. Rearranging gives $(0.99)^n\leq0.05$. Taking logarithms,

n\geq\frac{\ln(0.05)}{\ln(0.99)} =\frac{-2.995732}{-0.0100503} \approx298.07.

Since the sample size must be a whole number, the minimum is $n=299$. Checking the boundary, $1-(0.99)^{299}\approx0.9505$, whereas $n=298$ gives a probability just below 0.95. A common trap is to round 298.07 down. That fails the required probability threshold; always round a minimum sample size up. This result checks whether the sample reaches the quantile, not whether it contains enough information to estimate TVaR reliably.

Tail-Parameter Uncertainty Can Dominate the Arithmetic

A fitted tail model can produce a smooth numerical estimate even when different plausible parameter values lead to materially different results. The following comparison uses a Pareto model to make that sensitivity explicit. The model is an assumption, not a conclusion that the observations must follow a Pareto distribution.

Assume losses have a Pareto distribution with scale $x_m>0$ and shape $a>0$, with survival function $S_X(x)=(x_m/x)^a$ for $x\geq x_m$. For $0<\alpha<1$, solving $S_X(x)=1-\alpha$ gives $ \operatorname{VaR}_\alpha=x_m(1-\alpha)^{-1/a}$. When $a>1$, the mean excess above a threshold $q\geq x_m$ is $q/(a-1)$. Consequently, for a continuous Pareto loss, $ \operatorname{TVaR}_\alpha=\frac{a}{a-1}\operatorname{VaR}_\alpha$. If $a\leq1$, the model has infinite mean and finite TVaR is not available; the formula must not be used as if it yielded a finite value.

Why the Pareto TVaR formula holds: For $q=\operatorname{VaR}_\alpha$, the conditional survival function above $q$ is $P(X>x\mid X>q)=(q/x)^a$ for $x\geq q$. Integrating this survival function gives the mean excess $E[X-q\mid X>q]=\int_q^\infty(q/x)^a\,dx=q/(a-1)$, provided $a>1$. Adding $q$ gives $E[X\mid X>q]=q+q/(a-1)=aq/(a-1)$, which is the stated TVaR for this continuous model.

Worked Example: Comparing Plausible Pareto Tail Shapes

A hypothetical portfolio is modeled with Pareto scale $x_m=\$100{,}000$. Sparse observations do not distinguish confidently between shape $a=2$ and shape $a=3$. Calculate the 90% VaR and TVaR under each model, then compare the estimates. Round money to the nearest cent only at the final step.

Method and assumptions. Apply the Pareto formulas above, treating each shape as a separate candidate model. Both shapes exceed 1, so each model has finite TVaR. This is a sensitivity comparison, not evidence that either shape is correct.

For $a=2$: Since $1-\alpha=0.10$,

\operatorname{VaR}_{0.90} =\$100{,}000(0.10)^{-1/2} =\$100{,}000\sqrt{10} =\$316{,}227.7660\ldots \longrightarrow \$316{,}227.77.

The TVaR multiplier is $a/(a-1)=2$, so

\operatorname{TVaR}_{0.90} =2(\$316{,}227.7660\ldots) =\$632{,}455.5320\ldots \longrightarrow \$632{,}455.53.

For $a=3$:

\operatorname{VaR}_{0.90} =\$100{,}000(0.10)^{-1/3} =\$100{,}000(10)^{1/3} =\$215{,}443.4690\ldots \longrightarrow \$215{,}443.47.

The TVaR multiplier is $3/(3-1)=1.5$, so

\operatorname{TVaR}_{0.90} =1.5(\$215{,}443.4690\ldots) =\$323{,}165.2035\ldots \longrightarrow \$323{,}165.20.

The model with $a=2$ gives a VaR about $ \$100{,}784.30$ higher and a TVaR about $ \$309{,}290.33$ higher than the model with $a=3$. These differences arise solely from the shape assumption while the scale and confidence level are held fixed. A common mistake is to present one model’s cent-level result as though its tail parameter were known precisely. The correction is to show the sensitivity and explain that the data may not support choosing between these shapes. As checks, each TVaR exceeds its corresponding VaR, and the calculated dollar amounts are rounded from the unrounded formulas.

Choosing and Reporting a Measure Responsibly

Sparse or unreliable data do not make VaR or TVaR unusable, but they change what a reported estimate can reasonably claim. Before relying on a tail measure, ask what observations and assumptions support it. Useful checks include:

Count the effective tail observations. Translate the confidence level into an approximate tail count, $n(1-\alpha)$, for equally weighted data. A small count means the empirical tail measure is highly dependent on a few losses.
Check whether observations are comparable. Differences in exposure, policy terms, inflation, claim development, or reporting completeness can make raw historical losses a poor sample of the target loss distribution.
Separate empirical results from model-based results. State whether the estimate comes directly from ordered observations or from a fitted distribution. A model can extrapolate beyond observed losses, but that extrapolation depends on its assumptions.
Show sensitivity where tail assumptions are uncertain. Alternative plausible parameters or model forms can reveal whether the result is stable or driven by a particular tail specification.
Match reported precision to support. Extra decimal places do not reduce sampling or model uncertainty. Round consistently, and do not imply that the estimate is more certain than its inputs.

These checks help distinguish a precise calculation from a precise estimate. For example, calculating a quantile from a fully specified model may be straightforward, but uncertainty about the model’s tail parameter can remain substantial. Similarly, an empirical TVaR is exactly the mean of the selected sample tail, yet may still be a noisy estimate of future tail losses.

Key takeaway: Interpret VaR and TVaR alongside the data that support them. State the tail count, identify the distributional assumptions, and examine sensitivity when tail behavior is uncertain. Numerical precision is not evidence of statistical reliability.

Check Your Understanding

For each problem, state the key calculation and check whether the result is supported by the stated assumptions.

A hypothetical sample contains 40 equally weighted, comparable losses. The 38th, 39th, and 40th ordered values are $ \$45{,}000$, $ \$60{,}000$, and $ \$150{,}000$. Find empirical VaR and TVaR at 95%, and state the number of observations in the empirical tail. (Answer: VaR $=\$45{,}000$; TVaR $=\$105{,}000$; 2 tail observations.)
Assume a Pareto loss model with scale $ \$50{,}000$, shape $a=2$, and confidence level 95%. Find VaR and TVaR, rounding to cents. (Answer: VaR $=\$223{,}606.80$; TVaR $=\$447{,}213.60$.)
For 80 independent observations from a continuous, stable loss distribution, find the probability that the sample maximum reaches or exceeds the population 99th percentile. Give the probability to four decimals. (Answer: $1-(0.99)^{80}=0.5525$.)
Under independent observations from a continuous, stable distribution, find the minimum sample size giving at least a 99% probability that the sample maximum reaches or exceeds the population 95th percentile. (Answer: $n=90$.)
A sample has 100 equally weighted losses, with ordered values $X_{(98)}=\$80{,}000$, $X_{(99)}=\$100{,}000$, and $X_{(100)}=\$200{,}000$. Using the empirical quantile-integral definition, find VaR and TVaR at 97.5%. (Answer: VaR $=\$80{,}000$; TVaR $=\$136{,}000$.)

What You'll Learn