Interpreting Extreme Claims in Empirical Tail Estimates

Why Extreme Claims Need Careful Interpretation

A small number of unusually large claims can have a modest effect on an overall sample mean but a substantial effect on a tail estimate. This is not a calculation error: a tail estimate deliberately focuses on the largest observations. The important question is how many observations support that estimate. If a tail average is based on only two claims, changing one of them can materially change the result.

The calculations below describe hypothetical claim samples. Unless stated otherwise, interpret the claims as uncensored observations measured in the same units from a representative, independent sample of claims from one severity distribution. The empirical summaries can be calculated without specifying a parametric distribution; interpreting them as evidence about future claims requires the sample to be relevant to the future portfolio.

Definition: For observed claims $X_1,\ldots,X_n$, the empirical distribution function at $x$ is $\hat F_X(x)=n^{-1}\sum_{j=1}^n I(X_j\le x)$. At a fixed threshold $t$, the empirical exceedance probability is $\hat S_X(t)=n^{-1}\sum_{j=1}^n I(X_j>t)$. If $r=\sum_{j=1}^n I(X_j>t)>0$, the empirical conditional tail mean above $t$ is $\sum_{j:X_j>t}X_j/r$. If $r=0$, that conditional mean is undefined: the sample contains no claims above the threshold.

The strict inequality matters. A claim equal to $t$ is not counted as an exceedance of $t$. Also, the empirical exceedance probability uses all $n$ claims as its denominator, while the conditional tail mean uses only the $r$ exceedances. These quantities answer different questions: the first estimates how often a claim exceeds the threshold; the second describes the average size among the claims that do.

Fixed-Threshold Tail Estimates

At a fixed threshold, one unusually large claim changes the exceedance probability only if it changes whether that claim is above the threshold. But if it remains an exceedance, it can still move the conditional tail mean. The tail mean is an ordinary average of the exceedance amounts, so each exceedance has weight $1/r$. Its additive contribution is its amount divided by $r$; its share of the total mean is its amount divided by the sum of all exceedance amounts.

Worked Example: One Claim and a Fixed-Threshold Tail Mean

A hypothetical insurer records eight uncensored claims, in thousands of dollars: 12, 15, 18, 20, 22, 25, 30, and 78. Calculate the empirical probability of exceeding 25, the conditional mean claim above 25, and the effect of replacing 78 by 38. Use a strict exceedance threshold.

Method. Count claims strictly greater than 25, then average those claims. The observations above 25 are 30 and 78, so $n=8$ and $r=2$.

\hat S_X(25)=\frac{2}{8}=0.25,\qquad \widehat{E[X\mid X>25]}=\frac{30+78}{2}=\frac{108}{2}=54.

Thus, one quarter of this sample exceeds 25 thousand dollars, and the average exceedance is 54 thousand dollars. In the tail mean, the claim of 30 contributes $30/2=15$ thousand dollars to the average, and the claim of 78 contributes $78/2=39$ thousand dollars. Their shares of the resulting mean are $30/108\approx 27.78\%$ and $78/108\approx 72.22\%$, respectively.

Now replace 78 by 38. There are still two exceedances, so the empirical exceedance probability remains $2/8=0.25$. The revised conditional mean is $(30+38)/2=68/2=34$ thousand dollars. The change is $34-54=-20$ thousand dollars, which also equals the change in the affected observation, $(38-78)/2=-40/2=-20$.

Check. The revised exceedance amounts, 30 and 38, average to 34; both remain strictly above 25. The probability therefore stays fixed while the tail mean changes. This avoids the mistake of assuming that a fixed exceedance count means the tail estimate is unchanged.

A large claim can affect more than the tail mean. In the example, replacing 78 by 38 also reduces the overall sample total by 40 thousand dollars, so the overall sample mean falls by $40/8=5$ thousand dollars. The same claim has a larger effect on the tail mean because that average is based on two observations rather than eight.

Upper-Tail Averages and Fractional Observations

A fixed-threshold conditional mean depends on the chosen threshold and on how many claims exceed it. A related summary targets a specified fraction of the empirical distribution. For an upper-tail fraction $1-p$, the empirical upper-tail average can assign fractional weight to the observation at the boundary when the desired tail contains a non-integer number of claims. This makes the tail mass exactly $1-p$, rather than rounding the number of included observations up or down.

Definition: Let $X_{(1)}\le \cdots\le X_{(n)}$ be the ordered observations, and set $m=n(1-p)$. Write $m=k+\alpha$, where $k$ is the integer part of $m$ and $0\le\alpha<1$. For $m>0$, the empirical average of the upper $1-p$ fraction is $$ \widehat{\operatorname{TVaR}}_p= \frac{\displaystyle\sum_{j=n-k+1}^{n}X_{(j)}+\alpha X_{(n-k)}}{m}, $$ with the sum taken as zero when $k=0$. Thus the $k$ largest observations receive full weight and the next observation receives weight $\alpha$. This is the empirical integrated-quantile convention; it need not equal the simple conditional mean above a threshold.

This formula averages a total empirical mass of $m/n=1-p$. If $m$ is an integer, it is just the average of the $m$ largest observations. If $m$ is not an integer, the fractional boundary weight prevents the result from depending on an arbitrary rounding choice.

Worked Example: A Fractional Upper-Tail Average

A hypothetical sample of six claims, in thousands of dollars, is 10, 12, 14, 16, 20, and 60. Calculate the empirical average of the upper 25% of the sample, using the fractional-observation convention. Then recalculate it if the largest claim is 120 instead of 60.

Method. The target tail contains $m=6(1-0.75)=1.5$ observations. Thus $k=1$ and $\alpha=0.5$: include the largest observation fully and half of the next largest.

\widehat{\operatorname{TVaR}}_{0.75} =\frac{60+0.5(20)}{1.5} =\frac{60+10}{1.5} =\frac{70}{1.5} =46.666\ldots

The estimate is 46.67 thousand dollars, rounded to two decimals. With the largest claim changed to 120, the weighted numerator is $120+0.5(20)=130$, and the denominator remains 1.5. The revised estimate is $130/1.5=86.666\ldots$, or 86.67 thousand dollars to two decimals. The increase is 40 thousand dollars, equal to the claim increase of 60 divided by the tail weight of 1.5.

Check. In each calculation, the weights sum to $1+0.5=1.5$ observations, as required. Averaging only the single largest claim would give 60 thousand dollars initially, but that would represent the upper $1/6$, not the requested upper $1/4$. The fractional boundary observation is essential here.

Why Few Tail Observations Are Unstable

An empirical tail probability is a sample proportion. If claims are independent and identically distributed and the true probability of exceeding a fixed threshold is $q$, each exceedance indicator is Bernoulli with variance $q(1-q)$. The variance of their average is therefore $q(1-q)/n$. When $q$ is small, the expected number of exceedances, $nq$, may also be small. A single claim can then represent a large share of the evidence about the tail.

\operatorname{Var}\!\left(\hat S_X(t)\right)=\frac{q(1-q)}{n}, \qquad \frac{\operatorname{SE}(\hat S_X(t))}{q} =\sqrt{\frac{1-q}{nq}}.

The second expression is the relative standard error when $q>0$. It highlights why a small tail probability is hard to estimate precisely from a modest sample: precision depends strongly on the expected tail count $nq$, not just on the total number of claims. For a fixed threshold and a fixed positive exceedance count $r$, the conditional tail mean is an average of those $r$ observed exceedances. If the conditional severity variance is finite, its variance given $r$ is that variance divided by $r$. A small $r$ therefore makes the estimate sensitive to individual observations.

The sample maximum has a different interpretation from a fixed quantile or a tail mean. It is simply the largest observed claim. If the underlying claim distribution has distribution function $F_X$, and the $n$ observations are independent, then the probability that every claim is at most $x$ is the probability that each individual claim is at most $x$, multiplied across the sample.

Result: For independent claims with common distribution function $F_X$, the sample maximum $M_n=\max(X_1,\ldots,X_n)$ satisfies $P(M_n\le x)=F_X(x)^n$, and hence $P(M_n>x)=1-F_X(x)^n$. Increasing the sample size increases the probability of observing a claim above any fixed threshold for which $0<F_X(x)<1$. The observed maximum is therefore sample-size dependent; it is not a stable estimate of a fixed high quantile by itself.

This result follows directly from independence: the event $M_n\le x$ occurs exactly when all $n$ claims are at most $x$. Each has probability $F_X(x)$, so the joint probability is $F_X(x)^n$. Taking the complement gives the probability of at least one claim above $x$. A larger observed maximum in a larger sample does not, on its own, show that the underlying claim distribution has changed.

Worked Example: The Maximum and a Tail Mean Respond Differently

A hypothetical sample contains 20 claims. Eighteen are at most 50 thousand dollars, and the remaining two are 60 and 90 thousand dollars. Compare the empirical exceedance probability above 50, the conditional tail mean above 50, and the overall sample mean before and after replacing 90 by 300 thousand dollars.

Method. The threshold is fixed at 50, and both of the two specified claims exceed it before and after the change.

\hat S_X(50)=\frac{2}{20}=0.10,\qquad \widehat{E[X\mid X>50]}=\frac{60+90}{2}=\frac{150}{2}=75.

The initial maximum is 90 thousand dollars. After the replacement, the exceedance count is still two, so the empirical probability remains $2/20=0.10$. The new conditional tail mean is $(60+300)/2=360/2=180$ thousand dollars, and the new maximum is 300 thousand dollars.

The replacement increases the total sample by $300-90=210$ thousand dollars. Consequently, the overall sample mean increases by $210/20=10.5$ thousand dollars, while the tail mean increases by $210/2=105$ thousand dollars. The tail mean changes by ten times as much because it averages two observations rather than 20.

Check. The exceedance probability is unchanged because neither the count nor the sample size changed. The maximum equals the largest single observation, and the tail-mean change $180-75=105$ matches $210/2$. This separates the effects on frequency, tail severity, and the sample maximum.

Common Mistakes and Practical Interpretation

Confusing the denominators. Divide the exceedance count by $n$ for the empirical exceedance probability. Divide the sum of exceedance amounts by $r$ for the conditional tail mean.
Misdescribing an observation’s contribution. With $r$ exceedances, each observation has weight $1/r$ in the average; its additive contribution is its amount divided by $r$. It does not contribute $1/r$ of the mean unless all exceedance amounts are equal. Its share of the mean is its amount divided by the sum of the exceedance amounts.
Including the threshold claim by accident. For the definition $\hat S_X(t)=P(X>t)$ estimated empirically, a claim equal to $t$ is not an exceedance. Use the inequality stated in the question.
Treating a maximum as a quantile. The maximum is one order statistic and changes with sample size. A tail quantile or tail average is a different summary, with its own definition and data requirements.
Ignoring the number of tail observations. Report the exceedance count along with a tail estimate. A mean based on two claims should not be interpreted as if it were supported by hundreds.

These distinctions are useful when comparing periods or portfolios. The threshold, units, exposure basis, and treatment of claims must be comparable; otherwise a change in an empirical tail summary may reflect a change in the data definition rather than a change in underlying risk. This discussion is aligned with the SOA Exam FAM (Fundamentals of Actuarial Mathematics) syllabus.

Check Your Understanding

For each problem, distinguish the exceedance count, the relevant tail average, and the effect of an extreme observation.

A hypothetical sample of eight claims, in thousands of dollars, is 8, 12, 20, 35, 45, 50, 55, and 100. Using a strict threshold of 40, calculate the empirical exceedance probability and conditional tail mean. (Answer: $4/8=0.5000$; $(45+50+55+100)/4=62.50$ thousand dollars.)
A sample of five claims, in thousands, is 4, 7, 9, 12, and 28. What is the empirical average of the upper 20%? If 28 is replaced by 48, what is the change in that average? (Answer: 28 thousand dollars initially; the average increases by 20 thousand dollars to 48 thousand dollars.)
Of 12 hypothetical claims, exactly three exceed 100 thousand dollars, with amounts 110, 130, and 260 thousand dollars. Calculate the empirical exceedance probability and conditional tail mean. Then find the revised tail mean if 260 is replaced by 160. (Answer: Probability $3/12=0.2500$; initial mean $500/3=166.67$ thousand dollars; revised mean $400/3=133.33$ thousand dollars, a decrease of 33.33 thousand dollars.)
A sample of six claims, in thousands, is 5, 8, 10, 12, 20, and 50. Using the fractional-observation convention, calculate the empirical average of the upper 20%. (Answer: $m=6(0.20)=1.2$; $(50+0.2(20))/1.2=54/1.2=45$ thousand dollars.)
Assume 10 independent claims come from a common severity distribution with $F_X(100)=0.98$. Calculate the probability that the sample maximum is at most 100, and the probability that at least one claim exceeds 100. (Answer: $0.98^{10}=0.8171$; $1-0.98^{10}=0.1829$.)

What You'll Learn