Tail-Sensitive Descriptive Summaries

Why Summarize the Tail?

The overall mean describes the average loss across the entire distribution. It can be useful for pricing and planning, but it may conceal how often unusually large losses occur and how severe they are when they do occur. Tail-sensitive descriptive summaries focus on those questions: what loss level marks an upper part of the distribution, how frequently does loss exceed a chosen amount, and what is the average loss among exceedances?

These are descriptive views of a loss distribution, not interchangeable measures. An upper quantile gives a loss amount, an exceedance rate gives a probability, and a conditional average gives an amount among a restricted set of outcomes. The overall mean, by contrast, averages over all outcomes. Keeping the units and conditioning event clear is essential.

Definition: Let $X$ be a nonnegative loss with cumulative distribution function $F_X(x)=P(X\leq x)$ and survival function $S_X(x)=P(X>x)$. For $0<p<1$, an upper $p$-quantile is $$ q_p=\inf\{x:F_X(x)\geq p\}. $$ For a threshold $t$, the strict exceedance rate is $P(X>t)=S_X(t)$. When $P(X>t)>0$, the conditional average above $t$ is $E[X\mid X>t]$. Each summary describes a different feature of the loss distribution.

The quantile is a loss amount at a specified cumulative probability level. For example, $q_{0.90}$ is the smallest loss amount at which the cumulative probability is at least 0.90. The strict exceedance rate at $t$ counts only losses greater than $t$, not losses equal to $t$. That distinction matters in a discrete distribution or whenever there is a probability mass at the threshold.

A conditional average is defined only if exceedances have positive probability. It averages the losses in the event $X>t$, rather than averaging all losses and then adjusting the result. It is not automatically the mean of a fixed fraction of the largest observations: if there is a mass at a quantile, the probability strictly above that quantile may differ from $1-p$.

Calculating the Three Summaries

For a discrete loss distribution with possible values $x_j$ and probabilities $p_j$, the overall mean is $\sum_j x_jp_j$. For a threshold $t$, add the probabilities of the outcomes strictly greater than $t$ to find the exceedance rate. Then weight those same loss values by their probabilities and divide by the exceedance probability to find the conditional average.

P(X>t)=\sum_{x_j>t}p_j,\qquad E[X\mid X>t]=\frac{\sum_{x_j>t}x_jp_j}{\sum_{x_j>t}p_j}.

For a continuous loss with density $f_X$, the corresponding calculations use integrals:

P(X>t)=\int_t^\infty f_X(x)\,dx,\qquad E[X\mid X>t]=\frac{\int_t^\infty x f_X(x)\,dx}{P(X>t)},

provided the denominator is positive and the numerator is finite. In either case, the overall mean uses all outcomes; the conditional mean uses only the outcomes in the stated tail event.

Worked Example: Discrete Losses and a Strict Threshold

A hypothetical portfolio has annual loss $X$, measured in thousands of dollars, with $P(X=2)=0.20$, $P(X=5)=0.50$, and $P(X=9)=0.30$. Find the overall mean, the strict exceedance rate above 5, the conditional average above 5, and the 80th percentile. Report dollar amounts to the nearest cent and probabilities to four decimals.

Method. Use probability-weighted sums for the overall mean and conditional average, and the cumulative distribution for the percentile. This avoids treating a probability-weighted loss contribution as though it were itself a conditional average.

Overall mean. In thousands of dollars, $2(0.20)=0.40$, $5(0.50)=2.50$, and $9(0.30)=2.70$. Therefore, $E[X]=0.40+2.50+2.70=5.60$ thousand dollars, or $5,600.00.

Strict exceedance rate above 5. Only the outcome 9 is strictly greater than 5, so $P(X>5)=0.30$, or 0.3000.

Conditional average above 5. The only loss in the event $X>5$ is 9. Its weighted contribution is $9(0.30)=2.70$ thousand dollars. Divide by the event probability: $2.70/0.30=9.00$ thousand dollars, or $9,000.00.

80th percentile. The cumulative probabilities are $F_X(2)=0.20$, $F_X(5)=0.70$, and $F_X(9)=1.00$. The smallest loss with cumulative probability at least 0.80 is 9, so $q_{0.80}=$$9,000.00.

Check. The probabilities sum to $0.20+0.50+0.30=1.00$. The conditional average of $9,000.00 is at least the threshold of $5,000.00, as it must be.

Why the Conditional-Average Formula Holds

The conditional average follows directly from the definition of conditional expectation. In the discrete case, the probability of an outcome $x_j$ given $X>t$ is $P(X=x_j\mid X>t)=p_j/P(X>t)$ for each $x_j>t$. Averaging the eligible outcomes with these conditional probabilities gives:

E[X\mid X>t] =\sum_{x_j>t}x_j\frac{p_j}{P(X>t)} =\frac{\sum_{x_j>t}x_jp_j}{P(X>t)}.

The numerator is the contribution of exceedance outcomes to the unconditioned mean; it is not the conditional average by itself. Dividing by the probability of entering the exceedance set reweights those outcomes so their conditional probabilities sum to one. The same logic gives the density-and-integral formula for a continuous loss.

Worked Example: Empirical Tail Summaries With a Tie at the Threshold

A hypothetical insurer records ten annual losses, in thousands of dollars, sorted from smallest to largest: 1, 2, 2.5, 4, 7, 12, 20, 35, 50, and 80. Treat the observations as equally likely outcomes in an empirical distribution. Find the overall mean, the 90th percentile using $q_p=\inf\{x:F_X(x)\geq p\}$, and the strict exceedance rate and conditional average above $20,000. Report money to the nearest cent and the rate to four decimals.

Method. Treat each observation as having probability $1/10$, use the cumulative empirical distribution for the quantile, and count only observations strictly above the threshold. This avoids including the tied observation of $20,000 in the strict exceedance set.

Overall mean. The sum is $1+2+2.5+4+7+12+20+35+50+80=213.5$ thousand dollars. Dividing by ten gives $21.35$ thousand dollars, or $21,350.00.

90th percentile. Each observation increases the cumulative probability by 0.10. At the 9th ordered value, $50,000, the cumulative probability reaches 0.90. Thus $q_{0.90}=$$50,000.00.

Strict exceedance rate above $20,000. The observations strictly above 20 are 35, 50, and 80. Three of ten exceed, so the rate is $3/10=$0.3000.

Conditional average above $20,000. The eligible losses sum to $35+50+80=165$ thousand dollars. Their average is $165/3=55$ thousand dollars, or $55,000.00.

Check. The threshold observation of 20 is included in the overall mean but excluded from the strict exceedance calculations. The conditional average exceeds $20,000, consistent with its definition.

A Continuous-Model Example

For a continuous model, the same distinctions apply, but probabilities and conditional means are calculated with the density or survival function. The following example uses an exponential severity model, which assumes independent loss amounts with a constant rate of occurrence across the loss scale. The model is hypothetical and is used only to illustrate the calculations.

Worked Example: Exceedances Under an Exponential Loss Model

Suppose a hypothetical positive loss $X$ is exponentially distributed with mean $\theta=$$4,000, so its density is $f_X(x)=\theta^{-1}e^{-x/\theta}$ for $x\geq0$. Find the overall mean, the strict exceedance rate above $6,000, the conditional average above $6,000, and the 90th percentile. Report money to the nearest cent and the probability to four decimals.

Method. Use the exponential survival function and density. The model parameter is the mean, $\theta=$$4,000; the exponential model assumes a continuous, nonnegative loss. This example avoids confusing the threshold exceedance rate with the probability associated with a percentile.

Overall mean. By the parameterization, $E[X]=\theta=$$4,000.00.

Exceedance rate. For this model, $S_X(t)=e^{-t/\theta}$. At $t=$$6,000, $t/\theta=6,000/4,000=1.5$, so $P(X>6,000)=e^{-1.5}=0.223130\ldots$, or 0.2231.

Conditional average. The numerator is $\int_6^\infty x(1/4)e^{-x/4}\,dx$ when losses are measured in thousands of dollars. Integration by parts gives $e^{-6/4}(6+4)=10e^{-1.5}$ thousand dollars. Dividing by the exceedance probability $e^{-1.5}$ gives 10 thousand dollars, or $10,000.00.

90th percentile. Solve $F_X(q_{0.90})=0.90$. Since $F_X(x)=1-e^{-x/\theta}$, we have $e^{-q_{0.90}/4,000}=0.10$, and therefore $q_{0.90}=-4,000\ln(0.10)=9,210.340\ldots$ dollars. Rounded to the nearest cent, $q_{0.90}=$$9,210.34.

Check. The 90th percentile exceeds $6,000, and the conditional average above $6,000 exceeds that threshold. The overall mean is lower because it also includes losses below $6,000.

Interpreting Results and Avoiding Traps

Tail summaries are useful because they separate distinct questions. A quantile identifies a loss level at a specified cumulative probability. An exceedance rate measures the probability of crossing a selected threshold. A conditional average describes the size of losses in the exceedance set. The overall mean summarizes all losses. These quantities can be reported together, but one should not be substituted for another.

A frequent arithmetic and interpretation error is to stop after calculating the weighted contribution to a conditional mean. In the first example, $9(0.30)=2.70$ thousand dollars is the contribution of that outcome to the overall mean. It is not the conditional average above 5. The conditional average is $2.70/0.30=9.00$ thousand dollars. Omitting the division confuses an unconditional contribution with an average after conditioning.

Check the inequality. “Above 5” ordinarily means $X>5$; “at least 5” means $X\geq5$. They differ when $P(X=5)>0$.
Keep units visible. If losses are measured in thousands, a calculated value of 5.6 represents $5,600, not $5.60.
Use the correct denominator. A conditional average divides by the probability of the conditioning event, or by the number of observations in that event for an equally weighted empirical sample.
Do not infer a tail probability from a quantile without checking atoms. A discrete distribution can have a jump in its cumulative probability, so the probability strictly above $q_p$ need not equal $1-p$.
Check the direction of the result. If losses are nonnegative and $P(X>t)>0$, then $E[X\mid X>t]>t$. A result below the threshold signals an error or a different conditioning event.

Key takeaway: A quantile is a loss amount, an exceedance rate is a probability, and a conditional average is a loss amount within a specified event. State the threshold and whether the event is strict, then calculate and interpret each summary separately from the overall mean.

Check Your Understanding

For each problem, state which outcomes or events are included before calculating. Money is in thousands of dollars unless otherwise noted.

A discrete loss takes values 1, 4, and 10 with probabilities 0.50, 0.30, and 0.20. Find the overall mean, the 75th percentile under the stated quantile definition, the rate strictly above 4, and the conditional average strictly above 4. (Answer: mean = $3,700.00; 75th percentile = $4,000.00; exceedance rate = 0.2000; conditional average = $10,000.00.)
Six equally likely observed losses are 3, 5, 5, 8, 14, and 20. Find their overall mean, their empirical median using $q_p=\inf\{x:F_X(x)\geq p\}$, the strict exceedance rate above 8, and the average of those exceedances. (Answer: mean = $9,166.67; median = $5,000.00; exceedance rate = 0.3333; conditional average = $17,000.00.)
A continuous loss is uniform from $0 to $12,000. Find the overall mean, the 90th percentile, the probability strictly above $9,000, and the conditional average above $9,000. (Answer: mean = $6,000.00; 90th percentile = $10,800.00; exceedance rate = 0.2500; conditional average = $10,500.00.)
A hypothetical exponential loss has mean $3,000. Find the probability strictly above $6,000, the conditional average above $6,000, and the 95th percentile. (Answer: exceedance rate = $e^{-2}=0.1353$; conditional average = $9,000.00; 95th percentile = $8,987.20.)
A discrete loss takes values $0, $10,000, and $30,000 with probabilities 0.40, 0.45, and 0.15. Find the overall mean, the 80th percentile, the rate at least $10,000, and the average loss at least $10,000. (Answer: mean = $9,000.00; 80th percentile = $10,000.00; rate at least $10,000 = 0.6000; conditional average = $15,000.00.)

What You'll Learn