Empirical Exceedance Counts And Rates

Why Count Losses Above a Threshold?

A tail summary answers a direct question: in the observed data, how many losses exceed a specified amount, and what fraction of the observations does that represent? An insurer might use such a summary to describe how often hypothetical losses in a sample exceed a large-loss reporting point. The count gives the number of observations; the rate puts that count on a scale that can be compared with samples of different sizes.

The calculation is simple, but the wording matters. “Greater than $10,000” excludes losses exactly equal to $10,000; “at least $10,000” includes them. Because empirical loss data can contain ties, treating those phrases as interchangeable can change both the count and the rate.

Definition: For $n$ completely observed losses $x_1,\ldots,x_n$ and a specified threshold $c$, the strict exceedance count is the number of observations greater than $c$: $$ N_{>}(c)=\sum_{i=1}^{n} I(x_i>c). $$ The strict empirical exceedance rate is $$ \widehat{p}_{>}(c)=\frac{N_{>}(c)}{n}. $$ Here $I(\text{condition})$ equals 1 when the condition is true and 0 otherwise. For an inclusive threshold, replace $>$ by $\geq$ in both definitions. The denominator is the number of observations in the data set.

The strict empirical exceedance rate is the fraction of observed losses that exceed $c$. It is also the empirical survival estimate at $c$: $\widehat{S}_X(c)=\widehat{p}_{>}(c)$. In contrast, the empirical distribution function at $c$ counts observations less than or equal to $c$. Thus, for the same complete data,

\widehat{S}_X(c)=1-F_n(c)=\frac{\#\{i:x_i>c\}}{n}.

This identity follows because each observation belongs to exactly one of the disjoint groups $x_i\leq c$ and $x_i>c$. If the first group contains $n-N_{>}(c)$ observations, the second contains $N_{>}(c)$. Dividing the second count by $n$ gives the empirical exceedance rate, while subtracting the first group’s empirical proportion from 1 gives the same result.

Counting Correctly When Observations Tie

A reliable calculation starts by translating the threshold wording into a comparison sign. For a strict exceedance, count only values that satisfy $x_i>c$. For an inclusive exceedance, count values satisfying $x_i\geq c$. Values equal to the threshold contribute zero in the strict case and one in the inclusive case.

The rate always uses the number of observed losses as its denominator. Do not divide by the number of distinct loss amounts, the number of losses above the threshold, or the number of loss amounts appearing in a table. If observations are listed with frequencies, the total frequency is $n$, and each frequency contributes the stated number of observations to the count.

Worked Example: Strict and Inclusive Counts With Tied Losses

A hypothetical data set contains 12 complete losses, in thousands of dollars: $8,12,12,17,20,20,23,29,31,31,36,44$. Calculate the strict exceedance count and rate above $20$, the inclusive count and rate at $20$, and the strict count and rate above $30$. State which method you use and identify the tie-related mistake to avoid.

Method. Compare every observed loss with each threshold using the requested inequality, then divide the count by $n=12$. The values are stated in thousands, so the threshold $20$ means $20,000.

Strictly above 20. The values satisfying $x_i>20$ are $23,29,31,31,36,44$. There are 6 such losses, so

N_{>}(20)=6,\qquad \widehat{p}_{>}(20)=\frac{6}{12}=0.5000.

At least 20. The two losses equal to 20 now count as well. The qualifying values are $20,20,23,29,31,31,36,44$, so there are 8:

N_{\geq}(20)=8,\qquad \widehat{p}_{\geq}(20)=\frac{8}{12}=0.6667.

Strictly above 30. The qualifying values are $31,31,36,44$. Thus,

N_{>}(30)=4,\qquad \widehat{p}_{>}(30)=\frac{4}{12}=0.3333.

Check. At threshold 20, the inclusive count exceeds the strict count by exactly the two observations tied at 20. Raising the strict threshold from 20 to 30 reduces the count from 6 to 4, as expected. Rates are rounded to four decimal places.

Common mistake avoided. Using $8/12$ for “strictly above 20” incorrectly includes the two losses equal to 20. The correct strict calculation is $6/12$; $8/12$ answers the different question “at least 20.”

How Threshold Changes Affect the Tail Summary

For a fixed set of observations, increasing a strict threshold cannot increase the exceedance count. If $c_2>c_1$, every observation greater than $c_2$ is also greater than $c_1$, but some observations greater than $c_1$ may not exceed $c_2$. Therefore $N_{>}(c_2)\leq N_{>}(c_1)$. Since both rates use the same positive denominator $n$, the rates have the same ordering.

The count or rate may stay unchanged when the threshold moves if no observed loss lies between the two threshold values. It decreases when observations that previously exceeded the threshold no longer do. With tied losses, the size of a decrease can equal the number of observations at a value crossed by the threshold. These statements describe the observed data; they do not, by themselves, establish how often future losses will exceed either threshold.

c_2>c_1 \quad\Longrightarrow\quad N_{>}(c_2)\leq N_{>}(c_1) \quad\text{and}\quad \widehat{p}_{>}(c_2)\leq\widehat{p}_{>}(c_1).

Worked Example: Measuring the Change in Rate

A hypothetical portfolio has 10 complete losses, in thousands of dollars: $4,5,7,9,9,10,13,15,18,22$. Find the strict exceedance counts and rates above 9 and above 14. Then calculate the change in count, the change in rate in percentage points, and the relative reduction in rate when the threshold rises.

Method. Apply the strict comparison separately at each threshold. Use all 10 losses as the denominator for both rates so that the comparison is on the same basis.

Above 9. The observations greater than 9 are $10,13,15,18,22$. Therefore the count is 5 and the rate is

N_{>}(9)=5,\qquad \widehat{p}_{>}(9)=\frac{5}{10}=0.5000.

Above 14. The observations greater than 14 are $15,18,22$. The observation equal to 14 is not in the data; in any event, a value equal to the threshold would not count under a strict comparison. Thus,

N_{>}(14)=3,\qquad \widehat{p}_{>}(14)=\frac{3}{10}=0.3000.

The count change, final minus initial, is $3-5=-2$ losses. The rate changes from 50% to 30%, a decrease of 20 percentage points. Relative to the initial rate, the reduction is

\frac{0.5000-0.3000}{0.5000}=0.4000=40\%.

Check. The higher threshold has a lower count and rate, consistent with monotonicity. The 40% relative reduction is not the same quantity as the 20-percentage-point decrease. Rates are shown to four decimal places before expressing the percentage comparisons.

Common mistake avoided. Calling the difference $0.20$ a 20% relative reduction confuses percentage points with a relative change. Dividing the decrease by the initial rate gives $0.20/0.50=0.40$, or 40%.

Using Frequencies and Checking the Result

Loss observations are sometimes presented as distinct values with their frequencies rather than as a list of individual records. A frequency is the number of observations at that exact value. Add all frequencies to obtain $n$; then add only the frequencies for values that satisfy the threshold condition. This produces the same count and rate as expanding the frequency table into a list.

Worked Example: Exceedance Rates From a Frequency Table

A hypothetical set of 15 complete losses, in thousands of dollars, is summarized below. Calculate the strict exceedance count and rate above 10, and above 24. Find the decrease in rate between the two thresholds.

Loss (thousands of dollars)	Observed frequency
0	3
5	4
10	2
15	3
25	2
40	1

Method. First check the total frequency: $3+4+2+3+2+1=15$, matching the stated sample size. Apply the strict comparison to each loss value and include the full frequency whenever it qualifies.

Above 10. The values 15, 25, and 40 exceed 10. Their frequencies sum to $3+2+1=6$, so

N_{>}(10)=6,\qquad \widehat{p}_{>}(10)=\frac{6}{15}=0.4000.

Above 24. The values 25 and 40 exceed 24. Their frequencies sum to $2+1=3$, so

N_{>}(24)=3,\qquad \widehat{p}_{>}(24)=\frac{3}{15}=0.2000.

The rate decreases by $0.4000-0.2000=0.2000$, or 20 percentage points. Equivalently, the count decreases by $6-3=3$ observations, and $3/15=0.2000$. The equality check is consistent: the rate decrease equals the count decrease divided by the unchanged sample size.

Check. The threshold rises from 10 to 24, and the count falls from 6 to 3. The two observations at 10 do not count for a strict exceedance above 10; including them would answer an inclusive-threshold question instead.

Common mistake avoided. Dividing by the six distinct loss values would give the wrong denominator. The data contain 15 observations, so every empirical rate uses 15.

Common Checks Before Reporting a Tail Rate

Before reporting an empirical exceedance summary, check the comparison, the sample size, and the units. A useful calculation records the threshold and inequality alongside the count so a reader can tell whether ties were included. If data are reported in thousands of dollars, state that convention rather than silently treating a threshold of 20 as $20.

Check the wording. “Exceeds” means $>$; “at least” means $\geq$.
Check the denominator. Use the full number of observations represented in the data.
Check the threshold units. A threshold must use the same units as the losses.
Check the direction. Raising the threshold cannot increase a strict exceedance count or rate for fixed data.
Check the interpretation. The rate is the observed fraction in this data set, not a guarantee about a future sample.

A concise report can state the threshold, inequality, count, denominator, and rate together. For example: “Of 12 observed losses, 6 exceeded $20,000, giving an empirical exceedance rate of 0.5000.” This makes the treatment of equality and the basis of the rate explicit.

Check Your Understanding

For each problem, use the stated threshold wording and the full number of observed losses as the denominator.

A hypothetical data set of losses, in thousands of dollars, is $2,5,5,8,11,16$. Find the count and rate strictly above 8. (Answer: count $2$; rate $2/6=0.3333$.)
For the same losses, find the count and rate of losses at least 5. (Answer: count $5$; rate $5/6=0.8333$.)
A sample has 10 losses, in thousands: $3,6,8,10,10,12,14,17,19,25$. Compare the strict rates above 10 and above 15, and give the decrease in percentage points. (Answer: $5/10=0.5000$ and $3/10=0.3000$; decrease $0.2000$, or 20 percentage points.)
A frequency table has losses 4, 9, 16, and 30, with frequencies 2, 5, 3, and 2, respectively. Find the strict exceedance count and rate above 9. (Answer: $3+2=5$ of $12$; rate $5/12=0.4167$.)
In a complete sample of 20 losses, 7 exceed a threshold of 40. At a higher threshold of 50, 4 exceed it. Find both rates and the relative reduction in rate. (Answer: $7/20=0.3500$ and $4/20=0.2000$; relative reduction $(0.3500-0.2000)/0.3500=0.4286$, or 42.86%.)

What You'll Learn