Why Estimate Tail Co-Exceedance From Loss Data?
Two lines of business can have modest overall correlation yet experience unusually large losses in the same periods. Conversely, a strong overall association need not mean that their most severe losses coincide. Tail co-exceedance estimates focus on how often observations from one marginal tail also fall in the corresponding tail of the other margin.
With empirical losses, this is a counting exercise: select observations in each margin’s tail, count how many paired observations are in both selected tails, and compare that joint count with the count in the conditioning tail. The arithmetic is simple; interpretation requires care. In a small sample, a tail estimate may depend on only a handful of paired observations.
Constructing Empirical Tail Sets
For a sample of \(n\) values, write the ordered observations of \(X\) as \(X_{(1)}\le \cdots \le X_{(n)}\); define the ordering of \(Y\) similarly. For a chosen integer \(k\), the upper-tail set consists of the \(k\) largest observations in that margin. The lower-tail set consists of its \(k\) smallest observations. The corresponding empirical tail fraction is \(k/n\).
When all values in a margin are distinct, the upper tail can equivalently be selected by a threshold between \(X_{(n-k)}\) and \(X_{(n-k+1)}\), and the lower tail by a threshold between \(X_{(k)}\) and \(X_{(k+1)}\). These rank-based definitions avoid ambiguity about interpolated empirical quantiles. If ties occur at a boundary, state and apply a consistent tie rule; including every tied value can make a tail contain more than \(k\) observations, so the actual marginal tail count must then be used in the denominator.
Let \(I_U^X\) and \(I_U^Y\) be the sets of observation indices in the selected upper tails, and let \(I_L^X\) and \(I_L^Y\) be the lower-tail sets. With exactly \(k\) observations in each marginal tail:
These are sample conditional proportions: among the observations in the \(X\) tail, what fraction also lie in the corresponding \(Y\) tail? Each value lies between 0 and 1. The estimate need not be the same if the conditioning margin is switched in a sample with ties, unequal tail counts, or different selection conventions; under the equal-size rank construction without such complications, the joint count is the same in either direction.
Why the Count Ratio Is a Conditional Estimate
The empirical probability of joint upper-tail membership is the number of joint observations divided by \(n\). The empirical probability of membership in the \(X\) upper tail is \(k/n\). Dividing the first quantity by the second gives the conditional proportion:
The same argument applies to the lower tails. This derivation explains why the denominator is the marginal tail count, not the full sample size. Dividing by \(n\) instead gives the empirical joint tail probability, a different quantity.
A finite-sample estimate at a selected tail fraction is not itself a limiting tail-dependence coefficient. A limiting coefficient describes behavior as the probability level moves increasingly far into the tail. An empirical calculation uses the observed losses and a particular finite tail fraction; it can inform that question, but cannot establish the limiting behavior from a few extreme pairs alone.
Worked Example: Comparing Upper- and Lower-Tail Co-Exceedance
A hypothetical insurer records paired annual losses, in thousands of dollars, for two coverage segments over eight years. Use the two largest and two smallest observations in each margin to estimate upper- and lower-tail co-exceedance.
| Year | Segment X | Segment Y |
|---|---|---|
| 1 | 20 | 20 |
| 2 | 35 | 50 |
| 3 | 40 | 35 |
| 4 | 50 | 60 |
| 5 | 60 | 45 |
| 6 | 70 | 75 |
| 7 | 80 | 95 |
| 8 | 100 | 85 |
Method. Rank each margin separately, select \(k=2\), then count paired years in both selected sets. The tail fraction is \(2/8=0.25\).
Upper tails. Segment X’s two largest losses are 100 in year 8 and 80 in year 7. Segment Y’s two largest are 95 in year 7 and 85 in year 8. Both years 7 and 8 are in both upper-tail sets, so the joint count is 2. Therefore, \(\widehat{\lambda}_U(2)=2/2=1.0000\).
Lower tails. Segment X’s two smallest losses are 20 in year 1 and 35 in year 2. Segment Y’s two smallest are 20 in year 1 and 35 in year 3. Only year 1 belongs to both sets, so \(\widehat{\lambda}_L(2)=1/2=0.5000\).
Check. Each marginal tail contains exactly two years, so the denominators are 2. The estimates are between 0 and 1. They describe these selected empirical tails; they do not prove that the population has perfect upper-tail dependence.
Worked Examples With Ranked Observations
Worked Example: Estimating Both Tails at a Different Tail Fraction
A hypothetical reinsurer compares paired annual losses, in thousands of dollars, for two portfolios over ten years. The \(X\) values increase by year. The observed \(Y\) values, listed in year order, are 90, 20, 80, 30, 70, 40, 60, 50, 100, and 10. Use \(k=3\) and calculate both empirical estimates.
Method. Select the three largest and three smallest values in each margin by their ranks. Here the empirical tail fraction is \(3/10=0.30\).
Upper tail. Since \(X\) increases by year, its three largest observations are in years 8, 9, and 10. The three largest \(Y\) values are 100 in year 9, 90 in year 1, and 80 in year 3. The intersection contains year 9 only. Thus the joint count is 1, and \(\widehat{\lambda}_U(3)=1/3=0.3333\).
Lower tail. The three smallest \(X\) values are in years 1, 2, and 3. The three smallest \(Y\) values are 10 in year 10, 20 in year 2, and 30 in year 4. The intersection contains year 2 only. Thus \(\widehat{\lambda}_L(3)=1/3=0.3333\).
Check. Both estimates use a joint count of 1 and a conditioning-tail count of 3. The equality of the two estimates is a feature of these observed ranks, not a rule that upper and lower co-exceedance must agree. The common mistake avoided here is identifying tail observations by year rather than ranking each margin separately.
Worked Example: Interpreting a Sparse Extreme Tail
A hypothetical insurer has twelve paired annual loss observations, in thousands of dollars. The \(X\) values are 12, 18, 25, 31, 38, 44, 52, 61, 70, 78, 88, and 105 in years A through L, respectively. The corresponding \(Y\) values are 50, 15, 45, 20, 55, 30, 35, 25, 85, 40, 60, and 90. Estimate upper-tail co-exceedance using the two largest observations in each margin, then assess how much evidence the result provides.
Calculation. The two largest \(X\) observations are in years K and L, with values 88 and 105. The two largest \(Y\) observations are in years L and I, with values 90 and 85. Only year L appears in both upper-tail sets. There is one joint observation among two \(X\)-tail observations, so \(\widehat{\lambda}_U(2)=1/2=0.5000\). The selected tail fraction is \(2/12=0.1667\), rounded to four decimals.
Interpretation and check. The arithmetic is correct: the numerator is 1, the denominator is 2, and the estimate is between 0 and 1. But this estimate is determined by only two conditioning observations. With \(k=2\), possible estimates change in increments of \(1/2\): 0, 0.5000, or 1.0000. The value 0.5000 is therefore a coarse sample summary, not precise evidence that the underlying conditional probability is exactly one-half.
The common mistake is to describe this as strong evidence of a stable half-probability relationship. The correction is to report the estimate together with its tail count and sample size, and to acknowledge that one paired year changes the result by 0.5000.
What Small Samples Can and Cannot Show
The main limitation is not a complicated formula; it is the number of observations that actually determine the estimate. A dataset can contain hundreds of paired losses while a very extreme selected tail contains only a few pairs. If the joint count is zero, the estimate is zero at that chosen empirical threshold, but this does not establish that joint extremes are impossible. If every selected \(X\)-tail observation also falls in the \(Y\) tail, the estimate is one, but that does not establish perfect dependence beyond the observed sample.
A useful report identifies the sample size, the number of observations selected in each margin, the tail fraction, the joint count, the estimated proportion, and any tie convention. For example, “two of eight upper-tail years overlap” is more informative than reporting \(1.0000\) alone. The first statement makes the limited information visible.
Empirical estimates also inherit the properties of the observation set. Paired observations must correspond to the same years or exposure units; independently sorting the two columns and pairing the sorted values would destroy the observed dependence. Likewise, selecting the \(X\) tail correctly but using the full sample size as the denominator confuses a conditional proportion with a joint frequency.
Confirm that each \(X_i,Y_i\) pair represents the same period or unit and that both variables are measured consistently.
Choose \(k\), calculate \(k/n\), and define the upper or lower sets by marginal ranks. State how boundary ties are handled.
Find the paired observation indices present in both selected tail sets, without re-pairing or independently sorting the data rows.
Divide the joint count by the conditioning-tail count and report both counts. Interpret it as a finite-sample conditional proportion, not as proof of a limiting tail property.
Common Mistakes
- Using \(n\) as the denominator for a conditional estimate. The joint count divided by \(n\) is an empirical joint frequency. Divide by the number of observations in the conditioning tail to obtain the conditional co-exceedance proportion.
- Ranking only one margin. An observation is in the upper tail only if its value ranks among the largest in that particular margin. Identify each margin’s tail independently, then intersect the paired indices.
- Calling a finite-threshold estimate a limiting coefficient. The calculation describes the selected sample tail. It does not, by itself, establish what happens arbitrarily far into the population tail.
- Reporting a precise-looking decimal without the count. A value such as 0.5000 can be based on one joint year among two. Include the numerator, denominator, and tail fraction.
- Ignoring ties or changing quantile rules silently. A boundary tie can alter the number of selected observations. State the convention and use the actual conditioning-tail count.
Check Your Understanding
For each problem, identify the relevant tail sets before calculating. Answers are shown so you can check both the estimate and its denominator.
- A hypothetical portfolio has eight paired loss observations. The four largest \(X\) values occur in periods A, B, C, and D; the four largest \(Y\) values occur in periods B, D, E, and F. Using \(k=4\), calculate upper-tail co-exceedance. (Answer: 2/4 = 0.5000.)
- For a ten-year paired loss sample, the three smallest \(X\) values occur in years 1, 3, and 5; the three smallest \(Y\) values occur in years 3, 4, and 5. Calculate the lower-tail estimate and state its empirical tail fraction. (Answer: 2/3 = 0.6667; tail fraction = 3/10 = 0.3000.)
- A sample contains 24 paired observations. Using the five largest values in each margin, the upper-tail sets overlap in three observations. Calculate the conditional estimate and explain why dividing by 24 answers a different question. (Answer: \(3/5=0.6000\); \(3/24=0.1250\) is the joint empirical frequency, not the conditional estimate.)
- In a hypothetical sample of 15 paired annual losses, \(k=3\). The smallest \(X\) observations occur in years A, B, and C; the smallest \(Y\) observations occur in years C, D, and E. Calculate the lower-tail estimate and explain the effect of one additional overlapping year. (Answer: \(1/3=0.3333\); one additional overlap would increase it to \(2/3=0.6667\).)
- For \(n=20\), the two largest \(X\) losses and two largest \(Y\) losses share no observation. State the upper-tail estimate and one conclusion that is not justified by it. (Answer: \(0/2=0.0000\); it does not establish that population upper-tail co-exceedance is impossible.)