Why Threshold Choice Matters
An empirical tail-dependence estimate is based on observations beyond a selected threshold. Choosing a lower threshold gives more observations to estimate the conditional proportion, but includes less-extreme outcomes. Choosing a higher threshold focuses on more-extreme losses, but leaves fewer observations and can make the estimate sensitive to one or two pairs. A single estimate therefore cannot, by itself, show whether the apparent tail relationship is stable.
A useful diagnostic is to calculate the estimate at several nested thresholds and inspect a threshold plot: the selected quantile level on the horizontal axis and the empirical estimate on the vertical axis. The plot shows sensitivity to the threshold, not the true asymptotic coefficient. A flat-looking segment is evidence of local stability in the sample, not proof that the population tail-dependence coefficient has been identified.
This tutorial concerns nonnegative loss variables and their upper tails. The estimate at a fixed \(q\) is a finite-threshold quantity. It is distinct from the upper-tail dependence coefficient, which is defined as a limiting conditional probability as the threshold probability approaches one. Threshold plots help assess how much the finite-sample estimate changes as the analysis moves further into the tail.
What a Threshold Plot Can—and Cannot—Show
A threshold plot should be read together with the counts used to make it. A value of 0.60 based on 3 conditioning exceedances is not as informative as the same value based on 60 exceedances. As the threshold rises, the denominator typically shrinks, so the estimate becomes more variable. At the same time, a lower threshold may blend ordinary dependence with extreme-loss dependence. These competing effects mean that no threshold is automatically best for every data set.
Adjacent points on a threshold plot are usually not independent: a higher-threshold set is contained within a lower-threshold set. A change between adjacent estimates may be driven by only the few observations entering or leaving the exceedance set. Consequently, do not treat a succession of plotted points as repeated independent evidence.
A rough way to convey the impact of a small denominator is to calculate a binomial-style standard-error proxy, \(\sqrt{\hat p(1-\hat p)/k}\), where \(\hat p\) is the conditional estimate and \(k\) is its conditioning count. This is only a diagnostic approximation: in a rank-based threshold analysis the threshold is estimated from the same data, and observations or exceedance indicators may not satisfy the assumptions for a binomial standard error. It should not be presented as a formal confidence interval without an appropriate inferential method.
Worked Example: Reading a Nested-Threshold Plot
A hypothetical portfolio has 100 paired annual loss observations for two risk categories. The analyst uses empirical quantiles and obtains the following nested exceedance counts. At each threshold the marginal counts are equal, denoted by \(k\), and \(H\) is the number of pairs exceeding both marginal thresholds. Calculate the conditional estimate at each threshold and assess its stability.
Method. Because the two marginal exceedance counts are equal in this example, either conditional estimate is \(H/k\). Calculate each ratio and retain its denominator when interpreting the result.
| Quantile level \(q\) | Marginal count \(k\) | Joint count \(H\) | Estimate \(H/k\) |
|---|---|---|---|
| 0.80 | 20 | 8 | 0.4000 |
| 0.90 | 10 | 5 | 0.5000 |
| 0.95 | 5 | 3 | 0.6000 |
| 0.97 | 3 | 2 | 0.6667 |
At \(q=0.80\), the estimate is \(8/20=0.4000\). At \(q=0.90\), it is \(5/10=0.5000\). At \(q=0.95\), it is \(3/5=0.6000\). At \(q=0.97\), it is \(2/3=0.6667\), rounded to four decimal places. The estimates rise with the threshold in this sample, but the evidence at the highest threshold consists of only three conditioning exceedances.
For a rough variability comparison, the proxy at \(q=0.80\) is \(\sqrt{0.4(1-0.4)/20}=\sqrt{0.012}=0.1095\). At \(q=0.97\), it is \(\sqrt{(2/3)(1/3)/3}=\sqrt{2/27}=0.2722\). Both are rounded to four decimals. This illustrates why the highest-threshold point can move substantially when a single pair changes classification. The proxy is not a formal standard error for this rank-based analysis.
Check. Each joint count is no greater than its marginal count, and the joint counts decline or stay the same as the nested thresholds rise. The estimates are therefore valid proportions. The plot suggests threshold sensitivity, not a proven increase in the population’s asymptotic tail dependence. The common mistake avoided here is to report only the final estimate without showing how few observations support it.
Conditional Direction and Unequal Counts
The two conditional proportions need not be equal. If \(N_X(q)\) and \(N_Y(q)\) differ, the same joint count is divided by different denominators. Even when both thresholds use the same nominal quantile level, unequal empirical counts can occur because of ties or the chosen quantile and exceedance conventions. State which variable defines the conditioning event before reporting a conditional estimate.
Worked Example: Unequal Tail Counts and Zero Co-Exceedances
In a hypothetical sample at a selected threshold, \(N_X=4\), \(N_Y=5\), and \(H=2\). Calculate both conditional estimates. Then consider a second threshold where the marginal counts are 2 and 3 but there are no joint exceedances.
Method. For the probability of an \(X\)-exceedance conditional on a \(Y\)-exceedance, divide \(H\) by \(N_Y\). For the reverse conditional estimate, divide \(H\) by \(N_X\).
The first conditional estimate is \(H/N_Y=2/5=0.4000\). The reverse is \(H/N_X=2/4=0.5000\). They differ because their denominators differ. At the second threshold, \(H=0\), so the two estimates are \(0/3=0.0000\) and \(0/2=0.0000\), even though the marginal counts are unequal.
Check. Each joint count is at most either marginal count. The general correction is important: equal tail counts ensure that the two conditional proportions agree, but equal counts are not necessary. If \(H=0\), both proportions are zero whenever both denominators are positive. The common mistake is to claim that reversing the conditioning can give the same answer only when the tail counts are equal.
Why the Estimate Changes
At a fixed threshold, the empirical conditional estimate is a ratio of counts. Its behavior across thresholds follows directly from that ratio: both the joint count and conditioning count can change as observations enter or leave the exceedance set. The ratio may rise, fall, or remain unchanged. There is no rule that it must increase as the threshold rises.
Determine \(u_X(q)\) and \(u_Y(q)\) using a consistent empirical quantile convention, and record how ties are handled.
For an estimate conditional on \(Y\) exceeding its threshold, count \(N_Y(q)\). If this count is zero, that conditional estimate is undefined.
Count \(H(q)\) pairs for which both losses exceed their respective thresholds, then compute \(H(q)/N_Y(q)\).
Repeat for a prespecified sequence of thresholds. Plot the estimates alongside, or annotate them with, their conditioning counts.
The ratio has a simple justification. Among the \(N_Y(q)\) observations satisfying the conditioning event, exactly \(H(q)\) also satisfy the \(X\)-exceedance event. The empirical conditional proportion is therefore the number satisfying both events divided by the number satisfying the condition:
This calculation requires \(N_Y(q)>0\). It also shows why the marginal tail frequency cancels from the empirical conditional proportion. It does not make estimates at different thresholds independent, nor does it remove uncertainty caused by estimating the thresholds from the same sample.
Influential Observations and Rank Changes
An extreme observation can matter in two related ways. With fixed dollar thresholds, it can change the joint count if it crosses a threshold. With empirical quantile thresholds, it can also change which observations occupy the tail ranks. Thus, even if a newly added observation is extreme in only one loss category, it may displace another observation from that category’s selected tail and alter the joint count.
Worked Example: One Extreme Observation Changes a Rank-Based Estimate
In a hypothetical sample, the four largest \(X\) losses are observations \(a,b,c,d\), and the four largest \(Y\) losses are \(a,d,e,f\). At this threshold there are four exceedances in each margin. An additional observation \(g\) is then recorded with an \(X\) loss larger than every existing \(X\) loss and a \(Y\) loss below the selected \(Y\) threshold. Recalculate the overlap estimate using the four largest observations in each margin.
Method. List the selected tail observations in each margin before and after adding \(g\), then count their intersection.
Initially, the \(X\) tail is \(\{a,b,c,d\}\), and the \(Y\) tail is \(\{a,d,e,f\}\). Their intersection is \(\{a,d\}\), so \(H=2\), and the estimate conditional on a \(Y\)-exceedance is \(2/4=0.5000\).
After adding \(g\), the four largest \(X\) losses are \(\{g,a,b,c\}\): \(g\) enters and displaces \(d\). The four largest \(Y\) losses remain \(\{a,d,e,f\}\), because \(g\) is below the \(Y\) threshold. The new intersection is \(\{a\}\), so \(H=1\). The conditional estimate is \(1/4=0.2500\).
Check. Both marginal tail counts remain four, and the joint count falls from two to one, so the estimate falls from 0.5000 to 0.2500. The example isolates a rank-threshold effect: one extreme observation in \(X\) changes which existing observation is included in the empirical \(X\) tail. The common mistake is to assume that a point that does not itself exceed both thresholds cannot affect a rank-based joint estimate.
Practical Interpretation and Exam Traps
A defensible threshold analysis reports the threshold convention, the marginal exceedance counts, the joint count, and the direction of conditioning. If estimates vary sharply, describe the instability and identify whether it is associated with a small denominator, a small number of co-exceedances, ties, or an influential observation. Do not select a threshold solely because it gives the largest estimate or the smoothest-looking plot.
- Do not confuse finite-threshold estimates with a limit. An estimate at the 95th percentile is conditional co-exceedance at that threshold, not automatically the asymptotic tail-dependence coefficient.
- Do not ignore denominators. The same proportion based on three conditioning observations and on 300 observations has very different empirical support.
- Do not reverse conditioning without recalculating. Use the appropriate marginal count as the denominator. The estimates agree if the counts are equal, and also agree at zero when \(H=0\) and both counts are positive.
- Do not treat nearby thresholds as independent replications. Their nested exceedance sets overlap substantially.
- Do not overinterpret a plateau. A stable segment can be a useful diagnostic, but sampling variability, finite-threshold effects, and limited data remain.
For risk management, the distinction matters because an unstable estimate can create unwarranted confidence in how losses may co-occur under extreme conditions. A threshold plot is most useful when it makes that uncertainty visible and prompts scrutiny of the observations driving the result. It complements, rather than replaces, an appropriate dependence model and a broader assessment of the evidence.
Check Your Understanding
For each problem, calculate the requested estimate and interpret it using the counts given.
- At one threshold, a hypothetical loss sample has \(N_X=12\), \(N_Y=12\), and \(H=5\). At a higher threshold it has \(N_X=N_Y=6\) and \(H=3\). Calculate the conditional estimate at each threshold. (Answer: 0.4167 and 0.5000.)
- At a selected threshold, \(N_X=3\), \(N_Y=5\), and \(H=1\). Calculate the probability of an \(X\)-exceedance conditional on a \(Y\)-exceedance, and the reverse conditional estimate. (Answer: 0.2000 and 0.3333, respectively.)
- A threshold plot uses nested marginal counts \(k=25,10,4\), with joint counts \(H=10,6,3\), respectively; the marginal counts are equal at each threshold. Calculate the three estimates and state whether they are constant. (Answer: 0.4000, 0.6000, and 0.7500; they are not constant.)
- At a threshold, the conditional estimate is \(0.25\) from \(k=16\) conditioning exceedances. Calculate the binomial-style standard-error proxy \(\sqrt{\hat p(1-\hat p)/k}\), and state whether it is a formal confidence interval. (Answer: \(\sqrt{0.25(0.75)/16}=0.1083\); it is not a formal confidence interval.)
- Initially, the four largest \(X\) observations are \(\{a,b,c,d\}\), and the four largest \(Y\) observations are \(\{a,d,e,f\}\). A new observation enters the four largest \(X\) set and displaces \(d\), but does not enter the four largest \(Y\) set. What is the conditional estimate given a \(Y\)-exceedance before and after the change? (Answer: \(2/4=0.5000\) before and \(1/4=0.2500\) after.)