Finite-Threshold Tail Co-Exceedance

Why Finite-Threshold Co-Exceedance Matters

Risk decisions are made at finite thresholds: a capital team may ask how likely one loss is to exceed its 95th percentile when another loss is already above its 95th percentile. That is a different question from asking how dependence behaves in the limit as the threshold approaches the most extreme possible percentile. The finite-threshold probability can be material even when the asymptotic upper-tail dependence coefficient is zero.

For losses with continuous marginal distributions, let \(q\) be a percentile and let \(x_q=F_X^{-1}(q)\) denote the corresponding loss quantile. Exceeding \(x_q\) is equivalent, up to probability-zero boundary events, to having a percentile rank above \(q\). The same reasoning applies to a second loss and its own marginal quantile. This rank representation separates the threshold question from the units and scales of the two losses.

Definition: For two continuous losses \(X\) and \(Y\), their equal-threshold finite-level upper-tail co-exceedance probability at percentile \(q\) is \(P(X>F_X^{-1}(q),\,Y>F_Y^{-1}(q))\). The associated conditional exceedance probability is \(P(X>F_X^{-1}(q)\mid Y>F_Y^{-1}(q))\), which equals the joint probability divided by \(1-q\). The upper-tail dependence coefficient is the limit of this conditional probability as \(q\) approaches 1, when the limit exists.

A joint probability and a conditional probability have different denominators and answer different questions. For example, a joint probability of 0.01 means that both exceedances occur in 1% of all outcomes. If the conditioning loss exceeds its threshold only 5% of the time, the corresponding conditional probability is 0.01 divided by 0.05, or 0.20.

A Rank-Mixture Model

The first model gives a direct calculation and a clear separation between finite and asymptotic behavior. Let \(U\) and \(V\) be percentile ranks, each uniform on \([0,1]\). With probability \(\alpha\), set \(V=U\); with probability \(1-\alpha\), draw \(U\) and \(V\) independently from uniform distributions. The mixture choice and the independent draws are made as specified by this model. Each rank remains uniformly distributed, but the two ranks are not generally independent.

For equal percentile thresholds \(q\), the common-rank part exceeds the threshold with probability \(1-q\), while the independent part exceeds it jointly with probability \((1-q)^2\). Therefore,

P(U>q,V>q)=\alpha(1-q)+(1-\alpha)(1-q)^2.

Since \(P(V>q)=1-q\), dividing the joint probability by that conditioning probability gives

P(U>q\mid V>q)=\alpha+(1-\alpha)(1-q).

The term \(\alpha\) represents the common-rank component that continues to produce joint extremes even at very high thresholds. The other term shrinks as \(q\) approaches 1. Thus the limiting upper-tail dependence coefficient for this model is \(\alpha\), even though the conditional probability at any particular finite threshold also includes the independent component.

Worked Example: A 95th-Percentile Conditional Probability

Suppose two hypothetical loss ranks follow the rank-mixture model with \(\alpha=0.25\). Calculate the joint exceedance probability and the conditional probability that the first rank exceeds its 95th percentile given that the second exceeds its 95th percentile. Compare the conditional probability with the upper-tail dependence coefficient.

Method. First calculate the joint probability by weighting the common-rank and independent components. Then divide by the conditioning event probability.

At \(q=0.95\), each marginal exceedance probability is \(1-q=0.05\). In the common-rank component, the joint exceedance probability is 0.05. In the independent component, it is \(0.05^2=0.0025\). Therefore,

$$ P(U>0.95,V>0.95) =0.25(0.05)+0.75(0.0025) =0.0125+0.001875 =0.014375. $$

The conditioning event \(V>0.95\) has probability 0.05, so

P(U>0.95\mid V>0.95)=\frac{0.014375}{0.05}=0.2875.

The finite-threshold conditional probability is 0.2875, or 28.75%. The asymptotic upper-tail dependence coefficient is \(\alpha=0.25\), or 25%. The finite result is larger because the independent component still contributes \((1-\alpha)(1-q)=0.75(0.05)=0.0375\) to the conditional probability at this threshold. As \(q\) tends to 1, that contribution tends to zero. The calculation checks: multiplying 0.2875 by the conditioning probability 0.05 returns the joint probability 0.014375.

Common mistake avoided: 0.014375 is the probability of both exceedances in the full population, not the probability of the first exceedance conditional on the second. The conditional answer requires division by 0.05.

Different Thresholds and Different Conditioning Events

The thresholds need not be equal. In the same rank-mixture model, suppose the event of interest is \(U>q_1\), while the conditioning event is \(V>q_2\). In the common-rank component, both occur when \(U>\max(q_1,q_2)\). In the independent component, their joint probability is \((1-q_1)(1-q_2)\). Hence

P(U>q_1,V>q_2) =\alpha\bigl(1-\max(q_1,q_2)\bigr) +(1-\alpha)(1-q_1)(1-q_2),

and the conditional probability is this joint probability divided by \(1-q_2\). The denominator is determined by the conditioning event, not by the event whose probability is being predicted.

Worked Example: Unequal Quantiles in the Rank-Mixture Model

For a hypothetical pair of loss ranks, use the rank-mixture model with \(\alpha=0.40\). Find the probability that \(U\) exceeds its 90th-percentile threshold given that \(V\) exceeds its 97th-percentile threshold. Explain why the conditioning denominator is not 0.10.

Method. The thresholds are \(q_1=0.90\) and \(q_2=0.97\). In the common-rank component, both exceedances require \(U>0.97\), with probability 0.03. In the independent component, the joint probability is \(0.10(0.03)=0.003\). Thus,

$$ P(U>0.90,V>0.97) =0.40(0.03)+0.60(0.003) =0.012+0.0018 =0.0138. $$

The conditioning event is \(V>0.97\), whose probability is 0.03. Therefore,

P(U>0.90\mid V>0.97)=\frac{0.0138}{0.03}=0.46.

The answer is 0.46, or 46%. The denominator is 0.03 because the given information is that \(V\) exceeds its 97th-percentile threshold. Using 0.10 would condition on a different event. A check is \(0.46(0.03)=0.0138\), which reproduces the joint probability.

Common mistake avoided: using the exceedance probability for \(U>0.90\) as the denominator. In \(P(A\mid B)\), the denominator is always \(P(B)\), here the probability of the 97th-percentile exceedance by \(V\).

A Model With Zero Upper-Tail Dependence

A zero asymptotic coefficient does not imply that conditional exceedance is negligible at every practical threshold. The following explicitly specified joint distribution of percentile ranks illustrates the distinction. For a parameter \(\theta>0\), let the joint cumulative distribution function of \(U,V\) be

C(u,v)=\left(u^{-\theta}+v^{-\theta}-1\right)^{-1/\theta}, \qquad 0<u,v\leq 1.

This model has uniform marginal ranks: substituting \(v=1\) gives \(C(u,1)=u\), and substituting \(u=1\) gives \(C(1,v)=v\). For equal thresholds \(q\), inclusion-exclusion gives the joint exceedance probability \(1-2q+C(q,q)\). Therefore,

P(U>q\mid V>q) =\frac{1-2q+(2q^{-\theta}-1)^{-1/\theta}}{1-q}.

This formula is for the specified joint rank model; its use requires \(\theta>0\), uniform margins, and the stated dependence structure. It is not a formula for an arbitrary pair of losses. Because the ranks are continuous, whether strict or non-strict inequalities are used at a threshold does not change these probabilities.

Justification of the asymptotic limit: Put \(s=1-q\), so \(s\) decreases to zero as the threshold approaches 1. A second-order expansion gives \(C(q,q)=1-2s+(1+\theta)s^2+o(s^2)\). Thus the joint exceedance probability \(1-2q+C(q,q)\) is \((1+\theta)s^2+o(s^2)\). Dividing by \(P(V>q)=s\) gives \((1+\theta)s+o(s)\), which tends to zero. The upper-tail dependence coefficient is therefore zero for every fixed \(\theta>0\), despite potentially substantial conditional probabilities at finite thresholds.

Worked Example: Substantial Finite Exceedance With a Zero Limit

Suppose two hypothetical loss ranks follow the specified joint rank model with \(\theta=4\). Calculate their equal-threshold conditional exceedance probabilities at the 80th and 95th percentiles, then compare both with the asymptotic coefficient.

Method. At each percentile, calculate \(C(q,q)=(2q^{-4}-1)^{-1/4}\), use inclusion-exclusion to obtain the joint exceedance probability, and divide by the marginal exceedance probability \(1-q\).

At \(q=0.80\), \(0.80^{-4}=2.4414\) to four decimals, so \(2q^{-4}-1=3.8828\) to four decimals and \(C(0.80,0.80)\approx0.7124\). The joint exceedance probability is \(1-2(0.80)+0.7124=0.1124\), and the conditional probability is \(0.1124/0.20\approx0.5619\).

At \(q=0.95\), \(0.95^{-4}\approx1.2277\), so \(2q^{-4}-1\approx1.4555\) and \(C(0.95,0.95)\approx0.9104\). The joint exceedance probability is \(1-2(0.95)+0.9104\approx0.0104\), and the conditional probability is \(0.0104/0.05\approx0.2087\). Calculations use unrounded intermediate expressions; probabilities are rounded to four decimal places.

At the 80th percentile, the conditional probability is about 56.19%; at the 95th percentile, it is about 20.87%. Both are materially above zero, yet the asymptotic coefficient is zero. As the threshold moves still closer to 1, the conditional probability tends toward zero. The check at each threshold is that the conditional probability multiplied by \(1-q\) recovers the joint exceedance probability, subject to the displayed rounding.

Common mistake avoided: declaring that the 95th-percentile conditional probability must be zero because the asymptotic coefficient is zero. The coefficient is a limit, not a value assigned to every finite percentile.

Interpretation for Risk Aggregation

Finite-threshold co-exceedance is useful for describing how often losses simultaneously enter a specified high-loss region. It does not, by itself, determine an aggregate capital measure: the loss amounts above those thresholds and the full joint distribution also matter. Nor is a single finite-threshold value a universal summary of dependence. The model, threshold, marginal quantiles, and conditioning event all affect its interpretation.

When comparing a finite conditional probability with an asymptotic coefficient, keep the following distinctions explicit:

State the threshold. “At the 95th percentile” is a finite-level statement; “as the percentile tends to 1” describes a limit.
State the event and denominator. Identify which loss exceeds which quantile and which exceedance is conditioned upon.
Separate marginal and joint behavior. A joint probability can be small simply because the marginal exceedance events are rare; conditioning rescales by the probability of the conditioning event.
Use a model appropriate to the question. The formulas above follow from their stated rank models and should not be transferred to a different dependence model without justification.

This distinction complements the earlier material on upper and lower tail dependence coefficients and on tail dependence in the bivariate t distribution. A coefficient describes limiting behavior; finite-threshold calculations describe behavior at particular quantiles relevant to a stated analysis. Both can be informative, but they answer different questions.

Check Your Understanding

For each problem, identify the joint or conditional probability requested before calculating. Probabilities are rounded to four decimals where needed.

In the rank-mixture model, let \(\alpha=0.30\) and \(q=0.90\). Calculate the joint equal-threshold exceedance probability, the conditional exceedance probability, and the upper-tail dependence coefficient. (Answer: joint \(=0.0370\); conditional \(=0.3700\); coefficient \(=0.3000\))
In the rank-mixture model, let \(\alpha=0.50\). Calculate \(P(U>0.90\mid V>0.95)\). Show which threshold determines the denominator. (Answer: \(0.5500\))
For the specified joint rank model with \(\theta=1\), calculate the conditional exceedance probability at \(q=0.50\), then state its asymptotic upper-tail coefficient. (Answer: finite conditional probability \(=0.6667\); coefficient \(=0\))
For the specified joint rank model with \(\theta=2\), calculate the equal-threshold conditional exceedance probability at \(q=0.80\). Use \(C(q,q)=(2q^{-2}-1)^{-1/2}\). (Answer: \(0.4300\))
In the rank-mixture model with \(\alpha=0.20\), the equal-threshold conditional probability is observed to be 0.24 at a specified percentile \(q\). Solve for \(q\), then state the model's asymptotic coefficient. (Answer: \(q=0.95\); coefficient \(=0.2000\))

What You'll Learn