Tail Dependence With Signed Risk Variables

Why the Sign of a Risk Variable Matters

Tail dependence is about the chance that extreme outcomes occur together. But “extreme” does not always mean “large.” A large insurance loss is adverse, while a low investment return or a low operating gain may be adverse. The relevant joint event therefore depends on how each variable is defined and signed.

For example, if $X$ is a loss and $R$ is a return, the adverse combination is typically a high value of $X$ with a low value of $R$. Treating both high values as adverse would select the wrong part of the joint distribution. This tutorial makes that translation explicit. The treatment is aligned with the CAS Exam 9 (Risk Management for Actuaries) syllabus.

Definition: For continuous random variables $X$ and $Y$, define their marginal percentile ranks by $U_X=F_X(X)$ and $U_Y=F_Y(Y)$. Each rank is uniformly distributed on $(0,1)$. The upper-tail dependence coefficient is $$ \lambda_U(X,Y)=\lim_{q\uparrow1}P(U_Y>q\mid U_X>q), $$ if the limit exists. The lower-tail dependence coefficient is $$ \lambda_L(X,Y)=\lim_{q\downarrow0}P(U_Y\le q\mid U_X\le q), $$ if the limit exists. These coefficients describe whether extreme ranks in the same direction occur together. For opposite tails, define $$ \lambda_{UL}(X,Y)=\lim_{q\uparrow1}P(U_Y\le 1-q\mid U_X>q), $$ when the limit exists. A finite-threshold conditional probability is not itself an asymptotic coefficient.

The ranks put variables with different units on a common scale. An upper-tail event $U_X>q$ means that $X$ is above its $q$-quantile; a lower-tail event $U_X\le q$ means that it is at or below its $q$-quantile. For a continuous variable, the event $U_X>q$ has probability $1-q$. Using ranks also makes clear that the direction of adversity is a separate choice from the numerical size of a variable.

Translate Adversity Before Calculating

For loss variables, unusually large values are usually adverse, so the relevant event is in the upper tail. For return or gain variables, unusually small values may be adverse, so the relevant event is in the lower tail. When one variable is a loss and the other a return, joint adversity is generally an upper-tail event for the loss and a lower-tail event for the return.

Variable as defined	Typical adverse direction	Adverse tail event at threshold $q$
Loss $X$	Large loss	$U_X>q$
Return $R$	Low return	$U_R\le 1-q$
Gain $G$	Low or negative gain	$U_G\le 1-q$
Loss defined as $L=-G$	Large loss	$U_L>q$

The table describes common interpretations, not universal rules. A variable’s definition and the question being assessed determine which direction is adverse. A high value of a credit score, for example, need not be adverse simply because it is high. State the adverse event in context before choosing a tail coefficient.

Worked Example: A Large Loss With a Low Return

Suppose $X$ is a hypothetical quarterly loss and $R$ a hypothetical quarterly investment return. Their ranks follow this model: with probability $0.40$, $U_R=1-U_X$; with probability $0.60$, $U_R$ is an independent uniform random variable. In both cases $U_X$ is uniform, so both marginal ranks are uniform. Find the probability, conditional on $X$ being in its worst 1% of outcomes, that $R$ is in its lowest 1%. Also find the corresponding asymptotic opposite-tail coefficient.

Method. The adverse event is $U_X>0.99$ and $U_R<0.01$. The model’s first branch makes the ranks move in opposite directions; the independent branch contributes a separate joint probability.

On the first branch, if $U_X>0.99$, then $U_R=1-U_X<0.01$. The joint probability contributed by that branch is $0.40(0.01)=0.004$.

On the independent branch, the two rank events have probabilities $0.01$ and $0.01$, so the branch contributes $0.60(0.01)(0.01)=0.00006$.

The full joint probability is $0.004+0.00006=0.00406$. Since $P(U_X>0.99)=0.01$, the conditional probability is $0.00406/0.01=0.406$, or 40.60%.

For a general threshold $q$, the joint opposite-tail probability is $0.40(1-q)+0.60(1-q)^2$. Dividing by $P(U_X>q)=1-q$ gives $0.40+0.60(1-q)$. As $q\uparrow1$, this tends to $0.40$, so $\lambda_{UL}(X,R)=0.40$. The finite-threshold value at 99% is 0.406, not 0.40.

Check. The conditional probability is between 0 and 1, and its limit, 0.40, is the probability of the branch that creates persistent opposite-tail association. The common mistake avoided here is using upper-tail dependence for both variables just because both outcomes are extreme.

What Reversing a Variable Does

Multiplying a continuous variable by a negative number reverses its ordering. Thus the upper tail of the reversed variable corresponds to the lower tail of the original variable. This is the key translation when a return or gain is rewritten as a loss. In contrast, a strictly increasing transformation preserves the order and therefore preserves percentile ranks.

Define the reversed variable.
Let $Z=-X$, where $X$ is continuous.

Relate the distribution functions.
$F_Z(z)=P(-X\le z)=P(X\ge -z)=1-F_X(-z)$, using continuity to write the last equality.

Relate the percentile ranks.
Substitute $z=-X$: $U_Z=F_Z(-X)=1-F_X(X)=1-U_X$, almost surely.

Translate the tail event.
$U_Z>q$ is equivalent, apart from probability-zero boundaries, to $U_X<1-q$. Thus an upper-tail event for $-X$ is a lower-tail event for $X$.

This argument also works when reversing just one variable in a pair. Reversing both variables switches upper-upper dependence to lower-lower dependence. Reversing only one switches same-tail dependence to opposite-tail dependence. For example, a lower-tail event for a return $R$ becomes an upper-tail event for the loss-style variable $-R$.

Worked Example: Reversing Return Ranks in an Empirical Sample

In a hypothetical sample of 500 quarters, an analyst identifies the 25 quarters with the largest loss $X$ and the 25 quarters with the lowest return $R$. Nine quarters appear in both groups. Among the 25 largest-loss quarters, three instead fall in the 25 highest-return quarters. Calculate the finite-threshold adverse co-occurrence rate and show how the return event changes when $Z=-R$.

Method. There are $25/500=0.05$ observations in each selected tail. The adverse event is the largest-loss group intersected with the lowest-return group, so the conditional denominator is the 25 largest-loss observations.

The finite-threshold conditional rate is $9/25=0.36$, or 36%. The joint sample proportion is $9/500=0.018$; dividing it by the loss-tail proportion $25/500=0.05$ also gives $0.018/0.05=0.36$.

For $Z=-R$, the lowest 5% of $R$ are the highest 5% of $Z$. Therefore the same nine observations are counted by asking for high loss and high $Z$. The three high-return observations do not belong in this adverse count; they are a different return tail.

This is a finite-sample rate at a 5% threshold, not an estimate that automatically equals an asymptotic coefficient. The common mistake avoided is to count the highest returns as adverse or to describe an observed threshold rate as a limiting tail coefficient.

Gains Rewritten as Losses

For a gain variable $G$, the adverse lower tail can be expressed as an upper tail by defining $L=-G$. With two gains $G_1$ and $G_2$, reversing both variables maps joint lower-tail events to joint upper-tail events. The probability of joint adversity is unchanged; only the sign convention and tail labels change.

Worked Example: Two Gains and Their Corresponding Losses

Consider two hypothetical portfolio gains $G_1$ and $G_2$. Let $U_1$ be the rank of $G_1$. With probability $0.25$, set $U_2=U_1$; with probability $0.75$, let $U_2$ be an independent uniform rank. Find the probability that both gains fall in their lowest 2%, conditional on $G_1$ being in its lowest 2%. Then find the asymptotic lower-tail coefficient and translate it to losses $L_i=-G_i$.

Method. The adverse gain event is lower-lower: $U_1<0.02$ and $U_2<0.02$. Calculate each model branch’s contribution to the joint probability and divide by $P(U_1<0.02)=0.02$.

In the branch with probability $0.25$, the ranks are equal. The joint event occurs whenever $U_1<0.02$, so this branch contributes $0.25(0.02)=0.005$.

In the independent branch, both ranks must be below 0.02. The contribution is $0.75(0.02)(0.02)=0.0003$. The joint probability is therefore $0.005+0.0003=0.0053$.

The conditional probability is $0.0053/0.02=0.265$, or 26.50%. At a general lower-tail threshold $p$, the conditional probability is $0.25+0.75p$, which tends to $0.25$ as $p\downarrow0$. Hence $\lambda_L(G_1,G_2)=0.25$.

For $L_i=-G_i$, each upper 2% loss event is the corresponding lower 2% gain event. The finite-threshold conditional probability remains 0.265, and the upper-tail coefficient for the losses is 0.25. The check is direct: changing both signs does not change which pairs of observations are jointly adverse.

Interpretation and Common Traps

Tail coefficients measure limiting conditional co-occurrence, not the size of the losses, the average return, or the correlation at all parts of the distribution. At a chosen finite threshold, an empirical co-exceedance rate can be useful, but it should be reported with its threshold and sample context. It should not be silently substituted for a limiting coefficient.

Before calculating, write down the event in words and then in ranks. For a loss-return pair, that may be “loss above its $q$-quantile and return below its $(1-q)$-quantile,” represented by $U_X>q$ and $U_R\le1-q$. If a variable is reversed, update its tail direction before interpreting a result.

Common exam trap: Treating “extreme” as synonymous with “upper tail” for every variable. For a loss $X$ and a signed return $R$, the calculation $P(U_R>q\mid U_X>q)$ concerns high loss with high return; it does not measure high loss with low return. The correction is to use $P(U_R\le1-q\mid U_X>q)$, or to reverse the return sign and use its upper tail.

Key takeaway: Identify the adverse direction for each variable first. A sign reversal maps lower ranks to upper ranks, and upper ranks to lower ranks. Then state whether the quantity is a finite-threshold conditional rate or an asymptotic tail-dependence coefficient.

Check Your Understanding

For each problem, write the adverse event before calculating. Treat all stated rank distributions as continuous.

A loss $X$ and return $R$ have $P(U_R\le0.03\mid U_X>0.97)=0.28$. What is the joint adverse probability at this threshold, and which return tail is being used? (Answer: Joint probability $0.0084$; the lowest 3% of returns.)
For continuous $Y$, define $Z=-Y$. If the limiting coefficient for $U_X>q$ with $U_Y\le1-q$ is 0.22, what is the corresponding coefficient using $U_Z>q$ with $U_X>q$? (Answer: 0.22, an upper-upper coefficient for $X,Z$.)
In 200 hypothetical periods, 40 have losses above the 80th percentile, and 14 of those also have returns below the 20th percentile. Find the finite-threshold conditional adverse rate and the joint sample proportion. (Answer: Conditional rate $14/40=0.35$; joint proportion $14/200=0.07$.)
Two gain ranks follow a model in which, with probability 0.18, they are equal, and otherwise they are independent uniforms. What is the limiting lower-tail coefficient for the gains, and the corresponding upper-tail coefficient for the losses obtained by reversing both signs? (Answer: Both coefficients are 0.18.)
A high-loss event is paired with an event in the top 4% of a signed return $R$. An analyst wants joint adverse outcomes. Which percentile event should replace the return event, and what event is used if the return is first rewritten as $Z=-R$? (Answer: Use the bottom 4% of $R$; equivalently, use the top 4% of $Z$.)

Variable as defined	Typical adverse direction	Adverse tail event at threshold \(q\)
Loss \(X\)	Large loss	\(U_X>q\)
Return \(R\)	Low return	\(U_R\le 1-q\)
Gain \(G\)	Low or negative gain	\(U_G\le 1-q\)
Loss defined as \(L=-G\)	Large loss	\(U_L>q\)

What You'll Learn