Sensitivity to Tail Dependence in Aggregate Losses

Why Tail Dependence Matters

Two risks can have the same marginal loss distributions and the same ordinary correlation, yet differ in how often they produce large losses together. That distinction can matter greatly to an insurer: aggregate capital is often driven by the joint upper tail, not by average co-movement across all outcomes. A dependence model that matches ordinary correlation may therefore still give an inadequate description of extreme aggregate losses.

This tutorial compares aggregate tail measures when the component margins and ordinary correlation are held fixed while joint upper-tail behavior changes. The examples use explicit probability tables so the effect can be calculated directly. Copulas provide a general framework for making the same distinction with continuous loss distributions.

Definition: For continuous marginal distributions, the copula $C$ describes the dependence structure of risks after each component has been transformed to a uniform variable on $[0,1]$. The upper-tail dependence coefficient is $$ \lambda_U=\lim_{u\to 1^-}P(U>u\mid V>u) =\lim_{u\to 1^-}\frac{1-2u+C(u,u)}{1-u}, $$ when the limit exists. A positive $\lambda_U$ means that the conditional probability of one risk being extreme, given that the other is extreme at the same percentile threshold, does not vanish as the threshold approaches the maximum. Ordinary correlation instead summarizes linear co-movement across the full distribution.

For a Gaussian copula with latent correlation $\rho$ satisfying $-1<\rho<1$, $\lambda_U=0$. For a Student $t$ copula with degrees of freedom $\nu$ and latent correlation $\rho$, upper- and lower-tail dependence are positive and equal to $$ \lambda_U=\lambda_L=2t_{\nu+1}\left(-\sqrt{\frac{(\nu+1)(1-\rho)}{1+\rho}}\right), $$ where $t_{\nu+1}$ is the cumulative distribution function of a standard Student $t$ variable with $\nu+1$ degrees of freedom. These are properties of the copulas, not universal formulas for the Pearson correlation of the resulting losses.

In practice, changing copula families while keeping the same latent parameter does not generally keep the Pearson correlation of the losses fixed. That correlation depends on both the copula and the component margins. A fair model comparison should therefore check the resulting loss correlation and, when the purpose is to isolate tail behavior, calibrate parameters so that ordinary loss correlation is approximately matched. The finite discrete examples below make the comparison exact.

Same Margins and Correlation, Different Joint Extremes

Consider two component losses $X$ and $Y$, measured in thousands of dollars. Each has the same marginal distribution: it takes values $0$, $1$, and $10$, each with probability $1/3$. Under an independent joint distribution, every cell of the joint probability table has probability $1/9$. Now consider a second joint distribution formed by adding the following adjustments to those nine probabilities:

\frac{1}{1000} \begin{pmatrix} -81&81&0\\ 81&-80&-1\\ 0&-1&1 \end{pmatrix}, \qquad \text{rows and columns correspond to }0,1,10.

The adjustments leave every row sum and column sum unchanged, so both component margins remain the same. Each adjusted probability is nonnegative; the smallest is $1/9-81/1000=0.0301$ to four decimal places. The probability of the joint top outcome, $X=10,Y=10$, increases from $1/9$ to $1/9+0.001$.

The covariance also remains unchanged. The base table is independent, so its covariance is zero. The adjustments that affect the covariance can be grouped into two $2\times2$ changes. The change on the $0,1$ by $0,1$ block contributes $-81(1-0)(1-0)/1000=-0.081$. The change on the $1,10$ by $1,10$ block contributes $1(10-1)(10-1)/1000=0.081$. They cancel. Since each component variance is positive and unchanged, the correlation remains zero as well.

Key result: Matching the marginal distributions and ordinary correlation does not determine the joint upper tail. In this example the probability of both losses being at their maximum changes, even though the margins and correlation are exactly preserved.

Worked Example: A Small Dependence Change, a Different Aggregate VaR

A hypothetical portfolio has component losses $X$ and $Y$, each with the three-point marginal distribution just described. Compare independent losses with the adjusted joint distribution. Find the aggregate VaR at confidence level $88.8\%$, where $S=X+Y$. Use the convention $\operatorname{VaR}_{\alpha}(S)=\inf\{s:F_S(s)\geq\alpha\}$.

Method. First group the nine joint outcomes by their aggregate loss. Under independence, the possible sums $0,1,2,10,11,20$ have probabilities $1/9,2/9,1/9,2/9,2/9,1/9$, respectively. For example, the sum $11$ occurs for $(1,10)$ and $(10,1)$, so its probability is $2/9$.

Under the adjusted distribution, the probability changes follow directly from the table. Sum $0$ changes by $-0.081$. Sum $1$ changes by $0.162$, from the two cells $(0,1)$ and $(1,0)$. Sum $2$ changes by $-0.080$. Sum $10$ is unchanged. Sum $11$ changes by $-0.002$, and sum $20$ changes by $0.001$. Thus the aggregate probabilities, rounded to four decimals, are:

Aggregate loss $S$, in thousands	Independent probability	Adjusted probability
0	0.1111	0.0301
1	0.2222	0.3842
2	0.1111	0.0311
10	0.2222	0.2222
11	0.2222	0.2202
20	0.1111	0.1121

For independent losses, the cumulative probability through $11$ is $8/9=0.8889$, while the cumulative probability through $10$ is $6/9=0.6667$. Since $0.6667<0.8880\leq0.8889$, the $88.8\%$ VaR is $11$ thousand dollars.

For the adjusted losses, the cumulative probability through $11$ is $8/9-0.001=0.8879$, while the cumulative probability through $20$ is $1$. Therefore $0.8879<0.8880$, and the $88.8\%$ VaR is $20$ thousand dollars. The dependence change shifts this discrete quantile by $9$ thousand dollars, despite leaving the margins and ordinary correlation unchanged.

Check and common mistake. The adjusted probability changes sum to zero: $-0.081+0.162-0.080-0.002+0.001=0$. A common mistake is to round the cumulative probability $0.8879$ to $0.888$ before comparing it with the confidence level. The exact cumulative value is below $0.8880$, so the adjusted VaR is $20$, not $11$, thousand dollars.

Why the Covariance Can Stay Fixed

The table construction has a general explanation. For a two-by-two probability adjustment with entries $+c,-c,-c,+c$ on adjacent rows with loss values $x_i,x_{i+1}$ and adjacent columns with values $y_j,y_{j+1}$, the change in $E[XY]$ is $$ c(x_{i+1}-x_i)(y_{j+1}-y_j). $$ To see why, expand the four contributions: $$ c x_i y_j-c x_i y_{j+1}-c x_{i+1}y_j+c x_{i+1}y_{j+1} =c(x_{i+1}-x_i)(y_{j+1}-y_j). $$ The row and column adjustments sum to zero, so marginal means do not change. Consequently, the covariance change equals the change in $E[XY]$. In the earlier table, one block contributes $-81/1000$ and the other contributes $81/1000$; the total covariance change is zero even though the upper-right corner probability changes.

This argument also explains why matching a single dependence statistic cannot identify all features of a joint distribution. Covariance is a weighted average over all outcomes. Opposing changes in different parts of the table can cancel in that average while still changing the frequency of joint extremes.

Worked Example: TVaR Responds to the Added Joint Tail Probability

Use the same hypothetical portfolio and compare the two dependence structures at confidence level $80\%$. Calculate TVaR using the definition $$ \operatorname{TVaR}_{\alpha}(S)=\operatorname{VaR}_{\alpha}(S)+\frac{E[(S-\operatorname{VaR}_{\alpha}(S))_+]}{1-\alpha}, $$ where $z_+=\max(z,0)$. This definition handles discrete losses, including probability mass at the VaR threshold.

For both structures, the cumulative probability through $10$ is less than $0.80$, while the cumulative probability through $11$ is greater than $0.80$. Thus the $80\%$ VaR is $11$ thousand dollars in each case.

Only an aggregate loss of $20$ exceeds $11$, and its excess is $20-11=9$ thousand dollars. Under independence, $$ E[(S-11)_+]=9\left(\frac{1}{9}\right)=1 $$ thousand dollars, so $$ \operatorname{TVaR}_{0.80}(S)=11+\frac{1}{0.20}=16 $$ thousand dollars.

Under the adjusted dependence structure, the probability of $S=20$ is $1/9+0.001=0.1121$ to four decimal places. Using the unrounded probability, $$ E[(S-11)_+]=9\left(\frac{1}{9}+0.001\right)=1.009 $$ thousand dollars. Therefore $$ \operatorname{TVaR}_{0.80}(S)=11+\frac{1.009}{0.20}=16.045 $$ thousand dollars, or $\$16{,}045.00$.

The VaR at this confidence level does not change, but TVaR increases by $\$45.00$. This illustrates that a tail measure can respond to a changed extreme-loss probability even when the quantile lies below the extreme outcome.

Check and common mistake. The excess above VaR is $9$, not $20$, because TVaR uses the loss above the threshold in the numerator. Also, for a discrete distribution, TVaR should not automatically be calculated as $E[S\mid S\geq\operatorname{VaR}_{\alpha}(S)]$; the integrated-tail formula above correctly accounts for the fraction of probability mass needed at the quantile.

Interpreting Tail Dependence in Practice

For continuous copulas, $\lambda_U$ describes limiting behavior as the percentile threshold tends to one. For a discrete model or a finite data set, an asymptotic limit may not be the most useful diagnostic. A practical finite-threshold quantity is the conditional exceedance probability $P(X>x\mid Y>y)$, with thresholds chosen to represent material losses. This is a threshold-specific measure, not necessarily the copula tail-dependence coefficient.

Worked Example: Tail Co-Exceedance With Zero Correlation

A second hypothetical portfolio has component losses $X$ and $Y$, each taking values $0,2,12$ with probability $1/3$. Begin with independent components, then adjust their joint probabilities by $$ \frac{1}{1000} \begin{pmatrix} -25&25&0\\ 25&-24&-1\\ 0&-1&1 \end{pmatrix}. $$ Rows and columns correspond to $0,2,12$. Compare the probability of both components exceeding $2$, and verify that the ordinary correlation remains zero.

The row and column sums of the adjustment matrix are all zero, so the marginal probabilities remain $1/3$ at each loss value. The probability of $X>2,Y>2$ is the probability of the $(12,12)$ cell. It rises from $1/9=0.1111$ to $1/9+0.001=0.1121$, using four decimal places. Since $P(Y>2)=1/3$, the conditional exceedance probability is $$ P(X>2\mid Y>2)=\frac{1/9+0.001}{1/3} =\frac{1}{3}+0.003 =0.3363 $$ to four decimal places, compared with $1/3=0.3333$ under independence.

For the covariance check, the lower block adjustment contributes $-25(2-0)(2-0)/1000=-0.100$. The upper block contributes $1(12-2)(12-2)/1000=0.100$. These cancel, and the independent base covariance is zero. The component variance is positive: $E[X]=14/3$, $E[X^2]=148/3$, and $\operatorname{Var}(X)=148/3-(14/3)^2=248/9$. Thus the correlation is zero under both distributions.

Check and common mistake. The increase in the conditional exceedance probability is $0.003$, not $0.001$, because the joint probability change is divided by $P(Y>2)=1/3$. Zero correlation does not imply independence or equal tail co-exceedance probabilities.

Model Comparison and Exam Traps

When comparing dependence models for aggregate capital, keep the following distinctions explicit:

Margins versus dependence: verify that each component distribution is unchanged before attributing an aggregate difference to dependence.
Loss correlation versus copula parameter: a latent Gaussian or $t$-copula correlation is not automatically the Pearson correlation of transformed losses. Check the latter after applying the marginal distributions.
Ordinary co-movement versus joint extremes: matching covariance does not guarantee matching high-threshold conditional exceedance probabilities.
Quantile thresholds versus tail averages: a small change can move a discrete VaR across an outcome boundary, while TVaR can change even if VaR stays fixed.
Finite-threshold diagnostics versus asymptotic dependence: empirical or modeled exceedance probabilities at a selected threshold do not by themselves establish a limiting tail-dependence coefficient.

The comparison should be interpreted as a sensitivity analysis, not proof that one family is always preferable. A Gaussian copula can be appropriate in some applications, and a $t$ copula can be inappropriate if its joint-tail behavior is unsupported by the risk process. Model choice should reflect the available evidence and the purpose of the analysis. The key actuarial question is whether plausible changes in joint extremes materially alter aggregate VaR, TVaR, or capital conclusions.

Key takeaway: Ordinary correlation describes one aspect of dependence, not the full joint distribution. To understand aggregate tail risk, compare models with margins and ordinary correlation held fixed as far as practicable, then examine joint exceedances and aggregate tail measures directly.

Check Your Understanding

Work through each problem before checking its answer.

A hypothetical pair of losses each takes values $0,2,12$ with equal probability. A joint probability adjustment is $0.0005$ times the matrix $\begin{pmatrix}-9&9&0\\9&-8&-1\\0&-1&1\end{pmatrix}$, with rows and columns in increasing loss order. Find the change in the probability of both losses equaling $12$, and verify whether covariance changes. (Answer: The joint top probability increases by $0.0005$; covariance does not change, since the lower-block contribution is $-9(2)(2)(0.0005)=-0.018$ and the upper-block contribution is $1(10)(10)(0.0005)=0.050$.)
For the $0,1,10$ portfolio in the first worked example, calculate the independent aggregate cumulative probability through $10$. Is the $80\%$ VaR equal to $10$ or $11$ thousand dollars? (Answer: The cumulative probability is $6/9=0.6667$, so the $80\%$ VaR is $11$ thousand dollars.)
For the adjusted $0,1,10$ portfolio, use the $80\%$ VaR of $11$ and the top aggregate probability $1/9+0.001$. Calculate TVaR to the nearest dollar. (Answer: $11+9(1/9+0.001)/0.20=16.045$ thousand dollars, or $\$16{,}045.00$.)
A Gaussian copula and a Student $t$ copula have been fitted to the same component margins. An analyst uses the same numerical latent correlation in both and concludes that loss correlation is therefore matched and tail risk must be similar. Identify both errors. (Answer: The same latent parameter does not in general produce the same Pearson correlation after applying the margins; the loss correlations must be checked or recalibrated. Also, for a nondegenerate Gaussian copula upper-tail dependence is zero, while a Student $t$ copula has positive upper-tail dependence for finite degrees of freedom and admissible correlation, so similar ordinary correlation does not establish similar joint-tail risk.)

What You'll Learn