Diversification Under Stress Conditions

Why Stress Dependence?

Risks that appear only weakly related in ordinary conditions may move together when a common economic, operational or catastrophe driver becomes severe. If aggregate capital relies on ordinary-period dependence estimates, it can therefore understate the risk of simultaneous adverse outcomes. A dependence stress asks a focused question: how would aggregate risk change if relationships among losses were more adverse, while the marginal loss assumptions were held fixed?

That comparison is meaningful only when the assumptions are explicit. A higher correlation is not itself a complete stress model: it specifies a change in linear association, but it does not uniquely determine the joint tail behavior of non-normal losses. The earlier material on tail dependence addresses the importance of that distinction. Here the emphasis is on recalculating aggregate risk under a stated stressed dependence model and comparing its diversification benefit with the ordinary-period estimate.

Definition: A dependence stress is a specified change to the joint distribution of two or more losses, made to assess how adverse dependence assumptions affect aggregate loss or capital. A stress may alter correlations, joint-event probabilities, or another feature of the dependence model. Unless stated otherwise, a comparison should hold marginal distributions, confidence level, time horizon, and loss definitions fixed.

Aggregate Risk and Diversification Benefit

Let \(X_1,\ldots,X_n\) be losses measured on the same basis, and let \(L=\sum_{j=1}^n X_j\) be their aggregate. For two losses with finite variances,

\operatorname{Var}(X_1+X_2) =\operatorname{Var}(X_1)+\operatorname{Var}(X_2) +2\operatorname{Cov}(X_1,X_2).

If their standard deviations are \(\sigma_1,\sigma_2\) and their correlation is \(\rho\), then \(\operatorname{Cov}(X_1,X_2)=\rho\sigma_1\sigma_2\). Thus, changing the assumed correlation changes aggregate variance even when both individual loss distributions remain unchanged.

\operatorname{Var}(X_1+X_2) =\sigma_1^2+\sigma_2^2+2\rho\sigma_1\sigma_2.

For a confidence level \(p\), define the value-at-risk by \(\operatorname{VaR}_p(X)=\inf\{x:F_X(x)\ge p\}\). A commonly used comparison is the VaR diversification benefit:

\text{Diversification benefit at level }p =\sum_{j=1}^n\operatorname{VaR}_p(X_j) -\operatorname{VaR}_p\!\left(\sum_{j=1}^n X_j\right).

This definition compares the sum of standalone VaRs with aggregate VaR at the same confidence level and on the same basis. A positive value means aggregate VaR is below the sum of standalone VaRs; a zero or negative value is possible. It is a measure under the chosen model, not a guarantee that diversification will persist under stress.

Important condition: If the losses are jointly normal, their sum is normal, and its mean and variance determine its VaR. Specifically, if \(L\sim N(\mu_L,\sigma_L^2)\), then \(\operatorname{VaR}_p(L)=\mu_L+z_p\sigma_L\), where \(z_p\) is the \(p\)-quantile of the standard normal distribution. Outside a specified distributional model, aggregate variance and correlation alone do not determine VaR.

Why Covariance Enters: A Short Derivation

For finite-variance losses, variance is the expected squared deviation from the mean. Write \(X_1-\operatorname{E}[X_1]\) and \(X_2-\operatorname{E}[X_2]\) for their centered values. The centered aggregate is their sum. Squaring it produces two individual squared terms and a cross-product:

\begin{aligned} \operatorname{Var}(X_1+X_2) &=\operatorname{E}\!\left[ \big((X_1-\operatorname{E}[X_1])+(X_2-\operatorname{E}[X_2])\big)^2 \right]\\ &=\operatorname{Var}(X_1)+\operatorname{Var}(X_2) +2\operatorname{E}\!\left[ (X_1-\operatorname{E}[X_1])(X_2-\operatorname{E}[X_2]) \right]\\ &=\operatorname{Var}(X_1)+\operatorname{Var}(X_2) +2\operatorname{Cov}(X_1,X_2). \end{aligned}

The means do not appear in this variance formula: centering has removed them. The cross-product is exactly why a dependence assumption matters. Positive covariance adds to aggregate variance, while negative covariance reduces it.

Worked Example: Two Normal Losses Under a Correlation Stress

A hypothetical insurer models two annual losses as jointly normal. Loss \(X\) has mean \(40\) thousand dollars and standard deviation \(20\) thousand dollars; loss \(Y\) has mean \(30\) thousand dollars and standard deviation \(15\) thousand dollars. Compare the aggregate 95% VaR and VaR diversification benefit when the correlation is \(0\) under the ordinary-period model and \(0.60\) under stress. Use \(z_{0.95}=1.645\). Round money to the nearest \(0.01\) thousand dollars at the end.

Method and assumptions. The marginal distributions and means are held fixed, and each dependence assumption is represented by a jointly normal model with the specified correlation. Therefore the aggregate is normal. Calculate variance using standard deviations and correlation, not the losses' means.

Step 1: Calculate the aggregate mean. Under either dependence assumption, the mean is \(40+30=70\) thousand dollars.

Step 2: Calculate ordinary-period aggregate variance and VaR. At correlation \(0\), the variance is \(20^2+15^2+2(0)(20)(15)=625\), in squared thousands of dollars. The standard deviation is \(\sqrt{625}=25\). Thus the aggregate VaR is \(70+1.645(25)=111.125\), or \(111.13\) thousand dollars.

Step 3: Calculate stressed aggregate variance and VaR. At correlation \(0.60\), the variance is \(20^2+15^2+2(0.60)(20)(15)=400+225+360=985\). The standard deviation is \(\sqrt{985}=31.3847\ldots\). The aggregate VaR is \(70+1.645\sqrt{985}=121.6278\ldots\), or \(121.63\) thousand dollars.

Step 4: Compare diversification benefits. The standalone VaRs are \(40+1.645(20)=72.900\) and \(30+1.645(15)=54.675\) thousand dollars. Their unrounded sum is \(127.575\). The ordinary-period benefit is \(127.575-111.125=16.450\), or \(16.45\) thousand dollars. The stressed benefit is \(127.575-121.6278\ldots=5.9472\ldots\), or \(5.95\) thousand dollars.

Check. The stressed dependence raises aggregate VaR by \(121.6278\ldots-111.125=10.5028\ldots\) thousand dollars. The diversification benefit falls by the same amount, from \(16.45\) to \(5.95\) thousand dollars, subject to rounding.

Common mistake avoided: Adding the means \(40\) and \(30\) into the variance would be incorrect. Means determine the aggregate mean; variance uses the squared standard deviations and the covariance term.

Discrete Joint Losses: Stress the Joint Probability

A dependence stress need not be expressed as a correlation. For discrete losses, it may be more direct to specify a stressed joint-event probability. The marginal probabilities constrain which joint probabilities are possible. If \(A\) and \(B\) are loss events, then

P(A\text{ only})=P(A)-P(A\cap B),\qquad P(B\text{ only})=P(B)-P(A\cap B),

P(\text{neither})=1-P(A)-P(B)+P(A\cap B).

These four outcomes must have nonnegative probabilities summing to one. With equal marginal event probabilities \(q\), the joint probability \(r=P(A\cap B)\) must satisfy \(\max(0,2q-1)\le r\le q\).

Worked Example: A Discrete Catastrophe-Linked Loss Stress

In a hypothetical portfolio, each of two lines has an annual loss of \(40\) thousand dollars if its own event occurs and zero otherwise. Each event has probability \(0.08\). The ordinary-period model assigns a \(0.01\) probability to both events occurring; a stress scenario raises that joint probability to \(0.04\), without changing either marginal probability. Find the probability of exactly one event and of neither event under each model, then compare aggregate 95% VaR. Round probabilities to four decimals and money to the nearest \(0.01\) thousand dollars.

Method. Reconstruct the four joint outcomes from the two marginals and the specified joint probability. For aggregate VaR, use the discrete definition as the smallest loss amount whose cumulative probability reaches or exceeds \(0.95\).

Ordinary-period model. The probability of exactly one event is \((0.08-0.01)+(0.08-0.01)=0.14\). The probability of neither is \(1-0.08-0.08+0.01=0.85\). The probability of both is \(0.01\). These probabilities sum to \(0.14+0.85+0.01=1.00\), where the exactly-one probability includes both mutually exclusive one-event outcomes. The aggregate loss is \(0\) with probability \(0.85\), \(40\) with probability \(0.14\), and \(80\) with probability \(0.01\), in thousands. Its cumulative probability at \(40\) is \(0.85+0.14=0.99\), so the 95% VaR is \(40\) thousand dollars.

Stressed model. The probability of exactly one event is \((0.08-0.04)+(0.08-0.04)=0.08\). The probability of neither is \(1-0.08-0.08+0.04=0.88\), and the probability of both is \(0.04\). Again, \(0.88+0.08+0.04=1.00\). The cumulative probability at \(40\) is \(0.88+0.08=0.96\), so the stressed 95% VaR is also \(40\) thousand dollars.

Check and interpretation. The stress raises the chance of an \(80\)-thousand-dollar aggregate loss from \(0.0100\) to \(0.0400\), while the 95% VaR remains \(40\) thousand dollars. Each standalone 95% VaR is \(40\) thousand dollars, so the VaR diversification benefit is \(80-40=40\) thousand dollars in both cases. This example shows that a material change in joint outcomes need not change VaR at a particular confidence level: the quantile depends on whether cumulative probability crosses that level.

Common mistake avoided: The probability of neither event is not \(1-P(A)-P(B)\) when joint occurrence is possible. Inclusion-exclusion requires adding \(P(A\cap B)\).

Multiple Risks and Valid Stress Matrices

For several risks, a correlation stress must form a mathematically valid correlation matrix. In particular, the matrix must be positive semidefinite; otherwise it cannot be the correlation matrix of any random vector. Pairwise correlations each lying between \(-1\) and \(1\) is necessary but not sufficient. For three risks with correlations \(r_{12},r_{13},r_{23}\), the determinant condition is

1+2r_{12}r_{13}r_{23}-r_{12}^2-r_{13}^2-r_{23}^2\ge 0.

When the losses are jointly normal, the aggregate variance is the sum of individual variances plus twice the sum of distinct pairwise covariances. A valid matrix is therefore needed before using the assumed correlations to calculate aggregate risk.

Worked Example: Three-Risk Normal Aggregation Under Stress

A hypothetical insurer models three annual losses as jointly normal, in thousands of dollars. Their means are \(10,20,30\), and their standard deviations are \(8,6,10\), respectively. The ordinary-period correlations \((r_{12},r_{13},r_{23})\) are \((0.10,0.20,0.30)\); the stressed correlations are \((0.60,0.70,0.50)\). Compare the 95% aggregate VaR and diversification benefit using \(z_{0.95}=1.645\). Round monetary results to \(0.01\) thousand dollars.

Step 1: Check the correlation matrices. The ordinary determinant is \(1+2(0.10)(0.20)(0.30)-0.10^2-0.20^2-0.30^2=0.872\). The stressed determinant is \(1+2(0.60)(0.70)(0.50)-0.60^2-0.70^2-0.50^2=0.320\). Both are positive, and all pairwise correlations have magnitude below one; the corresponding three-risk correlation matrices are positive definite and valid.

Step 2: Find ordinary-period aggregate variance. The sum of individual variances is \(8^2+6^2+10^2=64+36+100=200\). The three covariances are \((0.10)(8)(6)=4.8\), \((0.20)(8)(10)=16\), and \((0.30)(6)(10)=18\). Therefore aggregate variance is \(200+2(4.8+16+18)=277.6\). Its standard deviation is \(\sqrt{277.6}=16.6613\ldots\). The aggregate mean is \(10+20+30=60\), so ordinary-period VaR is \(60+1.645\sqrt{277.6}=87.4079\ldots\), or \(87.41\) thousand dollars.

Step 3: Find stressed aggregate variance and VaR. The stressed covariances are \((0.60)(8)(6)=28.8\), \((0.70)(8)(10)=56\), and \((0.50)(6)(10)=30\). Aggregate variance is \(200+2(28.8+56+30)=429.6\), with standard deviation \(\sqrt{429.6}=20.7268\ldots\). Stressed aggregate VaR is \(60+1.645\sqrt{429.6}=94.0956\ldots\), or \(94.10\) thousand dollars.

Step 4: Compare benefits. The sum of standalone 95% VaRs is \(60+1.645(8+6+10)=99.48\) thousand dollars. The ordinary-period benefit is \(99.48-87.4079\ldots=12.0721\ldots\), or \(12.07\) thousand dollars. The stressed benefit is \(99.48-94.0956\ldots=5.3844\ldots\), or \(5.38\) thousand dollars.

Check. The stress increases aggregate variance by \(429.6-277.6=152.0\) squared thousands of dollars and increases aggregate VaR by approximately \(6.69\) thousand dollars. The benefit decreases by the same VaR difference, subject to rounding.

Common mistake avoided: Checking each correlation separately is not enough to establish that a multivariate stress is feasible. The joint matrix must also satisfy positive semidefiniteness.

Applying Stress Results Responsibly

A stressed result is conditional on both the stress design and the aggregation model. In the normal examples, increasing correlation raises aggregate variance and VaR because all standard deviations are positive and the normal VaR increases with aggregate standard deviation. That conclusion does not establish that every dependence stress can be summarized by correlation, or that correlation alone captures tail co-movement for a non-normal model.

For a useful comparison, document the ordinary and stressed assumptions side by side. State which marginal distributions remain unchanged, what dependence feature is altered, whether the stressed model is feasible, and which risk measure and confidence level are being compared. If the marginals also change, the result combines marginal and dependence stresses; it should not be described as the isolated effect of dependence.

The practical value of this analysis is not simply to produce a larger capital number. It reveals how much an aggregation result depends on ordinary-period relationships and whether a reported diversification benefit is robust to plausible adverse dependence. A stress scenario is a sensitivity assessment, not automatically a forecast or a probability-weighted estimate of future outcomes.

Key takeaway: Recalculate the joint model under the stressed dependence assumption, then recompute aggregate risk and diversification benefit on a consistent basis. Do not infer tail risk from correlation alone, omit covariance terms, or use an invalid correlation matrix.

Check Your Understanding

Use the definitions and methods above. Report probabilities to four decimals and monetary amounts to \(0.01\) thousand dollars where appropriate.

Two jointly normal losses have means \(12\) and \(18\), standard deviations \(3\) and \(4\), and correlation \(0.25\). Under stress the correlation is \(0.75\), with all other assumptions fixed. Using \(z_{0.95}=1.645\), calculate aggregate VaR under both assumptions. (Answer: Ordinary variance \(31\), VaR \(39.16\); stressed variance \(43\), VaR \(40.79\), in the loss units.)
Each of two hypothetical risks produces a loss of \(25\) thousand dollars with probability \(0.12\), and otherwise produces zero. The ordinary joint-loss probability is \(0.04\), rising to \(0.09\) under stress. Find the probability of exactly one loss and the 95% aggregate VaR under each model. (Answer: Exactly one is \(0.16\) ordinary and \(0.06\) stressed; VaR is \(25.00\) thousand dollars ordinary and \(50.00\) thousand dollars stressed.)
Three jointly normal losses have standard deviations \(5,7,9\) and correlations \(r_{12}=0.80\), \(r_{13}=0.80\), and \(r_{23}=0.30\). Check the three-risk determinant condition and calculate aggregate variance. (Answer: Determinant \(0.014\), so the matrix is positive definite; aggregate variance \(320.80\) squared loss units.)
Two event losses have marginal event probabilities \(0.10\) and \(0.10\). Their ordinary joint-event probability is \(0.01\), and their stressed joint-event probability is \(0.06\). For each assumption, calculate the probability of exactly one event and the probability of neither event. (Answer: Ordinary exactly one \(0.18\), neither \(0.81\); stressed exactly one \(0.08\), neither \(0.86\).)

What You'll Learn