Why Stress Dependence Directly?
Historical correlation summarizes linear co-movement over the observations and period used to estimate it. It does not, by itself, determine how often two risks will be severe together. A risk manager may therefore need to test a plausible adverse joint outcome even when historical Pearson correlation is modest, unstable, or estimated from a period with few extreme losses.
A dependence stress is a specified change to the joint behavior of risks used to assess the consequences of adverse co-movement. It may increase the probability that both risks exceed chosen thresholds, strengthen their association in a particular regime, or pair adverse marginal outcomes more often. The stress is an assumption for analysis, not a claim that historical data prove the stressed dependence will occur.
The focus here is not to replace historical estimates or to select a particular dependence model. It is to formulate transparent adverse assumptions, check their arithmetic, and measure their consequences. Historical correlation, rank measures, and finite-threshold co-exceedance are useful evidence, but none removes the need to state the stress itself.
Joint Exceedance Stresses and Feasibility
Suppose \(A\) and \(B\) are two adverse events with marginal probabilities \(p_X=P(A)\) and \(p_Y=P(B)\). Write \(q=P(A\cap B)\). The marginal probabilities constrain the possible value of \(q\):
The upper bound follows because the intersection cannot be more likely than either event. The lower bound follows from \(P(A\cup B)=p_X+p_Y-q\leq 1\). A proposed stressed joint probability outside these bounds cannot describe any joint distribution with the stated marginals.
When both events have positive probabilities and are not certain, their indicator correlation is
This follows because \(E[\mathbf{1}_A]=p_X\), \(E[\mathbf{1}_B]=p_Y\), and \(E[\mathbf{1}_A\mathbf{1}_B]=q\). Thus the covariance of the indicators is \(q-p_Xp_Y\), and their variances are \(p_X(1-p_X)\) and \(p_Y(1-p_Y)\). This formula describes the correlation of the specified event indicators; it is not necessarily the Pearson correlation of the full loss amounts.
Worked Example: Increasing Joint Extremes While Preserving Marginals
Method: Use a finite joint distribution and calculate its moments directly. A hypothetical insurer studies two loss measures, \(X\) and \(Y\), each of which has marginal values \(0,2,5\) with probability \(1/3\) each. Under a baseline assumption, \(X\) and \(Y\) are independent. Under a stress, first select one of two matchings with equal probability, then select one of its three pairs uniformly:
- Matching 1: \((0,0),(2,2),(5,5)\).
- Matching 2: \((0,2),(2,5),(5,0)\).
Find the baseline and stressed correlations and the probability that both losses equal 5.
Step 1: Check the marginals. Each matching uses each value \(0,2,5\) once in the \(X\) coordinate and once in the \(Y\) coordinate. Both matchings therefore preserve the stated marginal probabilities.
For either marginal, \(E[X]=E[Y]=(0+2+5)/3=7/3\), and \(E[X^2]=E[Y^2]=(0+4+25)/3=29/3\). Hence \(\operatorname{Var}(X)=\operatorname{Var}(Y)=29/3-(7/3)^2=38/9\).
Step 2: Calculate the baseline. Independence gives \(E[XY]=E[X]E[Y]=49/9\), so the covariance and correlation are both zero. Also, \(P(X=5,Y=5)=P(X=5)P(Y=5)=1/9\).
Step 3: Calculate the stressed product expectation. The product sums in Matching 1 and Matching 2 are respectively \(0+4+25=29\) and \(0+10+0=10\). The conditional average product in each matching is therefore \(29/3\) and \(10/3\). Averaging across the two equally likely matchings gives \(E[XY]=\tfrac12(29/3)+\tfrac12(10/3)=13/2\).
The stressed covariance is \(13/2-49/9=19/18\). Thus the stressed correlation is \((19/18)/(38/9)=1/4=0.25\).
Step 4: Calculate the stressed joint extreme probability. Both losses equal 5 only when Matching 1 is selected and its \((5,5)\) pair is selected. Therefore \(P(X=5,Y=5)=(1/2)(1/3)=1/6\).
Check and trap avoided. The marginals remain unchanged, and the probability of both losses equaling 5 rises from \(1/9\) to \(1/6\). Do not infer the stressed correlation by assuming the average product equals the product of the means. Under the stress, \(E[XY]=13/2\), not \(49/9\); the resulting correlation is \(1/4\), not zero.
From Dependence Stress to Aggregate-Loss Impact
A dependence stress matters because joint outcomes can change the distribution of aggregate loss even when each risk's marginal distribution is unchanged. The next example uses only two loss-event indicators, so every probability and aggregate outcome can be checked explicitly.
Worked Example: A Co-Exceedance Stress Changes Aggregate VaR
Method: Construct the four outcomes from the marginal event probabilities and the stressed intersection probability, then apply the quantile definition. A hypothetical portfolio has loss \(X=4\) million dollars if event \(A\) occurs and zero otherwise; loss \(Y=7\) million dollars if event \(B\) occurs and zero otherwise. Each event has probability \(0.10\). Under independence, \(P(A\cap B)=0.01\). A stress raises this probability to \(0.06\) while preserving both marginal probabilities. Calculate the event-indicator correlations and the 95th-percentile VaR of \(S=X+Y\) under each assumption.
Step 1: Check feasibility and construct the stressed table. For equal marginals of \(0.10\), the joint probability must lie between 0 and \(0.10\), so \(0.06\) is feasible. The stressed probabilities are: \(P(A\cap B)=0.06\), \(P(A\cap B^c)=0.10-0.06=0.04\), \(P(A^c\cap B)=0.10-0.06=0.04\), and \(P(A^c\cap B^c)=1-0.06-0.04-0.04=0.86\).
Step 2: Compare correlations. Each indicator has variance \(0.10(0.90)=0.09\). Under independence, the indicator correlation is zero. Under stress it is \((0.06-0.10(0.10))/0.09=0.05/0.09=5/9\), or \(0.5556\) to four decimal places.
Step 3: Find the aggregate-loss distributions. Under independence, the probabilities of aggregate losses \(0,4,7,11\) million dollars are \(0.81,0.09,0.09,0.01\), respectively. Under stress, the corresponding probabilities are \(0.86,0.04,0.04,0.06\).
Step 4: Apply the quantile definition. Define \(\operatorname{VaR}_{0.95}(S)=\inf\{s:P(S\leq s)\geq0.95\}\). Under independence, the cumulative probabilities at \(0,4,7,11\) are \(0.81,0.90,0.99,1.00\), so \(\operatorname{VaR}_{0.95}(S)=7\) million dollars. Under stress they are \(0.86,0.90,0.94,1.00\), so \(\operatorname{VaR}_{0.95}(S)=11\) million dollars.
Check and trap avoided. Both assumptions give \(E[S]=4(0.10)+7(0.10)=1.10\) million dollars, but the 95th-percentile VaR rises by 4 million dollars under stress. Do not conclude that unchanged marginal losses or unchanged expected aggregate loss imply unchanged tail capital.
Designing a Stress That Is Useful
A good dependence stress is specific enough to reproduce and interpret. It identifies the risks, the adverse events or thresholds, the feature being stressed, and the marginal assumptions held fixed. If several thresholds or regimes are relevant, the stress should say which ones change. A single Pearson correlation adjustment does not communicate all of this information.
State the loss variables and thresholds, for example \(A=\{X>x_0\}\) and \(B=\{Y>y_0\}\). Explain why those thresholds represent adverse outcomes for the risk assessment.
Summarize the historical estimates used, their observation period, and material limitations. Treat sampling uncertainty and regime instability as reasons to qualify the estimate, not as proof of a particular stressed value.
For example, raise \(P(A\cap B)\) while retaining \(P(A)\) and \(P(B)\), or describe a separate adverse regime. State the rationale and verify probability bounds.
Evaluate aggregate loss, quantiles, or another measure relevant to the risk decision under both baseline and stress assumptions. Distinguish model output from the assumption that generated it.
Explain what the stress reveals, how sensitive the result is to the chosen thresholds, and who reviews the assumptions. Do not describe a scenario result as a forecast unless a forecasting basis has been established.
Worked Example: An Asymmetric Stress for Unequal Event Probabilities
Method: Build the joint event table from the marginals and intersection probability, then compare conditional probabilities and an aggregate quantile. A hypothetical portfolio has an event \(A\) with probability \(0.10\), producing a loss of 10 million dollars, and an event \(B\) with probability \(0.20\), producing a loss of 6 million dollars. The baseline assumes independence; the stress sets \(P(A\cap B)=0.07\) and retains both marginal probabilities. Calculate \(P(B\mid A)\), the indicator correlation, and the 99th-percentile VaR of the total loss.
Step 1: Verify feasibility. The possible intersection range is from \(\max(0,0.10+0.20-1)=0\) to \(\min(0.10,0.20)=0.10\). Thus the stressed value \(0.07\) is feasible.
Step 2: Construct the stressed probabilities. Both events occur with probability \(0.07\). Event \(A\) alone has probability \(0.10-0.07=0.03\); event \(B\) alone has probability \(0.20-0.07=0.13\). Neither occurs with probability \(1-0.07-0.03-0.13=0.77\). These probabilities sum to 1.
Step 3: Calculate conditional probability and correlation. The conditional probability is \(P(B\mid A)=0.07/0.10=0.70\), compared with \(0.20\) under independence. The indicator covariance is \(0.07-(0.10)(0.20)=0.05\). The indicator variances are \(0.10(0.90)=0.09\) and \(0.20(0.80)=0.16\), so the correlation is \(0.05/\sqrt{0.09(0.16)}=0.05/0.12=0.4167\), to four decimal places.
Step 4: Calculate the 99th-percentile VaR. The stressed total loss is 16 million dollars with probability \(0.07\), 10 million dollars with probability \(0.03\), 6 million dollars with probability \(0.13\), and zero with probability \(0.77\). The cumulative probability through 10 million dollars is \(0.77+0.13+0.03=0.93\). It reaches 1.00 at 16 million dollars. Therefore \(\operatorname{VaR}_{0.99}=16\) million dollars.
Under independence, the four probabilities for both, \(A\) alone, \(B\) alone, and neither are \(0.02,0.08,0.18,0.72\). The cumulative probability through 10 million dollars is \(0.72+0.18+0.08=0.98\), so the baseline 99th-percentile VaR is 16 million dollars as well. The stress does not change this particular quantile, although it raises \(P(B\mid A)\) from \(0.20\) to \(0.70\). This is a useful check: a dependence stress need not change every reported risk measure.
Check and trap avoided. The stressed table preserves the marginals and sums to 1. Do not claim that every adverse dependence stress must increase every quantile. The effect depends on the loss amounts, threshold, and chosen risk measure.
Interpreting the Stress Alongside Historical Estimates
Historical estimates and stress assumptions answer different questions. A historical estimate summarizes observed dependence under the data and period used. A stress asks how results change under a specified adverse joint behavior. Presenting them together helps decision-makers distinguish evidence from a deliberate risk-management assumption.
| Item | What it describes | Question to ask |
|---|---|---|
| Historical correlation | Observed linear co-movement over a stated sample | Is the sample relevant and sufficiently informative for the risks being assessed? |
| Joint exceedance estimate | Observed or modeled frequency of specified events occurring together | Are the event thresholds and estimation period stated clearly? |
| Dependence stress | A specified adverse joint assumption used for sensitivity or scenario analysis | Are the marginals, stressed feature, and rationale explicit and internally consistent? |
A stress should not be presented as a new historical estimate merely because it is expressed numerically. Nor should a higher event-indicator correlation be mistaken for a complete specification of the full loss distribution. Two joint models can share a correlation while differing in other features that matter to extreme aggregate losses. The practical discipline is to state the stressed feature directly and then calculate its effect on the chosen risk measure.
Common Exam Traps
- Changing marginals unintentionally. If the stated purpose is to stress dependence alone, verify that each risk retains its marginal probability or distribution.
- Using an impossible intersection probability. Check the bounds before calculating conditional probabilities or correlations.
- Equating a correlation stress with a tail stress. Correlation is a summary, not a full joint distribution. Specify the adverse event or dependence feature directly.
- Assuming every capital metric must rise. Recalculate the requested measure. A stress may alter one part of the distribution without changing a particular quantile.
- Reporting an assumption as observed fact. Label historical estimates, judgmental stresses, and model outputs separately.
Check Your Understanding
For each problem, check the probability table and the requested calculation before comparing with the answer.
- Two adverse event indicators have marginal probabilities \(0.08\) and \(0.15\). A stress sets their joint probability to \(0.05\). Find the stressed conditional probability of the second event given the first and the indicator correlation. (Answer: \(P(B\mid A)=0.6250\); correlation \(=0.3923\).)
- Two risks each have an adverse-event probability of \(0.20\). A dependence stress sets the joint probability to \(0.12\), with marginals unchanged. Find the indicator correlation and \(P(B\mid A)\). (Answer: correlation \(=0.5000\); \(P(B\mid A)=0.6000\).)
- Events \(A\) and \(B\) have probabilities \(0.10\) and \(0.30\), and a stress sets \(P(A\cap B)=0.04\). Find the probabilities of \(A\) alone, \(B\) alone, and neither event. (Answer: \(0.06,\ 0.26,\ 0.64\), respectively.)
- Two events each have probability \(0.05\). Independence gives a joint probability of \(0.0025\); a stress raises it to \(0.03\) while preserving marginals. Find \(P(B\mid A)\) and the ratio of stressed to independent joint probability. (Answer: \(P(B\mid A)=0.6000\); ratio \(=12\).)
- A portfolio loses 8 million dollars if event \(A\) occurs and 5 million dollars if event \(B\) occurs. Their probabilities are \(0.10\) and \(0.20\), and a stress sets their joint probability to \(0.08\). Find the stressed indicator correlation and the 95th-percentile VaR of aggregate loss. (Answer: correlation \(=0.5000\); \(\operatorname{VaR}_{0.95}=13\) million dollars.)