Tutorials › CAS Exam 9 - Risk Management for Actuaries › Interpreting Correlated Market Shocks

Home  /  Tutorials  /  CAS Exam 9 / Market Risk Management and Stress

CAS Exam 9 · Market Risk Management and Stress · Tutorial 786 of 995

Interpreting Correlated Market Shocks

Assess how market factors moving together change portfolio and hedge outcomes under historical and deliberately adverse dependence assumptions.

Advanced 12 min read

What You'll Learn

  • Translate market-factor changes into portfolio gains and losses using signed sensitivities.
  • Calculate portfolio variance when market-loss components are correlated.
  • Compare historical dependence assumptions with deliberately adverse stress assumptions.
  • Evaluate how a hedge can help in one joint shock and hurt in another.
  • Identify key assumptions and limits when interpreting correlation-based risk estimates.

Why Joint Market Movements Matter

A portfolio can be exposed at the same time to interest rates, credit spreads, equity prices, and foreign exchange rates. Assessing each exposure separately can miss an important feature of a market event: several factors may move together. If those movements reinforce losses, the portfolio can lose more than an assessment based on diversification from ordinary historical relationships would suggest. If they offset, the joint result may be less severe.

This tutorial focuses on interpreting those joint movements for portfolio and hedge outcomes. The calculations use linear sensitivities and, where stated, a multivariate normal model for losses. They are not suitable without adjustment for positions whose value changes nonlinearly with market factors. The treatment is aligned with the CAS Exam 9 (Risk Management for Actuaries) syllabus.

Definition: A market shock is a change in one or more market factors over a specified horizon, such as a yield change, a spread change, an equity return, or an exchange-rate return. Dependence describes how the factors, or the portfolio losses associated with them, vary together. Correlation is a standardized measure of linear association between two random variables; it ranges from \(-1\) to \(1\), but does not by itself describe all forms of dependence.

From Market Changes to Portfolio Results

For small changes, a portfolio’s change in value can often be approximated by adding the effects of its market-factor exposures. For example, a bond portfolio may gain when yields fall and lose when credit spreads widen; an equity holding loses when its market price falls; and a foreign-currency asset loses in domestic-currency terms when that currency weakens. Signs matter: a long foreign-currency position and a short foreign-currency position respond in opposite directions to the same exchange-rate move.

Let \(L_i\) denote the loss contribution from market factor \(i\), with a positive value representing a loss. The total loss in a linear model is \(L=\sum_i L_i\). Its variance depends on both the variance of each contribution and the covariance between every pair:

$$ \operatorname{Var}(L) = \sum_i \operatorname{Var}(L_i) + 2\sum_{i<j}\operatorname{Cov}(L_i,L_j) $$

If the loss contributions have standard deviations \(\sigma_i\) and correlations \(\rho_{ij}\), then \(\operatorname{Cov}(L_i,L_j)=\rho_{ij}\sigma_i\sigma_j\). Positive correlation between two loss contributions increases aggregate variance; negative correlation reduces it. This statement concerns loss contributions. A correlation between market-factor changes cannot be interpreted without first considering the portfolio’s signed exposures.

Why the Covariance Formula Holds

The variance is the expected squared deviation from the mean. Write each loss contribution as \(L_i-E[L_i]\), so that \(L-E[L]=\sum_i(L_i-E[L_i])\). Squaring the sum produces each squared term and two copies of each cross-product:

$$ \begin{aligned} \operatorname{Var}(L) &=E\left[\left(\sum_i (L_i-E[L_i])\right)^2\right]\\ &=\sum_i E[(L_i-E[L_i])^2] +2\sum_{i<j}E[(L_i-E[L_i])(L_j-E[L_j])]\\ &=\sum_i\operatorname{Var}(L_i) +2\sum_{i<j}\operatorname{Cov}(L_i,L_j). \end{aligned} $$

The factor of 2 appears because the product for a pair \(i,j\) occurs twice when the square is expanded: once as \(L_iL_j\) and once as \(L_jL_i\). Omitting that factor understates or overstates the contribution of dependence.

Model condition: The variance identity is general when the variances and covariances exist. To convert a modeled standard deviation into a percentile loss using a normal quantile, an additional distributional assumption is required. In the examples below that assumption is that the total loss is normally distributed with mean zero. Normal-based tail estimates may be inappropriate for skewed, heavy-tailed, or nonlinear portfolio losses.

Worked Example: Comparing Historical and Adverse Dependence

Worked Example: Comparing Historical and Adverse Dependence

A hypothetical insurer models three linear loss components: rates, spreads, and equities. Their means are zero and their standard deviations are \(0.50\), \(0.60\), and \(0.40\) million dollars, respectively. Under a historical-dependence assumption, the pairwise correlations are \(0.30\), \(-0.24\), and \(0.083333\) for rates-spreads, rates-equities, and spreads-equities. A deliberately adverse stress assumption instead uses correlations \(0.70\), \(0.48\), and \(0.50\). Assume the component losses are jointly normal under each assumption. Calculate the variance and 99th-percentile loss in each case.

Method. Apply the covariance formula to the three loss components. The three individual variances sum to \(0.50^2+0.60^2+0.40^2=0.25+0.36+0.16=0.77\) million dollars squared.

Under historical dependence, the three covariance contributions, including the factor of 2, are:

$$ \begin{aligned} 2(0.30)(0.50)(0.60)&=0.180,\\ 2(-0.24)(0.50)(0.40)&=-0.096,\\ 2(0.083333)(0.60)(0.40)&\approx 0.040. \end{aligned} $$

Thus the total variance is \(0.77+0.180-0.096+0.040=0.894\) million dollars squared. The standard deviation is \(\sqrt{0.894}=0.9455\) million dollars. With a mean-zero normal loss, the 99th-percentile loss is \(2.3263\) standard deviations:

$$ \operatorname{VaR}_{0.99}=2.3263\sqrt{0.894}\approx 2.1996 \quad\text{million dollars}. $$

That is approximately $2,200,000 to the nearest thousand dollars.

Under the adverse stress correlations, the covariance contributions are \(2(0.70)(0.50)(0.60)=0.420\), \(2(0.48)(0.50)(0.40)=0.192\), and \(2(0.50)(0.60)(0.40)=0.240\). The variance is \(0.77+0.420+0.192+0.240=1.622\) million dollars squared, and the standard deviation is \(\sqrt{1.622}=1.2736\) million dollars. Therefore:

$$ \operatorname{VaR}_{0.99}=2.3263\sqrt{1.622}\approx 2.9627 \quad\text{million dollars}. $$

That is approximately $2,963,000 to the nearest thousand dollars. The adverse-dependence estimate is higher because the assumed pairwise relationships make the loss components reinforce one another more strongly. The common mistake is to use the correct standard deviation but miscalculate its product with the normal quantile; retaining extra digits until the final rounding gives the values shown.

Historical Relationships and Deliberately Adverse Assumptions

Historical dependence estimates summarize observed co-movement over a chosen period. They are useful evidence, but they are not a guarantee that the same relationships will persist in a severe event. Estimates depend on the data window, factor definitions, and observation frequency. Relationships can also change when market participants respond to a shock. A historical correlation is therefore a model input, not a fixed law of markets.

A stress assumption asks a different question: what would the portfolio outcome be if selected market movements occurred together under a specified adverse pattern? The assumed dependence should be deliberate, documented, and consistent with a plausible scenario. Setting every loss correlation to \(1\) may provide a simple upper bound under some fixed marginal distributions, but it can be implausible, may not represent the relevant market-factor directions, and is not a substitute for scenario design.

A correlation matrix must also be mathematically valid. In particular, it must be positive semidefinite; otherwise it cannot be the correlation matrix of a jointly distributed set of factors. For three factors, pairwise correlations that each lie between \(-1\) and \(1\) are not automatically a valid joint specification. The example’s historical and adverse sets are valid: their correlation-matrix determinants are approximately \(0.8335\) and \(0.3656\), respectively, and their diagonal entries are one. In practice, an invalid matrix must be repaired or replaced before it is used in a joint model.

Worked Example: Joint Shocks and a Hedge

Worked Example: Joint Shocks and a Hedge

Consider a hypothetical insurer portfolio with a rate sensitivity of \(-$80,000\) per basis-point increase in yield, a spread sensitivity of \(-$40,000\) per basis-point widening, $15 million of equity exposure, and an $8 million long foreign-currency position. Use the linear approximations \(\Delta V_{\text{rate}}=-80{,}000\Delta y\), \(\Delta V_{\text{spread}}=-40{,}000\Delta s\), \(\Delta V_{\text{equity}}=15{,}000{,}000R_e\), and \(\Delta V_{\text{FX}}=8{,}000{,}000R_f\), where yield and spread changes are in basis points and returns are decimals. Evaluate two joint shocks: a historical-style move of yields down 15 basis points, spreads wider by 20 basis points, equities down 6%, and the foreign currency down 4%; and an adverse move of yields up 25 basis points, spreads wider by 30 basis points, equities down 12%, and the foreign currency down 7%. Then evaluate a hedge that reduces the rate sensitivity to \(-$30,000\) per basis point and the currency exposure to $3 million. Assume the hedge has no transaction cost, basis risk, or nonlinear effects.

Historical-style shock, unhedged. The rate contribution is \(-80{,}000(-15)=+$1,200,000\). The spread contribution is \(-40{,}000(20)=-$800,000\). The equity contribution is \(15{,}000{,}000(-0.06)=-$900,000\), and the currency contribution is \(8{,}000{,}000(-0.04)=-$320,000\). Total portfolio change is \(1{,}200{,}000-800{,}000-900{,}000-320{,}000=-$820,000\), a loss of $820,000.

Adverse shock, unhedged. The four contributions are \(-80{,}000(25)=-$2,000,000\), \(-40{,}000(30)=-$1,200,000\), \(15{,}000{,}000(-0.12)=-$1,800,000\), and \(8{,}000{,}000(-0.07)=-$560,000\). The total is \(-$5,560,000\), a loss of $5,560,000.

Adverse shock, hedged. The rate hedge changes the rate contribution to \(-30{,}000(25)=-$750,000\), and the currency hedge changes the currency contribution to \(3{,}000{,}000(-0.07)=-$210,000\). The spread and equity contributions are unchanged. The total is \(-750{,}000-1{,}200{,}000-1{,}800{,}000-210{,}000=-$3,960,000\). The hedge reduces the adverse-scenario loss by \(5{,}560{,}000-3{,}960{,}000=$1,600,000\).

The same hedge does not necessarily improve every outcome. In the historical-style shock, the hedged rate gain is \(30{,}000(15)=$450,000\), and the hedged currency loss is \(3{,}000{,}000(0.04)=$120,000\). The total change is \(450{,}000-800{,}000-900{,}000-120{,}000=-$1,370,000\), a loss $550,000 greater than the unhedged loss. The sign of a market move and the direction of each exposure determine whether a hedge helps in a particular scenario. The common mistake is to label a hedge as beneficial without stating the shock against which its benefit is being measured.

Worked Example: Dependence and a Rate Hedge’s VaR

Worked Example: Dependence and a Rate Hedge’s VaR

A hypothetical insurer has two mean-zero normal loss components, one associated with rates and one with spreads. Their standard deviations are \(0.60\) and \(0.80\) million dollars. Historical dependence is modeled with correlation \(-0.25\); an adverse stress assumption uses correlation \(0.50\). A rate hedge reduces the rate-loss standard deviation to \(0.30\) million dollars without changing the spread standard deviation or the assumed correlation. Calculate the 99th-percentile loss before and after hedging under both dependence assumptions.

Unhedged, historical dependence. The variance is \(0.60^2+0.80^2+2(-0.25)(0.60)(0.80)=0.36+0.64-0.24=0.76\) million dollars squared. The 99th-percentile loss is \(2.3263\sqrt{0.76}\approx2.0280\) million dollars.

Unhedged, adverse dependence. The variance is \(0.36+0.64+2(0.50)(0.60)(0.80)=1.48\) million dollars squared. The 99th-percentile loss is \(2.3263\sqrt{1.48}\approx2.8301\) million dollars.

Hedged, historical dependence. Replacing the rate standard deviation by \(0.30\) gives variance \(0.30^2+0.80^2+2(-0.25)(0.30)(0.80)=0.09+0.64-0.12=0.61\) million dollars squared. The 99th-percentile loss is \(2.3263\sqrt{0.61}\approx1.8169\) million dollars.

Hedged, adverse dependence. The variance is \(0.09+0.64+2(0.50)(0.30)(0.80)=0.97\) million dollars squared. The 99th-percentile loss is \(2.3263\sqrt{0.97}\approx2.2911\) million dollars.

The hedge lowers modeled 99th-percentile loss under both dependence assumptions, but the adverse dependence estimate remains higher than the historical one. Each result follows from the stated normal, mean-zero model and linear hedge effect; it should not be presented as a distribution-free tail estimate.

Interpreting Results Carefully

Correlation-based aggregation is useful for comparing portfolio configurations and identifying where joint market movements may amplify risk. It is only as informative as the exposure mapping, marginal loss distributions, dependence assumptions, and horizon used. A historical covariance estimate can be unstable, particularly when the data contain few severe observations. A stress dependence assumption is a scenario choice and should be labeled as such rather than reported as a measured historical fact.

  • Check the exposure signs. Correlation of market changes does not automatically equal correlation of portfolio losses. A short position reverses the effect of a factor move.
  • Distinguish scenarios from distributions. A set of simultaneous shocks gives a scenario result; it does not by itself define a VaR or other percentile.
  • Include the factor of 2. Each pairwise covariance appears twice in the expansion of total variance.
  • Do not assume linearity is exact. Large changes may invalidate duration-like sensitivities or other local approximations. Nonlinear positions require repricing or a model that captures their response.
  • Do not treat correlation as causation or tail dependence. Correlation measures linear co-movement; it does not establish why factors moved together or guarantee that their extreme outcomes will follow the same pattern.
Key takeaway: A sound joint-market assessment states the portfolio’s signed exposures, the market shocks or dependence assumptions, and the model used to translate them into losses. Historical relationships and adverse stress assumptions answer different questions; compare them without confusing one for the other.

Check Your Understanding

For each problem, retain full precision through intermediate calculations and round dollar amounts to the nearest cent unless the answer is stated in millions.

  1. Two mean-zero loss components have standard deviations of $0.40 million and $0.30 million and correlation \(0.50\). Under a joint normal model, calculate the aggregate variance and 95th-percentile loss using the standard normal quantile \(1.6449\). (Answer: variance \(0.37\) million dollars squared; 95th-percentile loss approximately $1.0006 million.)
  2. Two loss components have standard deviations of \(0.50\) and \(0.70\) million dollars. Compare their aggregate variances at correlations \(-0.20\) and \(0.60\). (Answer: \(0.60\) and \(1.16\) million dollars squared, respectively.)
  3. A portfolio has rate sensitivity \(-$45,000\) per basis point, spread sensitivity \(-$20,000\) per basis point, $4 million of equity exposure, and a $2 million long foreign-currency position. Yields rise 12 basis points, spreads widen 18 basis points, equities fall 5%, and the currency rises 3%. Calculate the total change in value. (Answer: a loss of $1,040,000.)
  4. For the portfolio in question 3, a hedge reduces the rate sensitivity to \(-$15,000\) per basis point and the currency exposure to $0.5 million. Recalculate the joint-shock result, leaving other exposures unchanged. (Answer: a loss of $725,000.)
  5. Two mean-zero normal loss components have standard deviations \(0.60\) and \(0.80\) million dollars. Compare their 99th-percentile aggregate losses using correlations \(0.25\) and \(0.75\), with normal quantile \(2.3263\). (Answer: approximately $2.5905 million at correlation \(0.25\), and $3.0509 million at correlation \(0.75\).)

Previous

Next

Back to Tutorials
Clinical Biostats