Correlated Multi-Risk Scenario Design

Why Multi-Risk Scenarios Need Dependence

A long-term insurance stress can affect several assumptions at once. A health event may increase mortality and disability incidence; an economic downturn may coincide with weaker investment returns, higher expenses, and a change in policyholder lapse behavior. Treating every assumption as independent can miss a plausible combination—or produce a set of “adverse” assumptions that is not coherent with the relationships being modeled.

A multi-risk scenario is a specified joint movement in assumptions. It is not necessarily a forecast, a probability quantile, or the simultaneous worst possible movement in every variable. Its purpose is to test a defined combination of conditions. The scenario should make clear which variables are linked, the direction of each movement, and the assumptions used to represent that dependence. The methods below are aligned with the SOA Exam ALTAM (Advanced Long-Term Actuarial Mathematics) syllabus.

Definition: A joint risk scenario specifies movements in two or more risk variables together. A dependence assumption describes how those variables vary together, often through their covariance or correlation. A scenario is coherent when its movements follow the stated dependence structure and are interpreted consistently with the product and the chosen risk definitions.

Choose Risk Coordinates Before Choosing Shocks

Correlation is meaningful only after the variables and their directions have been defined. For example, “investment return” rising may be favorable for one product and unfavorable for another. A useful convention for a particular stress exercise is to define each coordinate so that a positive value represents movement in a specified adverse direction. For a guarantee-sensitive product, that might mean a positive investment-risk score represents a fall in investment return. For a product exposed to persistent guarantee costs, a positive lapse-risk score might represent a fall in lapses.

These directions are scenario-specific, not universal. Higher lapses can increase or decrease the value of a product’s obligations depending on its benefits and guarantees. The scenario designer must state the interpretation rather than assuming that “higher” always means “worse.”

For random risk amounts $X_1,\ldots,X_n$, define the standard deviation of $X_j$ as $\sigma_j$, and the correlation between $X_j$ and $X_k$ as $\rho_{jk}$. Correlation is dimensionless and lies between $-1$ and $1$. It describes linear association; it does not, by itself, specify the full joint distribution or prove a causal relationship.

Theorem: For a sum $T=\sum_{j=1}^n X_j$ with finite variances, $$ \operatorname{Var}(T)=\sum_{j=1}^n \sigma_j^2+2\sum_{j<k}\rho_{jk}\sigma_j\sigma_k. $$ This identity does not require normality. If the components are independent, their pairwise covariances are zero, but dependence generally requires the covariance terms.

The identity follows by expanding the centered square. Since $T-E[T]=\sum_j(X_j-E[X_j])$, squaring creates each component’s squared deviation and, for every distinct pair, two identical cross-products. Taking expectations gives each variance once and each covariance twice. Substituting $\operatorname{Cov}(X_j,X_k)=\rho_{jk}\sigma_j\sigma_k$ yields the stated result.

Worked Example: Aggregating Three Correlated Risk Components

A hypothetical portfolio has three centered annual loss components, measured in millions of dollars. Their standard deviations are $1$, $2$, and $2$, and every pair has correlation $0.25$. Calculate the standard deviation of the aggregate loss. Assume only that these variances and correlations are the intended model inputs; no distributional assumption is needed.

Method. Apply the covariance formula, including all three pairwise terms. This avoids the common mistake of adding variances as if the components were independent.

The individual variance terms sum to $$ 1^2+2^2+2^2=1+4+4=9. $$

The three pairwise covariance terms, before the factor of $2$, are $$ 0.25(1)(2)=0.5,\qquad 0.25(1)(2)=0.5,\qquad 0.25(2)(2)=1. $$ Their sum is $0.5+0.5+1=2$. Therefore, $$ \operatorname{Var}(T)=9+2(2)=13 $$ in squared millions of dollars, and $$ \operatorname{SD}(T)=\sqrt{13}=3.605551\ldots $$ million dollars. To four decimal places, the aggregate standard deviation is $\$3.6056$ million. Checking the result, the positive correlations increase the variance above the independent-components value of $9$, as expected.

Representing Dependence with Standardized Drivers

A direct way to define correlated shocks is to start with independent standardized drivers and form weighted combinations. If $U$ and $E$ are independent with mean zero and variance one, then $$ Z_1=U,\qquad Z_2=\rho U+\sqrt{1-\rho^2}\,E $$ gives each $Z_j$ mean zero and variance one, with correlation $\rho$. This construction works for $-1\leq\rho\leq1$. If the drivers are independent standard normal variables, the resulting pair is jointly normal as well. Without that distributional assumption, the construction still establishes the stated means, variances, and correlation, but does not establish joint normality.

A chosen scenario sets values for the underlying drivers and then calculates all risk scores from them. The chosen driver values need not be expected values or percentiles. In particular, a “two-standard-deviation” coordinate in one variable does not imply that every other coordinate must also equal two.

Worked Example: A Joint Mortality and Investment Shock

For a hypothetical product, let $Z_M$ be a standardized mortality-increase score and $Z_I$ a standardized investment-adversity score, where a higher $Z_I$ means a larger fall in investment return. Assume both have mean zero and variance one, with correlation $0.6$. Use independent standardized drivers $U$ and $E$ to represent the pair. In the selected scenario, set $U=2$ and $E=-0.5$. The base annual mortality assumption is $1.00\%$, with an absolute standard deviation of $0.20$ percentage points for its stress coordinate. The base annual investment return is $4.00\%$, with a standard deviation of $5.00$ percentage points for the investment-adversity coordinate. Calculate the stressed assumptions.

Method. Construct the correlated scores first, then convert each score to its assumption movement. This avoids applying a correlation directly to rates with different units.

Set $Z_M=U=2$. For the investment score, $$ Z_I=0.6U+\sqrt{1-0.6^2}\,E =0.6(2)+0.8(-0.5) =1.2-0.4 =0.8. $$ The mortality rate increases by $2(0.20)=0.40$ percentage points, so it becomes $1.00\%+0.40\%=1.40\%$. The investment return falls by $0.8(5.00)=4.00$ percentage points, so it becomes $4.00\%-4.00\%=0.00\%$.

The result is internally consistent with the chosen adverse-direction scores: mortality is above its base assumption and the return is below its base assumption. The correlation is a property of the modeled variables, not a rule that requires their realized scenario scores to be equal.

Check the Correlation Structure

For three or more variables, plausible-looking pairwise correlations may still be mutually impossible. A valid correlation matrix must be symmetric, have ones on its diagonal, and be positive semidefinite. Positive semidefiniteness means that $a^\mathsf{T}Ra\geq0$ for every real vector $a$, where $R$ is the correlation matrix. Equivalently, it cannot imply a negative variance for some linear combination of the variables.

For example, a proposed three-variable matrix with every off-diagonal correlation equal to $-0.9$ fails this condition. Its determinant is $1+2(-0.9)^3-3(-0.9)^2=-2.888$, which is negative. Thus the three pairwise statements cannot all describe one valid correlation structure. A factor construction from independent drivers provides a convenient way to guarantee a valid structure, because the resulting covariance matrix is built from actual linear combinations of those drivers.

Worked Example: Building a Five-Risk Scenario from Shared Drivers

Consider a hypothetical guarantee-sensitive long-term product. Define positive scores as adverse movements: higher mortality, higher disability incidence, higher expense inflation, a larger fall in investment return, and lower lapses. Let $U,V,W,Q$ be independent, mean-zero, variance-one drivers. Construct the scores as $$ Z_M=U,\quad Z_D=0.5U+\sqrt{0.75}V,\quad Z_E=0.4U+\sqrt{0.84}W,\quad Z_I=0.6U+0.8W,\quad Z_L=0.3U+\sqrt{0.91}Q. $$ In the selected scenario, set $U=V=W=Q=1$. The base assumptions are mortality $1.00\%$, disability incidence $3.00\%$, expense inflation $3.00\%$, investment return $4.00\%$, and lapse rate $8.00\%$. Their respective stress-coordinate standard deviations are $0.15$, $0.60$, $1.00$, $5.00$, and $2.00$ percentage points. Convert the scores into stressed assumptions.

Method. Calculate the scores from the shared drivers, then apply each score in its stated adverse direction. This avoids treating each assumption as a separately selected extreme.

The scores are $$ Z_M=1,\quad Z_D=0.5+\sqrt{0.75}=1.366025\ldots,\quad Z_E=0.4+\sqrt{0.84}=1.316515\ldots, $$ $$ Z_I=0.6+0.8=1.4,\qquad Z_L=0.3+\sqrt{0.91}=1.253939\ldots. $$ Thus the mortality assumption is $1.00\%+1(0.15\%)=1.15\%$. Disability incidence is $3.00\%+1.366025(0.60\%)=3.819615\ldots\%$, or $3.8196\%$ to four decimal places. Expense inflation is $3.00\%+1.316515(1.00\%)=4.316515\ldots\%$, or $4.3165\%$. Investment return is reduced by $1.4(5.00\%)=7.00$ percentage points, giving $-3.00\%$. The lapse rate falls by $1.253939(2.00\%)=2.507878\ldots$ percentage points, giving $5.492122\ldots\%$, or $5.4921\%$.

Each score has variance one: for example, $\operatorname{Var}(Z_D)=0.5^2+0.75=1$, and $\operatorname{Var}(Z_I)=0.6^2+0.8^2=1$. As a dependence check, $\operatorname{Corr}(Z_M,Z_D)=0.5$, while $\operatorname{Corr}(Z_D,Z_I)=0.5(0.6)=0.3$; the latter pair shares only the $U$ driver. The investment and expense scores share both $U$ and $W$, so their correlation is $0.4(0.6)+0.8\sqrt{0.84}=0.9732$ to four decimals. The construction therefore states not just five adverse values, but the driver relationships that make them a defined joint scenario.

Practical Checks and Common Traps

A scenario is easier to interpret and reproduce when it records both its structure and its assumptions. At a minimum, state the risk coordinates, their adverse directions, the dependence model, the selected driver values, and the resulting assumption values. If a risk is measured in percentage points, keep that unit explicit when applying a standardized score.

Assuming independence by default. Independence sets covariance terms to zero; it is a modeling assumption, not a consequence of having separate assumption categories.
Adding standard deviations. Variances combine through covariance terms. In general, neither the variances nor the standard deviations can simply be added.
Confusing correlation with causation. Correlation describes modeled co-movement. It does not explain why the variables move together.
Using “adverse” without defining direction. A higher lapse rate is not necessarily adverse. State whether the score represents higher or lower lapses for the product being tested.
Choosing pairwise correlations without a matrix check. A collection of pairwise values must form a positive semidefinite correlation matrix to be jointly feasible.

A joint scenario is a disciplined way to combine risks, not a shortcut to a single “worst case.” The dependence model controls which combinations are represented, while product mechanics determine how the resulting assumptions are interpreted. Keeping those steps separate makes the scenario auditable and helps prevent accidental double counting or contradictory directions.

Check Your Understanding

For each problem, state the calculation or dependence check used, then compare your result with the answer.

Two centered annual loss components have standard deviations of $0.5$ and $1.5$ million dollars and correlation $-0.2$. Calculate the aggregate standard deviation. (Answer: $\sqrt{2.2}=1.4832$ million dollars, to four decimal places.)
Let $Z_1=U$ and $Z_2=0.4U+\sqrt{0.84}E$, where $U,E$ are independent standardized drivers. For $U=1.5$ and $E=-1$, calculate $Z_2$. (Answer: $0.6-\sqrt{0.84}=-0.3165$, to four decimals.)
A proposed correlation matrix for three risks has all three pairwise correlations equal to $-0.6$. Is it a valid correlation matrix? (Answer: No; its determinant is $1+2(-0.6)^3-3(-0.6)^2=-0.512<0$, so it is not positive semidefinite.)
A mortality assumption is $2.00\%$, with an absolute stress-coordinate standard deviation of $0.30$ percentage points. Its adverse score is $1.2$. A return assumption is $3.00\%$, with an adverse-coordinate standard deviation of $4.00$ percentage points, and its adverse score is $0.5$, where a positive score means a fall in return. Find both stressed assumptions. (Answer: Mortality $2.36\%$; investment return $1.00\%$.)
Three centered loss components have standard deviations $1,2,2$ million dollars, with pairwise correlations $0.25,0.25,0.25$. Find the aggregate variance and standard deviation, and explain why independence would give a different result. (Answer: Variance $13$ squared millions of dollars; standard deviation $\sqrt{13}=3.6056$ million dollars. Under independence the variance would be $9$, because all three positive covariance terms would be omitted.)

What You'll Learn