Dependence From Shared Risk Drivers

Why Shared Risk Drivers Matter

Two lines of business can respond to the same underlying environment: weather, economic conditions, litigation climate, or a level of system activity. Within any one environment, their losses might be generated independently. Across environments, however, both lines may tend to have higher or lower expected losses together. That shared movement can make the losses dependent overall.

For risk aggregation, this distinction matters. Adding the separate expected losses gives the expected aggregate loss, but assessing the variability of the aggregate also requires accounting for covariance. A model that treats conditionally independent losses as unconditionally independent can therefore understate or overstate aggregate risk.

Definition: Let $Z$ be a shared risk driver and let $X$ and $Y$ be losses with finite second moments. The losses are conditionally independent given $Z$ if, for every relevant value $z$, their conditional joint distribution factors into the product of their conditional marginal distributions: $$ P(X\in A,\,Y\in B\mid Z=z) = P(X\in A\mid Z=z)P(Y\in B\mid Z=z) $$ for all sets $A$ and $B$. This means that once the driver is known, observing one loss gives no additional information about the other. It does not, by itself, mean that $X$ and $Y$ are independent without conditioning on $Z$.

A shared driver creates overall dependence when it makes the conditional loss distributions move together. In particular, conditional independence implies a useful covariance identity. Here, covariance measures linear co-movement: positive covariance means the losses tend to be above or below their respective means together, while negative covariance indicates opposite movement.

The Covariance Result

Write $m_X(Z)=E[X\mid Z]$ and $m_Y(Z)=E[Y\mid Z]$. The law of total covariance separates overall co-movement into a within-driver component and a between-driver component.

\operatorname{Cov}(X,Y) = E[\operatorname{Cov}(X,Y\mid Z)] + \operatorname{Cov}(E[X\mid Z],E[Y\mid Z])

When $X$ and $Y$ are conditionally independent given $Z$, their conditional covariance is zero. Thus:

Theorem: If $X$ and $Y$ have finite second moments and are conditionally independent given $Z$, then $$ \operatorname{Cov}(X,Y) = \operatorname{Cov}(E[X\mid Z],E[Y\mid Z]). $$ Consequently, if the two conditional means tend to be high in the same driver states, the resulting covariance is positive; if one tends to be high when the other is low, it is negative. Conditional independence alone does not determine the sign.

The identity follows directly from conditional expectation. Conditional on $Z$, the product expectation is the conditional covariance plus the product of the conditional means. Taking expectations and subtracting $E[X]E[Y]$ gives the result.

\begin{aligned} \operatorname{Cov}(X,Y) &=E[XY]-E[X]E[Y]\\ &=E\!\left[E[XY\mid Z]\right]-E[m_X(Z)]E[m_Y(Z)]\\ &=E[\operatorname{Cov}(X,Y\mid Z)] +E[m_X(Z)m_Y(Z)]-E[m_X(Z)]E[m_Y(Z)]\\ &=E[\operatorname{Cov}(X,Y\mid Z)] +\operatorname{Cov}(m_X(Z),m_Y(Z)). \end{aligned}

Under conditional independence, the conditional joint distribution factors, so $E[XY\mid Z]=E[X\mid Z]E[Y\mid Z]$. Therefore $\operatorname{Cov}(X,Y\mid Z)=0$, and only the covariance between conditional means remains. This derivation also shows why independence within each environment does not eliminate co-movement caused by changes between environments.

Worked Example: Two Lines Across Calm and Stress Regimes

A hypothetical insurer models annual losses for two lines, in thousands of dollars. The shared driver $Z$ is the environment: calm with probability $0.75$, or stress with probability $0.25$. Conditional on the environment, the two losses are independent. In a calm year, $X$ is equally likely to be 0 or 4 and $Y$ equally likely to be 0 or 2. In a stress year, $X$ is equally likely to be 0 or 20 and $Y$ equally likely to be 0 or 14. Calculate the expected losses and their covariance, and check whether the losses are independent overall.

Method. First calculate the conditional means. Then use the shared-regime covariance result. All loss amounts below are in thousands of dollars.

In a calm year, $E[X\mid Z=\text{calm}]=(0+4)/2=2$ and $E[Y\mid Z=\text{calm}]=(0+2)/2=1$. In a stress year, the corresponding means are $(0+20)/2=10$ and $(0+14)/2=7$.

The unconditional expected losses are $E[X]=0.75(2)+0.25(10)=1.5+2.5=4$ and $E[Y]=0.75(1)+0.25(7)=0.75+1.75=2.5$.

Conditional independence gives $E[XY\mid Z]=E[X\mid Z]E[Y\mid Z]$. Therefore, $E[XY]=0.75(2\cdot1)+0.25(10\cdot7)=1.5+17.5=19$. The covariance is $\operatorname{Cov}(X,Y)=E[XY]-E[X]E[Y]=19-(4)(2.5)=9$ thousand-dollars squared.

Check. The conditional means rise together from $(2,1)$ in calm conditions to $(10,7)$ in stress conditions, so positive covariance is consistent with the scenario. Since the covariance is nonzero, $X$ and $Y$ are not independent overall. The calculation avoids the mistake of setting overall covariance to zero just because conditional covariance is zero.

Interpreting the Shared-Driver Covariance

For a driver with a finite set of regimes $z_j$, probabilities $p_j$, and conditional means $m_{X,j}$ and $m_{Y,j}$, the covariance of conditional means can be calculated as

\operatorname{Cov}(E[X\mid Z],E[Y\mid Z]) = \sum_j p_j(m_{X,j}-E[X])(m_{Y,j}-E[Y]).

This expression focuses on how the conditional means vary from regime to regime. If high values of one conditional mean generally coincide with high values of the other, the products tend to be positive. If high values coincide with low values, they tend to be negative. This is a statement about covariance, not a claim that every possible feature of the joint distribution is captured by the conditional means.

Worked Example: A Shared Driver with Random Claim Counts

A hypothetical company models claim counts in two lines. A shared activity driver $Z$ equals 1 or 3, each with probability $0.5$. Conditional on $Z$, the counts are independent, with $X\mid Z\sim\operatorname{Poisson}(2Z)$ and $Y\mid Z\sim\operatorname{Poisson}(3Z)$. Find the expected counts, covariance, variances, and correlation.

Method. Use the conditional Poisson means and variances, then separate each overall variance into expected conditional variance plus variance of the conditional mean. The shared-driver covariance formula gives the covariance between the counts.

The driver has $E[Z]=0.5(1)+0.5(3)=2$, and $E[Z^2]=0.5(1^2)+0.5(3^2)=5$, so $\operatorname{Var}(Z)=5-2^2=1$. The unconditional means are $E[X]=E[2Z]=2(2)=4$ and $E[Y]=E[3Z]=3(2)=6$.

The conditional means are $2Z$ and $3Z$. Hence $\operatorname{Cov}(X,Y)=\operatorname{Cov}(2Z,3Z)=6\operatorname{Var}(Z)=6$. For a Poisson variable, conditional variance equals conditional mean. Thus $\operatorname{Var}(X)=E[2Z]+\operatorname{Var}(2Z)=4+4(1)=8$, and $\operatorname{Var}(Y)=E[3Z]+\operatorname{Var}(3Z)=6+9(1)=15$.

The correlation is $\operatorname{Corr}(X,Y)=6/\sqrt{8\cdot15}=6/\sqrt{120}\approx0.5477$, rounded to four decimal places. The covariance is positive because both conditional claim-count means increase with the activity driver.

Check. Directly, conditional independence gives $E[XY\mid Z]=(2Z)(3Z)=6Z^2$, so $E[XY]=6(5)=30$. Subtracting $E[X]E[Y]=(4)(6)=24$ again gives covariance $30-24=6$. This check avoids treating the random driver as fixed at its average value.

Worked Example: A Shared Driver Can Produce Negative Covariance

Consider two hypothetical lines, with losses in thousands of dollars, and a shared economic regime that is favorable or adverse with equal probability. Conditional on the favorable regime, $X$ is equally likely to be 0 or 4 and $Y$ equally likely to be 7 or 11. Conditional on the adverse regime, $X$ is equally likely to be 6 or 10 and $Y$ equally likely to be 1 or 5. The losses are conditionally independent in each regime. Calculate their covariance.

Method. Calculate the conditional means and compare their movement across the two regimes. This is the same shared-driver identity, but the conditional means move in opposite directions.

In the favorable regime, the means are $E[X\mid Z]=2$ and $E[Y\mid Z]=9$. In the adverse regime, they are $E[X\mid Z]=8$ and $E[Y\mid Z]=3$. Therefore $E[X]=0.5(2)+0.5(8)=5$, and $E[Y]=0.5(9)+0.5(3)=6$.

The conditional-mean covariance is $0.5(2-5)(9-6)+0.5(8-5)(3-6)=0.5(-9)+0.5(-9)=-9$. Thus $\operatorname{Cov}(X,Y)=-9$ thousand-dollars squared.

Check. Conditional independence gives $E[XY\mid Z]$ equal to the product of conditional means. Hence $E[XY]=0.5(2\cdot9)+0.5(8\cdot3)=9+12=21$, and $E[XY]-E[X]E[Y]=21-(5)(6)=-9$. A shared driver creates dependence here, but not positive dependence: the conditional means move in opposite directions.

Common Exam Trap and Risk-Management Use

A frequent error is to reason, “The losses are independent conditional on the regime, so their covariance is zero.” That conclusion applies to $\operatorname{Cov}(X,Y\mid Z)$, not automatically to $\operatorname{Cov}(X,Y)$. For the first worked example, the incorrect calculation would set $\operatorname{Cov}(X,Y)=0$. The correct calculation is $\operatorname{Cov}(X,Y)=\operatorname{Cov}(E[X\mid Z],E[Y\mid Z])=9$ thousand-dollars squared.

The distinction carries directly into aggregate variance. For an aggregate $X+Y$,

\operatorname{Var}(X+Y) = \operatorname{Var}(X)+\operatorname{Var}(Y)+2\operatorname{Cov}(X,Y).

Ignoring a positive shared-driver covariance understates aggregate variance; ignoring a negative one overstates it. The expected aggregate remains $E[X+Y]=E[X]+E[Y]$, regardless of dependence, but its variance depends on the covariance. In applications, the risk driver should represent a plausible common influence and the conditional model should reflect what is assumed independent after that influence is accounted for.

The shared-driver result is a covariance statement. Zero covariance does not in general prove independence, and conditional independence is a stronger distributional condition than zero conditional covariance. When applying the result, state the driver, the conditional distributions or conditional means, and the conditional independence assumption explicitly. This topic is aligned with the CAS Exam 9 (Risk Management for Actuaries) syllabus.

Check Your Understanding

For each problem, identify the shared-driver role and use conditional means or conditional distributions as appropriate.

A hypothetical insurer has a calm regime with probability $0.7$ and a severe regime with probability $0.3$. Conditional losses $X,Y$, in thousands of dollars, are independent and have conditional means $(2,5)$ in calm conditions and $(8,11)$ in severe conditions. Find both unconditional means and the covariance. (Answer: $E[X]=3.8$, $E[Y]=6.8$, and $\operatorname{Cov}(X,Y)=7.56$ thousand-dollars squared.)
A common activity driver $Z$ is 1 or 3 with equal probability. Given $Z$, independent claim counts satisfy $X\sim\operatorname{Poisson}(Z)$ and $Y\sim\operatorname{Poisson}(2Z)$. Find their covariance and the variance of $X+Y$. (Answer: $\operatorname{Cov}(X,Y)=2$ and $\operatorname{Var}(X+Y)=15$.)
A shared regime is equally likely to be low or high. Conditional on the regime, independent losses have means $E[X\mid\text{low}]=3$, $E[X\mid\text{high}]=9$, $E[Y\mid\text{low}]=10$, and $E[Y\mid\text{high}]=4$. Find the unconditional covariance and state its sign. (Answer: $\operatorname{Cov}(X,Y)=-9$; it is negative.)
A driver is low with probability $0.8$ and high with probability $0.2$. Conditional on it, $X$ and $Y$ are independent, with conditional means for $X$ of 1 and 6 and conditional means for $Y$ of 2 and 5. Their conditional variances are respectively 4 and 9 for $X$, and 1 and 4 for $Y$, in low and high regimes. Calculate $\operatorname{Var}(X+Y)$. (Answer: $16.84$.)
A driver $Z$ is 0 or 1, with $P(Z=1)=0.25$. Given $Z$, independent binary indicators $I_X,I_Y$ have success probabilities $P(I_X=1\mid Z=0)=0.1$, $P(I_X=1\mid Z=1)=0.7$, $P(I_Y=1\mid Z=0)=0.2$, and $P(I_Y=1\mid Z=1)=0.8$. Find $P(I_X=1,I_Y=1)$, the product of the marginal probabilities, and their covariance. (Answer: $0.1550$, $0.0875$, and $0.0675$, respectively.)

What You'll Learn