EPV and VHM as Structural Parameters

Why Separate the Two Sources of Variation?

Suppose two policyholders have different expected annual losses because they have different underlying risk characteristics. Even if each policyholder’s risk characteristics remain fixed, their realized annual losses can vary from year to year. A credibility model separates these two sources of uncertainty: variation within a risk and variation between risks.

The expected process variance, abbreviated EPV, summarizes the average variation of outcomes around their risk-specific means. The variance of hypothetical means, abbreviated VHM, summarizes how much those risk-specific means vary across the population of risks. These are structural parameters of the model: they describe the assumed risk population and its conditional outcome distributions, rather than the variation in one particular observed sample.

The calculations require the distribution of the risk parameter and the conditional mean and variance of the outcome given that parameter. They do not require a fitted credibility premium. The focus here is on defining and calculating the two structural quantities and checking them against the total variance.

Definition: Let $\Theta$ denote a risk parameter and let $X$ be a random outcome, such as a hypothetical annual loss. Define the risk-specific mean $m(\Theta)=E[X\mid\Theta]$ and the conditional variance $\sigma^2(\Theta)=\operatorname{Var}(X\mid\Theta)$. Provided the relevant expectations are finite, $$ \mathrm{EPV}=E[\operatorname{Var}(X\mid\Theta)] =E[\sigma^2(\Theta)], \qquad \mathrm{VHM}=\operatorname{Var}(E[X\mid\Theta]) =\operatorname{Var}(m(\Theta)). $$ The grand mean is $E[X]=E[m(\Theta)]$. The expectation and variance in these definitions are taken over the distribution of $\Theta$.

For a discrete risk parameter taking values $\theta_j$ with probabilities $p_j$, where $\sum_j p_j=1$, the definitions become weighted sums:

E[X]=\sum_j p_jm(\theta_j),\qquad \mathrm{EPV}=\sum_jp_j\sigma^2(\theta_j),\qquad \mathrm{VHM}=\sum_jp_j\bigl(m(\theta_j)-E[X]\bigr)^2.

The probabilities are essential. Unless the risk parameter is uniformly distributed over its possible values, treating each listed value as equally likely changes the model and generally gives incorrect structural quantities. For a continuous $\Theta$, the corresponding expectations are integrals with respect to its probability distribution.

Why EPV and VHM Add to the Unconditional Variance

The identity linking the two structural parameters to the overall variation is the law of total variance. It provides both an interpretation and a useful arithmetic check. Conditional on $\Theta$, $X$ fluctuates around $m(\Theta)$; across risks, the center $m(\Theta)$ itself varies around the grand mean $E[X]$.

Theorem: If $X$ has finite variance, then $$ \operatorname{Var}(X) =E[\operatorname{Var}(X\mid\Theta)] +\operatorname{Var}(E[X\mid\Theta]) =\mathrm{EPV}+\mathrm{VHM}. $$

To justify the identity, write $\mu=E[X]$ and add and subtract $m(\Theta)$ inside the centered outcome:

X-\mu=\bigl(X-m(\Theta)\bigr)+\bigl(m(\Theta)-\mu\bigr).

Square both sides and take the conditional expectation given $\Theta$. The first squared term has conditional expectation $\operatorname{Var}(X\mid\Theta)$, because $E[X-m(\Theta)\mid\Theta]=0$. The second squared term is fixed given $\Theta$. The cross term is zero because it contains the conditional expectation $E[X-m(\Theta)\mid\Theta]$:

E[(X-\mu)^2\mid\Theta] =\operatorname{Var}(X\mid\Theta) +\bigl(m(\Theta)-\mu\bigr)^2 +2\bigl(m(\Theta)-\mu\bigr)E[X-m(\Theta)\mid\Theta] =\operatorname{Var}(X\mid\Theta)+\bigl(m(\Theta)-\mu\bigr)^2.

Taking expectations over $\Theta$ gives $E[(X-\mu)^2]=E[\operatorname{Var}(X\mid\Theta)]+E[(m(\Theta)-\mu)^2]$. The left side is $\operatorname{Var}(X)$; the two terms on the right are EPV and VHM. No assumption that $X$ is normally distributed is needed. The identity requires finite second moments, not a particular distributional shape.

Calculating the Structural Parameters

Worked Example: Unequal Probabilities Across Three Risk Types

A hypothetical insurer models annual loss $X$, measured in thousands of dollars, using a risk parameter $\Theta$. The parameter takes values $0,1,2$ with probabilities $0.2,0.3,0.5$, respectively. Given $\Theta=\theta$, the conditional mean is $E[X\mid\Theta=\theta]=3+3\theta$ thousand dollars, and the conditional variance is $\operatorname{Var}(X\mid\Theta=\theta)=4+3\theta$ in squared thousands of dollars. Find the grand mean, EPV, VHM, and unconditional variance.

Method. Evaluate the conditional mean and variance at every risk type, then take probability-weighted averages. The probability weights sum to $0.2+0.3+0.5=1$.

Step 1: Conditional means and variances. At $\theta=0$, the mean is $3+3(0)=3$ and the variance is $4+3(0)=4$. At $\theta=1$, they are $6$ and $7$. At $\theta=2$, they are $9$ and $10$. Thus the mean/variance pairs are $(3,4),(6,7),(9,10)$.

Step 2: Grand mean. Weight each conditional mean by the probability of its risk type:

$$ E[X]=0.2(3)+0.3(6)+0.5(9)=0.6+1.8+4.5=6.9. $$

The grand mean is $6.9$ thousand dollars.

Step 3: EPV. Average the conditional variances using the same probabilities:

\mathrm{EPV}=0.2(4)+0.3(7)+0.5(10)=0.8+2.1+5.0=7.9.

EPV is $7.9$ squared thousands of dollars.

Step 4: VHM. Measure each conditional mean’s squared deviation from the grand mean, then apply the risk probabilities:

\begin{aligned} \mathrm{VHM} &=0.2(3-6.9)^2+0.3(6-6.9)^2+0.5(9-6.9)^2\\ &=0.2(15.21)+0.3(0.81)+0.5(4.41)\\ &=3.042+0.243+2.205=5.490. \end{aligned}

VHM is $5.49$ squared thousands of dollars.

Step 5: Check. The law of total variance gives $\operatorname{Var}(X)=7.9+5.49=13.39$ squared thousands of dollars. The grand mean, EPV, and VHM are $6.9$, $7.9$, and $5.49$, respectively. The addition check confirms the decomposition.

Common mistake avoided: The three risk types do not have equal probability. Averaging the three conditional means or variances with weights of one third would not describe the stated model.

Worked Example: Structural Variation in a Hypothetical Portfolio

For a hypothetical portfolio, annual loss $X$ is measured in dollars. A policy’s risk type is low, medium, or high, with probabilities $0.25,0.50,0.25$. The conditional means are $40,60,80$ dollars and the conditional variances are $100,144,196$ dollars squared, respectively. Calculate the grand mean, EPV, VHM, and total variance. Explain which component is larger.

Method. Use the portfolio’s risk-type probabilities to average the conditional variances and to measure the dispersion of the conditional means around their grand mean.

Step 1: Grand mean.

E[X]=0.25(40)+0.50(60)+0.25(80)=10+30+20=60\text{ dollars}.

Step 2: EPV.

\mathrm{EPV}=0.25(100)+0.50(144)+0.25(196)=25+72+49=146\text{ dollars}^2.

Step 3: VHM. The low and high conditional means are each $20$ dollars from the grand mean, while the medium mean equals it:

\mathrm{VHM}=0.25(40-60)^2+0.50(60-60)^2+0.25(80-60)^2 =0.25(400)+0+0.25(400)=200\text{ dollars}^2.

Step 4: Check and interpret. The unconditional variance is $146+200=346$ dollars squared. Here between-risk variation, $200$, exceeds average within-risk variation, $146$. This comparison is about variance components, not a comparison of dollar amounts with squared-dollar amounts: both components are variances and therefore have the same squared-dollar units.

Common mistake avoided: VHM is the variance of the conditional means, not the variance of the conditional variances. The values $100,144,196$ are used to calculate EPV; the values $40,60,80$ are used to calculate VHM.

Worked Example: Continuous Risk Parameter

A hypothetical annual loss $X$ is measured in dollars. The risk parameter $\Theta$ has a uniform distribution on $[0,2]$. Conditional on $\Theta=\theta$, the mean is $30+5\theta$ dollars and the variance is $10+2\theta+\theta^2$ dollars squared. Find the grand mean, EPV, VHM, and unconditional variance.

Method. Integrate the conditional mean and variance over the uniform risk-parameter distribution. For $\Theta\sim\operatorname{Uniform}(0,2)$, its density is $1/2$ on that interval, $E[\Theta]=1$, and $E[\Theta^2]=4/3$. Hence $\operatorname{Var}(\Theta)=4/3-1^2=1/3$.

Step 1: Grand mean.

E[X]=E[30+5\Theta]=30+5E[\Theta]=30+5(1)=35\text{ dollars}.

Step 2: EPV. Average the conditional variance:

\mathrm{EPV}=E[10+2\Theta+\Theta^2] =10+2(1)+\frac{4}{3} =\frac{40}{3}\text{ dollars}^2 \approx13.3333\text{ dollars}^2.

Step 3: VHM. The conditional mean is $30+5\Theta$, so its variance is $5^2\operatorname{Var}(\Theta)$:

\mathrm{VHM}=\operatorname{Var}(30+5\Theta) =25\operatorname{Var}(\Theta) =25\left(\frac{1}{3}\right) =\frac{25}{3}\text{ dollars}^2 \approx8.3333\text{ dollars}^2.

Step 4: Check. The unconditional variance is $40/3+25/3=65/3$, or approximately $21.6667$ dollars squared. The grand mean is $35$ dollars. The fractions add exactly, which is a useful check before rounding.

Common mistake avoided: The EPV is not the conditional variance evaluated at the average risk parameter. That shortcut would give $10+2(1)+1^2=13$, whereas the correct EPV is $E[10+2\Theta+\Theta^2]=10+2E[\Theta]+E[\Theta^2]=40/3$. In particular, $E[\Theta^2]$ is not $(E[\Theta])^2$.

Common Interpretation and Calculation Traps

The definitions are short, but several mix-ups can change the answer. First, EPV averages conditional variances; it is not the variance of the conditional means. Second, VHM measures the dispersion of the conditional means; it is not the average conditional variance. Third, the same distribution of $\Theta$ must be used throughout: each risk type’s probability weights the grand mean, EPV, and VHM.

For example, in the three-risk-type model above, an incorrect equal-weight calculation would give a grand mean of $(3+6+9)/3=6$ rather than $6.9$. It would give an EPV of $(4+7+10)/3=7$ rather than $7.9$, and a VHM of $\{(3-6)^2+(6-6)^2+(9-6)^2\}/3=6$ rather than $5.49$. The arithmetic is internally consistent for a different model in which the three risk types are equally likely; it is wrong for the stated probabilities $0.2,0.3,0.5$. Retaining the weights is not optional.

A further useful check is that both EPV and VHM are nonnegative, since they are expectations of variances or squared deviations. If either is negative, there is an arithmetic or modeling error. Their sum must equal the unconditional variance under the stated model. If only conditional means and variances are supplied, this identity determines the unconditional variance without requiring the full conditional distribution of $X$.

Key takeaway: EPV is the probability-weighted average of the variation within each risk; VHM is the variation across risk-specific means. Calculate them from the same risk-parameter distribution, and verify that $\operatorname{Var}(X)=\mathrm{EPV}+\mathrm{VHM}$.

Check Your Understanding

For each hypothetical model, calculate the requested quantities from the stated risk distribution. Keep the probabilities attached to their corresponding risk types.

A risk parameter $\Theta$ takes values $1,2,4$ with probabilities $0.2,0.5,0.3$. Given $\Theta=\theta$, $E[X\mid\Theta=\theta]=20+2\theta$ dollars and $\operatorname{Var}(X\mid\Theta=\theta)=9+\theta$ dollars squared. Find the grand mean, EPV, VHM, and unconditional variance. (Answer: grand mean $=24.8$ dollars; EPV $=11.4$ dollars squared; VHM $=4.96$ dollars squared; unconditional variance $=16.36$ dollars squared.)
A hypothetical risk type is A with probability $0.4$ and B with probability $0.6$. Their conditional means are $12$ and $18$ dollars, and their conditional variances are $5$ and $8$ dollars squared. Find the grand mean, EPV, VHM, and unconditional variance. (Answer: grand mean $=15.6$ dollars; EPV $=6.8$ dollars squared; VHM $=8.64$ dollars squared; unconditional variance $=15.44$ dollars squared.)
Let $\Theta\sim\operatorname{Uniform}(0,3)$. Given $\Theta=\theta$, let $E[X\mid\Theta=\theta]=10+2\theta$ dollars and $\operatorname{Var}(X\mid\Theta=\theta)=4+\theta^2$ dollars squared. Find the grand mean, EPV, VHM, and unconditional variance. (Answer: grand mean $=13$ dollars; EPV $=7$ dollars squared; VHM $=3$ dollars squared; unconditional variance $=10$ dollars squared.)
A hypothetical risk parameter takes values $-1,0,2$ with probabilities $0.25,0.50,0.25$. Given $\Theta=\theta$, the conditional mean is $5+\theta^2$ dollars and the conditional variance is $2+\theta$ dollars squared. Find the grand mean, EPV, VHM, and unconditional variance. (Answer: grand mean $=6.25$ dollars; EPV $=2.25$ dollars squared; VHM $=2.6875$ dollars squared; unconditional variance $=4.9375$ dollars squared.)

What You'll Learn