Expected Process Variance

Why Average Conditional Variance?

In a credibility model, the risk parameter represents a characteristic of a risk that is treated as fixed when describing that risk’s outcomes. For example, it might represent a claim frequency, a claim probability, or a mean loss level. Even after that parameter is fixed, an individual outcome can still vary from one observation period to another. Expected process variance measures the average size of that remaining, within-risk variation.

The word “expected” matters: different risks can have different conditional variances. To summarize the process variation for a randomly selected risk, we average those conditional variances using the distribution of the risk parameter. This is distinct from measuring how the conditional means vary across risks, the subject of a separate Bühlmann model component.

Definition: Let $\Theta$ be a risk parameter with a specified distribution, and let $X$ be an outcome with finite conditional second moments. The expected process variance is $$ \operatorname{EPV}=E[\operatorname{Var}(X\mid\Theta)]. $$ For a fixed value $\Theta=\theta$, $\operatorname{Var}(X\mid\Theta=\theta)$ measures the variation in outcomes for risks with that parameter value. The expectation averages those variances over the distribution of $\Theta$. The result is measured in squared units of $X$.

The conditional variance itself is defined by $$ \operatorname{Var}(X\mid\Theta)=E\left[(X-E[X\mid\Theta])^2\mid\Theta\right]. $$ The conditional mean is subtracted before squaring. Thus, this variance measures deviations of outcomes from the mean for the same fixed risk parameter; differences between the means of different risks are not part of that conditional variance.

How the Averaging Works

If the risk parameter has a finite set of possible values $\theta_1,\ldots,\theta_k$, with probabilities $w_1,\ldots,w_k$, then the weights are nonnegative and sum to one. The expected process variance is the probability-weighted average of the conditional variances:

\operatorname{EPV} = \sum_{j=1}^{k} w_j\operatorname{Var}(X\mid\Theta=\theta_j), \qquad \sum_{j=1}^{k}w_j=1.

If $\Theta$ is continuous with density $g(\theta)$, the corresponding average is an integral:

\operatorname{EPV} = \int \operatorname{Var}(X\mid\Theta=\theta)g(\theta)\,d\theta.

These expressions follow directly from the definition of expectation. For a discrete risk parameter, the random quantity being averaged is $\operatorname{Var}(X\mid\Theta)$. It equals $\operatorname{Var}(X\mid\Theta=\theta_j)$ when $\Theta=\theta_j$, an event with probability $w_j$. Therefore its expectation is the sum of each value times its probability. For a continuous parameter, the same expectation is obtained by integrating the conditional variance against the density of $\Theta$.

Interpretation: Expected process variance is an average of within-risk variances, not an average of standard deviations. It has squared outcome units. Taking its square root gives a single standard-deviation scale for the average process variance, but that is not generally the average of the risks’ conditional standard deviations.

Worked Applications

Worked Example: Claim Counts Across Two Risk Types

A hypothetical portfolio contains two types of risks. A randomly selected risk is Type A with probability $0.40$ and Type B with probability $0.60$. Conditional on its type, its annual claim count $X$ has a Poisson distribution: Type A has mean $2$, and Type B has mean $5$. Find the expected process variance of annual claim count.

Method. Use the finite-risk weighted-average formula. For a Poisson random variable, its conditional variance equals its conditional mean. The probabilities and conditional distributions here fully specify the risk mixture.

Step 1: Find each conditional variance. For Type A, $X\mid A\sim\operatorname{Poisson}(2)$, so $\operatorname{Var}(X\mid A)=2$ claims squared. For Type B, $X\mid B\sim\operatorname{Poisson}(5)$, so $\operatorname{Var}(X\mid B)=5$ claims squared.

Step 2: Weight by the probabilities of the risk types.

\operatorname{EPV} =0.40(2)+0.60(5) =0.80+3.00 =3.80\ \text{claims}^2.

Check and common mistake avoided. The weights sum to $0.40+0.60=1.00$, and the answer lies between the two conditional variances, as a weighted average should. A common mistake is to use the difference between the conditional means, $5-2=3$, as the process variance. That difference describes separation between the two type-specific means; it is not the average within-type variance. The expected process variance is $3.80$ claims squared.

Worked Example: Binomial Outcomes With a Beta Risk Parameter

For a hypothetical group of insured devices, let $\Theta$ be the unknown failure probability for a device. Across devices, $\Theta\sim\operatorname{Beta}(2,2)$. Conditional on $\Theta=\theta$, the number $X$ of failures in the next four independent operating cycles has a $\operatorname{Binomial}(4,\theta)$ distribution. Find the expected process variance of $X$.

Method. First use the binomial conditional variance, then average it over the beta distribution of the risk parameter. The conditional independence of the four cycles is what makes the conditional count binomial.

Step 1: Write the conditional variance. Given $\Theta=\theta$, a binomial count with four trials and failure probability $\theta$ has variance $4\theta(1-\theta)$. Hence

\operatorname{EPV} =E[4\Theta(1-\Theta)] =4\left(E[\Theta]-E[\Theta^2]\right).

Step 2: Calculate the two moments. For a $\operatorname{Beta}(\alpha,\beta)$ variable, $E[\Theta]=\alpha/(\alpha+\beta)$ and $E[\Theta^2]=\alpha(\alpha+1)/[(\alpha+\beta)(\alpha+\beta+1)]$. With $\alpha=\beta=2$,

E[\Theta]=\frac{2}{4}=0.5, \qquad E[\Theta^2]=\frac{2(3)}{4(5)}=\frac{6}{20}=0.3.

Step 3: Evaluate the expected conditional variance.

\operatorname{EPV} =4(0.5-0.3) =4(0.2) =0.8\ \text{failures}^2.

Check and common mistake avoided. For any fixed $\theta$ between 0 and 1, $4\theta(1-\theta)$ cannot exceed $1$, so an average of these conditional variances of $0.8$ is plausible. A common error is to substitute the mean failure probability into the conditional variance and stop: $4(0.5)(1-0.5)=1.0$. That computes the conditional variance at the average parameter, not the average conditional variance. The requested quantity averages $4\Theta(1-\Theta)$ over the distribution of $\Theta$, giving $0.8$.

Worked Example: Process Variance of Exposure-Adjusted Claim Rates

A hypothetical portfolio has two risk groups. A randomly selected policy belongs to Group A with probability $0.60$, where annual exposure is $20$ vehicle-years and the claim rate parameter is $\theta_A=0.03$ claims per vehicle-year. It belongs to Group B with probability $0.40$, where exposure is $50$ vehicle-years and $\theta_B=0.04$. Conditional on the group, the annual claim count $N$ is Poisson with mean equal to exposure times claim rate. Define the observed rate $X=N/e$, where $e$ is the group’s exposure. Find the expected process variance of $X$.

Method. Calculate the conditional variance of the rate, not the variance of the raw count. Exposure is fixed within each group, so scaling the count by $1/e$ scales its variance by $1/e^2$.

Step 1: Find each group’s conditional count mean and variance. In Group A, $e\theta_A=20(0.03)=0.60$, so $N\mid A\sim\operatorname{Poisson}(0.60)$ and $\operatorname{Var}(N\mid A)=0.60$. In Group B, $e\theta_B=50(0.04)=2.00$, so $N\mid B\sim\operatorname{Poisson}(2.00)$ and $\operatorname{Var}(N\mid B)=2.00$.

Step 2: Convert each count variance to a rate variance. Since $X=N/e$,

\operatorname{Var}(X\mid A) =\frac{0.60}{20^2} =\frac{0.60}{400} =0.0015, \qquad \operatorname{Var}(X\mid B) =\frac{2.00}{50^2} =\frac{2.00}{2500} =0.0008.

The units are squared claims per vehicle-year. The rate variance formula can also be written $\theta/e$, because a Poisson count with mean $e\theta$ has variance $e\theta$, and $\operatorname{Var}(N/e\mid e,\theta)=e\theta/e^2$.

Step 3: Average across the two groups.

\operatorname{EPV} =0.60(0.0015)+0.40(0.0008) =0.0009+0.00032 =0.00122.

Check and common mistake avoided. The result is between the two group-specific rate variances, $0.0008$ and $0.0015$. A common mistake is to average the raw count variances and divide by the square of the average exposure. That mixes outcome scales: the groups have different exposures, and each rate is divided by its own group’s exposure. The correct expected process variance is $0.00122$ squared claims per vehicle-year.

What the Component Does—and Does Not—Measure

Expected process variance is useful because an individual risk’s observed outcome is not determined by its risk parameter alone. Random outcomes fluctuate even when the underlying risk characteristic is held fixed. In the claim-count example, a Type A risk can have different annual counts in different years despite having the same Poisson parameter. The conditional variance quantifies that fluctuation for Type A; expected process variance averages this quantity across the portfolio’s risk types.

Keep three distinctions clear when setting up a calculation:

Condition first. Identify the outcome distribution after fixing the risk parameter. Use its conditional variance, not an unconditional variance.
Average using the risk distribution. The weights are probabilities of the parameter values or risk groups, not arbitrary equal weights unless the groups are equally likely.
Preserve the outcome scale. A count, a rate, and a dollar loss have different units. If the outcome is rescaled, transform its variance accordingly before averaging.

A further trap is averaging conditional standard deviations instead of conditional variances. In general, $E[\sqrt{\operatorname{Var}(X\mid\Theta)}]$ is not the same quantity as $\sqrt{E[\operatorname{Var}(X\mid\Theta)]}$. The definition calls for averaging variances first. If a standard-deviation summary is desired, take the square root after that average.

Specify the outcome and risk parameter.
State what $X$ measures and what is treated as fixed when conditioning on $\Theta$.

Find the conditional variance.
Use the distribution of $X\mid\Theta$, retaining any exposure or other fixed scaling relevant to the outcome.

Average across risks.
Weight the conditional variances by the probabilities of the risk types, or integrate them against the distribution of $\Theta$.

Check units and range.
The result has squared outcome units and, for a finite mixture, lies between the smallest and largest conditional variances with positive weight.

Key takeaway: Expected process variance is $E[\operatorname{Var}(X\mid\Theta)]$: it averages the random variation that remains within risks after their parameters are fixed. It is not a measure of how different the risks’ conditional means are.

Check Your Understanding

For each problem, calculate the expected process variance of the stated outcome, carefully preserving its scale.

A hypothetical risk is Type C with probability $0.25$ and Type D with probability $0.75$. Conditional annual claim counts are Poisson with means $1.2$ and $3.2$, respectively. Find the expected process variance of annual claim count. (Answer: $2.7$ claims squared)
A hypothetical policy has a claim count $X$ that, conditional on $\Theta=\theta$, is $\operatorname{Binomial}(5,\theta)$. The risk parameter is $0.2$ with probability $0.5$ and $0.6$ with probability $0.5$. Find the expected process variance. (Answer: $1.0$ claims squared)
A single hypothetical exposure has a Bernoulli outcome $X$, with $P(X=1\mid\Theta=\theta)=\theta$. The risk parameter is $0.1$ or $0.5$, each with probability $0.5$. Find the expected process variance. (Answer: $0.17$ outcome units squared)
A hypothetical policy has $30$ vehicle-years of exposure and annual count $N\mid\theta\sim\operatorname{Poisson}(30\theta)$. Its rate parameter is $0.02$ with probability $0.4$ and $0.06$ with probability $0.6$. For $X=N/30$, find the expected process variance. (Answer: $0.0014667$ squared claims per vehicle-year)
A hypothetical severity outcome $X$, measured in dollars, has a conditional normal distribution. For 70% of risks, $X\mid\Theta=A\sim N(1{,}200,250^2)$; for 30%, $X\mid\Theta=B\sim N(1{,}800,400^2)$. Find the expected process variance. (Answer: $\$91{,}750$ squared)

What You'll Learn