The Bühlmann-Straub Risk Model With Exposure Weights

Introduction

In credibility work, observations may represent very different amounts of experience. A rate based on a large exposure can be more stable than a rate based on a small exposure, even when both relate to the same underlying risk. The Bühlmann-Straub model represents this difference by assigning each observation an exposure weight.

The central modeling assumption is that conditional process variance decreases in inverse proportion to exposure. This tutorial defines that assumption, explains its connection to exposure-weighted aggregate amounts, and distinguishes conditional process variation from variation among risks. The model is aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus.

Definition: Let $\Theta_i$ denote the risk parameter for risk $i$, and let $X_{ij}$ be an observed experience rate for that risk in period $j$. The observation has specified exposure weight $w_{ij}>0$. In the Bühlmann-Straub model, $$ E[X_{ij}\mid\Theta_i]=\mu(\Theta_i),\qquad \operatorname{Var}(X_{ij}\mid\Theta_i)=\frac{\sigma^2(\Theta_i)}{w_{ij}}. $$ Here $\mu(\Theta_i)$ is the risk-specific conditional mean rate, and $\sigma^2(\Theta_i)$ is the conditional process variance at one unit of exposure. Observations for the same risk are assumed conditionally independent given $\Theta_i$. The model does not require a particular conditional distribution, such as a normal distribution.

An exposure weight is a relative measure of the amount of experience represented by an observation. It may count policies, exposure-years, or another appropriate exposure unit. In the formulas below, $w_{ij}$ is treated as a positive numerical weight measured in a consistent exposure scale. A rate $X_{ij}$ is measured per exposure unit; its variance is therefore in squared rate units. The parameter $\sigma^2(\Theta_i)$ has those same squared rate units, and dividing it by a dimensionless weight preserves them.

Why the Rate Variance Has an Exposure Weight in the Denominator

The inverse-weight relationship has a direct explanation when an experience rate is an average. Suppose an observation with integer weight $w$ is the average of $w$ conditionally independent, equal-exposure measurements, each with conditional mean $\mu(\Theta)$ and conditional variance $\sigma^2(\Theta)$. Conditional on $\Theta$, the average has the same mean as each measurement. Its variance is the variance of a sum divided by $w^2$.

\operatorname{Var}\left(\left.\frac{1}{w}\sum_{k=1}^{w} Z_k\right|\Theta\right) =\frac{1}{w^2}\sum_{k=1}^{w}\operatorname{Var}(Z_k\mid\Theta) =\frac{w\sigma^2(\Theta)}{w^2} =\frac{\sigma^2(\Theta)}{w}.

The middle equality uses conditional independence, which makes the conditional covariances zero. The standard Bühlmann-Straub specification uses this inverse-weight variance relationship directly, including when weights are not integer counts. For a fractional or otherwise constructed exposure weight, the relationship is a modeling assumption; it need not come from a literal average of that many independent observations.

Larger weight reduces process variance for the rate, but it does not change the risk-specific conditional mean. The model therefore expresses a precision difference, not a claim that high-exposure risks have intrinsically different expected rates.

Worked Example: Conditional Rate Variance at One Exposure Weight

A hypothetical coverage portfolio uses an experience rate measured in dollars per standardized exposure unit. For a particular risk, suppose $\mu(\Theta)=120$ dollars per exposure unit and $\sigma^2(\Theta)=240$ (dollars per exposure unit)$^2$. One observed rate has weight $w=4$. Find its conditional mean, variance, and standard deviation. Also find the conditional mean and variance of the corresponding exposure-weighted aggregate amount.

Method. Apply the model to the rate first. Then define the corresponding aggregate amount as $S=wX$, so its conditional moments follow by multiplying by $w$ and $w^2$, respectively.

The conditional rate mean is $$ E[X\mid\Theta]=120\text{ dollars per exposure unit}. $$ The conditional rate variance is $$ \operatorname{Var}(X\mid\Theta)=\frac{240}{4} =60\text{ (dollars per exposure unit)}^2. $$ Thus the conditional rate standard deviation is $\sqrt{60}=7.7460$ dollars per exposure unit, rounded to four decimal places.

The aggregate amount has conditional mean $$ E[S\mid\Theta]=E[4X\mid\Theta]=4(120)=480\text{ dollars}. $$ Its conditional variance is $$ \operatorname{Var}(S\mid\Theta) =\operatorname{Var}(4X\mid\Theta) =4^2(60)=960\text{ dollars}^2. $$ This also equals $w\sigma^2(\Theta)=4(240)=960\text{ dollars}^2$, a useful check. The common mistake is to divide the aggregate variance by $w$, as if it were a rate. The aggregate amount is $w$ times the rate, so its variance is $w^2$ times the rate variance.

Rates and Exposure-Weighted Aggregate Amounts

For an observation with weight $w_{ij}$, define its exposure-weighted aggregate amount by $S_{ij}=w_{ij}X_{ij}$. Conditional on the risk parameter, the rate and aggregate moments are related as follows:

E[S_{ij}\mid\Theta_i]=w_{ij}\mu(\Theta_i),\qquad \operatorname{Var}(S_{ij}\mid\Theta_i) =w_{ij}^2\frac{\sigma^2(\Theta_i)}{w_{ij}} =w_{ij}\sigma^2(\Theta_i).

This conversion is useful because the rate variance decreases with weight while the aggregate variance increases with weight. There is no contradiction: a higher-weight aggregate is a larger amount, but its rate is more stable. Keep the quantity being modeled explicit before applying a variance formula.

Worked Example: Combining Conditional Aggregate Amounts

A hypothetical insurer records experience for one risk in three periods, with weights $1$, $4$, and $2$. Assume the conditional mean rate is $75$ dollars per exposure unit, the conditional process variance is $180$ (dollars per exposure unit)$^2$, and the three period observations are conditionally independent given the risk parameter. Find the conditional mean and variance of each period’s aggregate amount and of their sum.

Method. For each period, multiply the rate mean by its weight and the rate variance by the square of its weight. Equivalently, the aggregate variance for weight $w$ is $w\sigma^2(\Theta)$. Conditional independence then permits the aggregate variances to be added.

For weight $1$, the aggregate mean is $1(75)=75$ dollars and the aggregate variance is $1(180)=180$ dollars$^2$. For weight $4$, the aggregate mean is $4(75)=300$ dollars and the aggregate variance is $4(180)=720$ dollars$^2$. For weight $2$, the aggregate mean is $2(75)=150$ dollars and the aggregate variance is $2(180)=360$ dollars$^2$.

The conditional mean of the sum is $$ 75+300+150=525\text{ dollars}. $$ By conditional independence, its conditional variance is $$ 180+720+360=1{,}260\text{ dollars}^2. $$ As a check, the total weight is $1+4+2=7$, so the sum has conditional mean $7(75)=525$ dollars and conditional variance $7(180)=1{,}260$ dollars$^2$. The mistake to avoid is adding rate variances $180+45+90$ and reporting that as the aggregate total variance; those are variances of rates, not of aggregate amounts.

Conditional and Unconditional Variation

The conditional model describes process variation for a fixed risk parameter. Across a population of risks, however, the conditional mean $\mu(\Theta)$ can itself vary. These are distinct sources of variation. As in the earlier material on EPV and VHM, the expected process variance is the expected value of $\sigma^2(\Theta)$, while the variation in risk-specific means is $\operatorname{Var}(\mu(\Theta))$.

For a rate with weight $w$, the law of total variance gives $$ \operatorname{Var}(X) =E[\operatorname{Var}(X\mid\Theta)] +\operatorname{Var}(E[X\mid\Theta]) =\frac{E[\sigma^2(\Theta)]}{w}+\operatorname{Var}(\mu(\Theta)). $$ The first term is process variation and is reduced by exposure. The second is variation among the underlying risk means and remains even at high exposure.

Conditional independence does not generally mean unconditional independence. Two observations from the same risk can share the same unknown $\Theta$. Their conditional process errors are independent, but their conditional means move together as $\Theta$ varies. If the observations have weights $w_1,w_2$, then $$ \operatorname{Cov}(X_1,X_2) =E[\operatorname{Cov}(X_1,X_2\mid\Theta)] +\operatorname{Cov}(E[X_1\mid\Theta],E[X_2\mid\Theta]) =\operatorname{Var}(\mu(\Theta)), $$ where the first term is zero by conditional independence.

Worked Example: Shared Risk Variation and Unconditional Covariance

Suppose a hypothetical group of risks has $E[\mu(\Theta)]=100$ dollars per exposure unit and $\operatorname{Var}(\mu(\Theta))=64$ (dollars per exposure unit)$^2$. For every risk, let $\sigma^2(\Theta)=200$ (dollars per exposure unit)$^2$. Two observations for the same risk have weights $1$ and $4$ and are conditionally independent given $\Theta$. Find their unconditional variances and covariance, and the variance of the sum of their exposure-weighted aggregate amounts.

Method. Apply the law of total variance to each rate and the law of total covariance to the pair. Then use $S_1=X_1$ and $S_2=4X_2$ to calculate the aggregate-sum variance.

For the first rate, $$ \operatorname{Var}(X_1)=\frac{200}{1}+64=264\text{ (dollars per exposure unit)}^2. $$ For the second rate, $$ \operatorname{Var}(X_2)=\frac{200}{4}+64=50+64=114\text{ (dollars per exposure unit)}^2. $$ Their unconditional covariance is the variance of the shared conditional mean: $$ \operatorname{Cov}(X_1,X_2)=64\text{ (dollars per exposure unit)}^2. $$ The aggregate sum is $S_1+S_2=X_1+4X_2$. Therefore, $$ \operatorname{Var}(S_1+S_2) =264+4^2(114)+2(1)(4)(64) =264+1{,}824+512 =2{,}600\text{ dollars}^2. $$ An independent check uses total weight $1+4=5$. Conditional on $\Theta$, the aggregate sum has variance $5(200)=1{,}000$ dollars$^2$ and mean $5\mu(\Theta)$, whose variance is $5^2(64)=1{,}600$ dollars$^2$. Their sum is $1{,}000+1{,}600=2{,}600$ dollars$^2$. Treating the two rates as unconditionally independent would omit the covariance term and understate the result.

Assumptions and Exam Traps

The exposure-weighted model is compact, but using it correctly requires attention to what each observation measures and to the conditioning assumptions.

Specify a rate or an aggregate. The model’s $X_{ij}$ is an experience rate. Its variance is $\sigma^2(\Theta_i)/w_{ij}$. For the aggregate $S_{ij}=w_{ij}X_{ij}$, the variance is $w_{ij}\sigma^2(\Theta_i)$.
Use consistent exposure units. The weights must use a common exposure scale for the observations being compared. A weight of $4$ means four times the model’s reference exposure, not necessarily four policies unless the reference unit is one policy.
Condition on the same risk parameter. Conditional independence is an assumption about observations given $\Theta_i$; it does not remove their unconditional dependence through the shared risk mean.
Do not assume a distribution unnecessarily. The conditional mean and variance specification does not, by itself, say that observations are normal, Poisson, or any other particular distribution.
Write variance units without ambiguous squares. For example, $\sigma^2(\Theta)=240$ (dollars per exposure unit)$^2$ has numerical variance value $240$, not $240^2$. At weight $4$, the rate variance is $60$ (dollars per exposure unit)$^2$, and its standard deviation is $\sqrt{60}$, not $\sqrt{60^2}$.

The main value of the model is its separation of experience volume from risk level. Exposure determines the precision of the observed rate through the inverse-weight process variance; the latent risk parameter determines the rate’s conditional mean. Keeping those roles and their units distinct is essential for sound credibility calculations.

Key takeaway: For a Bühlmann-Straub rate, $E[X_{ij}\mid\Theta_i]=\mu(\Theta_i)$ and $\operatorname{Var}(X_{ij}\mid\Theta_i)=\sigma^2(\Theta_i)/w_{ij}$. Multiplying the rate by its weight gives an aggregate amount with conditional mean $w_{ij}\mu(\Theta_i)$ and conditional variance $w_{ij}\sigma^2(\Theta_i)$.

Check Your Understanding

For each problem, identify whether the requested quantity is a rate or an aggregate amount before calculating.

A rate has conditional process variance $315$ (dollars per exposure unit)$^2$ and weight $5$. Find its conditional variance and standard deviation. (Answer: variance $=63$ (dollars per exposure unit)$^2$; standard deviation $=\sqrt{63}\approx7.9373$ dollars per exposure unit.)
A risk has conditional mean rate $48$ dollars per exposure unit, process variance $96$ (dollars per exposure unit)$^2$, and weight $3$. Find the corresponding aggregate amount’s conditional mean and variance. (Answer: mean $=\$144$; variance $=288$ dollars$^2$.)
Two conditionally independent aggregate observations for the same risk have weights $2$ and $3$. Their conditional mean rate is $70$ dollars per exposure unit and their process variance is $150$ (dollars per exposure unit)$^2$. Find the conditional mean and variance of the sum of the two aggregate amounts. (Answer: mean $=\$350$; variance $=750$ dollars$^2$.)
Across risks, $E[\mu(\Theta)]=80$ dollars per exposure unit, $\operatorname{Var}(\mu(\Theta))=25$ (dollars per exposure unit)$^2$, and $E[\sigma^2(\Theta)]=100$ (dollars per exposure unit)$^2$. Find the unconditional variance of a rate with weight $4$. (Answer: $50$ (dollars per exposure unit)$^2$.)
For the same risk, two conditionally independent rates have weights $1$ and $2$. Suppose $E[\mu(\Theta)]=90$ dollars per exposure unit, $\operatorname{Var}(\mu(\Theta))=36$ (dollars per exposure unit)$^2$, and $\sigma^2(\Theta)=120$ (dollars per exposure unit)$^2$. Find the covariance of the rates and the variance of the sum of their weighted aggregate amounts. (Answer: covariance $=36$ (dollars per exposure unit)$^2$; aggregate-sum variance $=684$ dollars$^2$.)

What You'll Learn