How K Controls Credibility Across Volumes

Why Compare K Across Risks?

A Bühlmann credibility weight depends on two quantities: the number of comparable observations $n$ and the structural parameter $K$. Comparing $K$ values alone does not tell you which risk receives more weight on its own experience. The volume matters too: a risk with a larger $K$ can still have the same, or greater, credibility if it has enough additional observations.

The purpose of this comparison is to separate the effects. Holding $n$ fixed isolates the effect of $K$; holding $K$ fixed isolates the effect of volume. When both change, comparing the ratio $n/K$ gives a direct way to compare the resulting weights.

Definition: Let $\Theta$ represent a risk’s unknown risk parameter and let $X_1,\ldots,X_n$ be equal-volume observations that are conditionally independent and identically distributed given $\Theta$. Write $\mu(\Theta)=E[X_j\mid\Theta]$, and let $m=E[\mu(\Theta)]$ be the collective mean. The expected process variance is $\mathrm{EPV}=E[\operatorname{Var}(X_j\mid\Theta)]$; the variance of hypothetical means is $\mathrm{VHM}=\operatorname{Var}(\mu(\Theta))$. When $\mathrm{EPV}>0$ and $\mathrm{VHM}>0$, $K=\mathrm{EPV}/\mathrm{VHM}$, and the Bühlmann weight on the individual sample mean $\bar X$ is $Z=n/(n+K)$.

The units must match: $n$ counts observations on the chosen, equal-volume basis, and $K$ is measured in that same observation unit. If observations have unequal exposure, the raw count cannot generally be substituted for $n$; the exposure basis must first be handled appropriately. The formula also presumes the same structural model applies to the observations being combined.

A larger EPV means more variation among observations from the same risk, making a given sample mean less informative about that risk’s expected value. A larger VHM means greater variation in expected values between risks, making individual experience more useful for distinguishing a risk from the collective. Since $K$ is EPV divided by VHM, more process noise relative to risk heterogeneity produces a larger $K$.

How the Two Inputs Control the Weight

For positive $n$ and $K$, the weight can also be written as $Z=1/(1+K/n)$. This form makes clear that the comparison depends on volume relative to $K$, not on either number in isolation. If two risks have the same $n/K$, they have the same $Z$. If they do not, the risk with the larger $n/K$ has the larger weight on its sample mean.

Z=\frac{n}{n+K}=\frac{1}{1+K/n}

The direction of each effect can be checked directly. Treating $K$ as fixed, increasing $n$ increases $Z$. Treating $n$ as fixed, increasing $K$ decreases $Z$. These are comparisons within the stated model: if a change in observation volume also changes the underlying risk mix or structural parameters, the comparison must account for those changes as well.

\frac{\partial Z}{\partial n}=\frac{K}{(n+K)^2}>0, \qquad \frac{\partial Z}{\partial K}=-\frac{n}{(n+K)^2}<0

Why the Ratio Produces This Weight

The comparison formula follows from the mean-squared-error criterion for a linear credibility estimate. Under the model above, the estimate is $m+Z(\bar X-m)$. Its error balances two sources: shrinking toward $m$ can leave the risk-specific mean partly unrecognized, while putting weight on $\bar X$ introduces sampling variation.

To see the balance, write $\bar X=\mu(\Theta)+(\bar X-\mu(\Theta))$. Conditional independence and equal conditional variances give $E[\bar X\mid\Theta]=\mu(\Theta)$ and $E[\operatorname{Var}(\bar X\mid\Theta)]=\mathrm{EPV}/n$. The process deviation has conditional mean zero, so its cross-product with the deviation of $\mu(\Theta)$ from $m$ has expectation zero. The mean-squared error is therefore

E[(m+Z(\bar X-m)-\mu(\Theta))^2] =(1-Z)^2\mathrm{VHM}+Z^2\frac{\mathrm{EPV}}{n}.

Differentiate with respect to $Z$, set the derivative to zero, and solve. The second derivative is positive, so this stationary point is the minimum.

-2(1-Z)\mathrm{VHM}+2Z\frac{\mathrm{EPV}}{n}=0 \quad\Longrightarrow\quad Z=\frac{\mathrm{VHM}}{\mathrm{VHM}+\mathrm{EPV}/n} =\frac{n}{n+\mathrm{EPV}/\mathrm{VHM}} =\frac{n}{n+K}.

This justification shows why $K$ is not a stand-alone measure of credibility. It summarizes the relative size of two kinds of variation, while $n$ determines how much process variation is averaged away. In particular, $\mathrm{EPV}/n$ is the expected sampling variance of the sample mean under equal-volume observations.

Worked Comparisons

The following examples use hypothetical annual aggregate loss-cost observations as a normal proxy. In each, conditional observations are independent with $X_j\mid\Theta\sim N(\Theta,\mathrm{EPV})$, where the second normal parameter is the variance. The risk parameter $\Theta$ has two equally likely values, $m-\sqrt{\mathrm{VHM}}$ and $m+\sqrt{\mathrm{VHM}}$, so its mean is $m$ and its variance is VHM. This proxy can produce negative values; it is used here to illustrate the credibility calculations, not to model individual claim severities. No discounting is applied (interest rate 0%); mortality is not modeled and is not applicable to these annual observations.

Worked Example: Same Volume, Different K

Two hypothetical risk groups each have $n=4$ annual observations and collective mean $m=\$1{,}000$. Group A has EPV $\$800^2$ and VHM $\$400^2$; Group B has EPV $\$800^2$ and VHM $\$100^2$. For each group, the four-observation sample mean is $\$1{,}200$. Compare their credibility estimates.

Method and structural parameters. In the stated conditional normal model, the conditional variance is $800^2=640{,}000$ for either group. Group A’s risk parameter is equally likely to be $1{,}000-400=\$600$ or $1{,}000+400=\$1{,}400$, giving VHM $400^2=160{,}000$. Group B’s risk parameter is equally likely to be $\$900$ or $\$1{,}100$, giving VHM $100^2=10{,}000$. Thus both groups have the same process variance, but different between-risk variation.

Group A. Its structural parameter is $K=640{,}000/160{,}000=4$. With $n=4$, its weight is $Z=4/(4+4)=0.5$. The credibility estimate is $0.5(\$1{,}200)+0.5(\$1{,}000)=\$1{,}100$.

Group B. Its structural parameter is $K=640{,}000/10{,}000=64$. Its weight is $Z=4/(4+64)=4/68=1/17$. The estimate is $(1/17)(\$1{,}200)+(16/17)(\$1{,}000)=\$1{,}011.7647$, or $\$1{,}011.76$ to the nearest cent.

The check is that both estimates lie between the collective mean and the sample mean. Group A receives more individual weight because its VHM is greater, while EPV and volume are unchanged. The common mistake is to assume equal volumes imply equal credibility; the structural $K$ values still differ.

Worked Example: Same Volume, Different Process and Risk Variation

Two other hypothetical groups each have $n=12$ annual observations and collective mean $m=\$1{,}000$. Group C has EPV $900$ and VHM $300$; Group D has EPV $1{,}200$ and VHM $200$. Each group’s observed sample mean is $\$1{,}150$. Find and compare their credibility estimates. The conditional normal variance is the stated EPV, not its square.

Group C. Its risk parameter is equally likely to be $1{,}000-\sqrt{300}$ or $1{,}000+\sqrt{300}$, so the mean is $\$1{,}000$ and the VHM is 300. Its conditional distribution is $X_j\mid\Theta\sim N(\Theta,900)$. Hence $K=900/300=3$, and $Z=12/(12+3)=0.8$. The estimate is $0.8(\$1{,}150)+0.2(\$1{,}000)=\$1{,}120$.

Group D. Its risk parameter is equally likely to be $1{,}000-\sqrt{200}$ or $1{,}000+\sqrt{200}$, and $X_j\mid\Theta\sim N(\Theta,1{,}200)$. Thus $K=1{,}200/200=6$, and $Z=12/(12+6)=2/3$. The estimate is $(2/3)(\$1{,}150)+(1/3)(\$1{,}000)=\$1{,}100$.

The estimates are between $\$1{,}000$ and $\$1{,}150$, as required by their weights. Group D has both greater process variance and smaller VHM, so its $K$ is larger and its credibility weight is smaller at the same volume. The common mistake is to compare EPV alone: the denominator, VHM, matters just as directly through the ratio.

Worked Example: Different Volumes, Equal Credibility

Group E has $n=6$, EPV $900$, and VHM $300$. Group F has $n=12$, EPV $1{,}200$, and VHM $200$. Both have collective mean $\$1{,}000$; their sample means are $\$1{,}300$ for E and $\$1{,}150$ for F. Compare the weights and estimates.

Specify the hypothetical models. For E, $X_j\mid\Theta\sim N(\Theta,900)$, with $\Theta$ equally likely to be $1{,}000-\sqrt{300}$ or $1{,}000+\sqrt{300}$. For F, $X_j\mid\Theta\sim N(\Theta,1{,}200)$, with $\Theta$ equally likely to be $1{,}000-\sqrt{200}$ or $1{,}000+\sqrt{200}$. As above, interest is 0% and mortality is not modeled.

For E, $K=900/300=3$, so $Z=6/(6+3)=2/3$. Its estimate is $(2/3)(\$1{,}300)+(1/3)(\$1{,}000)=\$1{,}200$. For F, $K=1{,}200/200=6$, so $Z=12/(12+6)=2/3$. Its estimate is $(2/3)(\$1{,}150)+(1/3)(\$1{,}000)=\$1{,}100$.

The equal weights are checked by the equal ratios $n/K$: $6/3=12/6=2$. The estimates nevertheless differ because the sample means differ. The common mistake is to infer that equal credibility weights imply equal credibility estimates; the weight specifies how each group’s own sample mean is blended with the collective mean.

Common Exam Trap and Practical Comparison

A frequent error is to reverse the roles of $n$ and $K$. For example, if $n=4$ and $K=8$, the incorrect calculation $Z=K/(n+K)=8/12=2/3$ assigns more weight to individual experience as $K$ increases. The correct calculation is $Z=n/(n+K)=4/12=1/3$. This direction is consistent with the variance interpretation: a large EPV relative to VHM means each observation is relatively noisy about which risk-specific mean applies.

A reliable comparison proceeds in a fixed order: establish the observation unit and volume, calculate EPV and VHM using the model’s definitions, form $K$, and then calculate $Z$. When comparing two cases, state what is held fixed. If both $n$ and $K$ change, compare $n/K$, or calculate both weights explicitly. Do not conclude that the larger-volume risk necessarily has the larger weight if its $K$ also changes.

Confirm comparable observations.
Check that the count $n$ is based on the same equal-volume observation unit and that the assumed structural model is appropriate.

Calculate the structural ratio.
Use $K=\mathrm{EPV}/\mathrm{VHM}$; do not treat $K$ as an observed claim count or as a credibility weight.

Compare the relative volume.
Use $Z=n/(n+K)$, equivalently $Z=1/(1+K/n)$, and explain whether a difference comes from volume, structure, or both.

For ratemaking interpretation, a lower individual weight does not mean that a risk’s experience is irrelevant. It means that, under the specified model and observation basis, its sample mean receives less weight in the linear credibility estimate. The comparison is useful only when the collective mean, experience measure, exposure unit, and structural variance definitions are aligned across the cases.

Check Your Understanding

Use the equal-volume Bühlmann model. Unless stated otherwise, treat each EPV and VHM as the model moments for independent conditional observations.

A risk has $n=5$, EPV $=600$, and VHM $=120$. Calculate $K$ and $Z$. (Answer: $K=5$, $Z=0.5000$.)
Two groups each have $n=9$. Group A has EPV $720$ and VHM $240$; Group B has EPV $900$ and VHM $300$. Compare their $K$ values and weights. (Answer: Both have $K=3$ and $Z=0.7500$.)
Group A has $n=8$, EPV $600$, and VHM $200$; Group B has $n=12$, EPV $900$, and VHM $300$. Which has the larger weight, and by how much? (Answer: A has $Z=8/11=0.7273$; B has $Z=0.8000$, larger by $0.0727$.)
Risk P has $n=6$, EPV $1{,}000$, and VHM $250$. Risk Q has $n=10$, EPV $1{,}500$, and VHM $250$. Calculate both weights and identify the larger one. (Answer: P has $K=4$, $Z=0.6000$; Q has $K=6$, $Z=0.6250$; Q’s weight is larger by $0.0250$.)
Risk R has $n=9$, EPV $800$, and VHM $400$. Risk S has EPV $1{,}200$ and VHM $200$. What volume must S have to match R’s credibility weight? (Answer: R has $K=2$, $Z=9/11$; S has $K=6$, so $n/ (n+6)=9/11$ gives $2n=54$, or $n=27$.)

What You'll Learn