The Buhlmann Credibility Factor from Bayesian Updating

Why a Credibility Factor Appears in Bayesian Updating

A premium for one risk should reflect both the risk's observed experience and what is known about the wider class. With little experience, the class mean should usually carry more weight; with more experience, the individual experience should usually matter more. Bayesian updating makes this trade-off explicit under a specified prior and likelihood. In the normal-normal model, the posterior mean is exactly a weighted average of the prior mean and the observed average. Its weight on observed experience has the same form as the Buhlmann credibility factor.

The connection is useful, but its scope matters. In a general Buhlmann model, the credibility premium is the best linear predictor of the risk's hypothetical mean. It need not equal the full Bayesian posterior mean. The exact equality holds for particular models, including the normal-normal model.

Definition: Let $\Theta$ represent a risk's latent type, and let $X_1,\ldots,X_n$ be conditionally independent observations with the same conditional distribution given $\Theta$. Define the hypothetical mean by $\mu(\Theta)=E[X_j\mid\Theta]$ and the conditional process variance by $v(\Theta)=\operatorname{Var}(X_j\mid\Theta)$. The expected process variance is $\operatorname{EPV}=E[v(\Theta)]$, and the variance of hypothetical means is $\operatorname{VHM}=\operatorname{Var}(\mu(\Theta))$. When both are finite and $\operatorname{VHM}>0$, set $K=\operatorname{EPV}/\operatorname{VHM}$. For $n$ equally informative observations, the Buhlmann factor is $Z_n=n/(n+K)$.

The grand mean is $m=E[\mu(\Theta)]=E[X_j]$, and the sample mean is $\bar X=(X_1+\cdots+X_n)/n$. The Buhlmann linear credibility premium is $Z_n\bar X+(1-Z_n)m$. Here the observations are equally informative and conditionally identically distributed; unequal exposures require a different treatment.

Deriving the Factor from a Normal Prior and Likelihood

First consider the normal-normal model. Suppose the latent claim-cost level has prior distribution $\Theta\sim N(\mu,\tau^2)$, and, conditional on $\Theta=\theta$, the observations are independent with $X_j\mid\Theta=\theta\sim N(\theta,\sigma^2)$. Both variances are positive and known. The prior mean is $\mu$, the prior variance is $\tau^2$, and each observation has conditional process variance $\sigma^2$.

Conditional independence gives a likelihood proportional to $\exp[-\sum_{j=1}^n(x_j-\theta)^2/(2\sigma^2)]$. The prior density contributes a factor proportional to $\exp[-(\theta-\mu)^2/(2\tau^2)]$. Multiplying them and collecting the terms involving $\theta$ gives a quadratic whose precision is $1/\tau^2+n/\sigma^2$, and whose precision-weighted center is $\mu/\tau^2+n\bar x/\sigma^2$. Completing the square therefore yields the posterior distribution and mean below.

\Theta\mid x_1,\ldots,x_n\sim N\left( \frac{\mu/\tau^2+n\bar x/\sigma^2}{1/\tau^2+n/\sigma^2}, \frac{1}{1/\tau^2+n/\sigma^2} \right), \qquad E[\Theta\mid x_1,\ldots,x_n] = \frac{\mu/\tau^2+n\bar x/\sigma^2}{1/\tau^2+n/\sigma^2}.

Multiplying numerator and denominator by $\sigma^2\tau^2$ rewrites the posterior mean as a weighted average. Its weight on the observed average is $n\tau^2/(n\tau^2+\sigma^2)$. Since this model has $\operatorname{EPV}=\sigma^2$ and $\operatorname{VHM}=\tau^2$, the Buhlmann ratio is $K=\sigma^2/\tau^2$. Thus the Bayesian weight and the Buhlmann factor coincide:

E[\Theta\mid x_1,\ldots,x_n] =(1-Z_n)\mu+Z_n\bar x, \qquad Z_n=\frac{n\tau^2}{n\tau^2+\sigma^2} =\frac{n}{n+\sigma^2/\tau^2} =\frac{n}{n+K}.

Justification: In the normal-normal model, the likelihood contributes $n/\sigma^2$ units of precision and the prior contributes $1/\tau^2$ units. The posterior mean weights each source in proportion to its precision. The resulting data weight is $n\tau^2/(n\tau^2+\sigma^2)$, which is larger when there are more observations or when the prior variance is larger, and smaller when individual observations are noisier.

Worked Example: Normal-Normal Updating and the Buhlmann Factor

A hypothetical insurer models a policy's annual claim cost $\Theta$ as $N(1000,400)$, where the second parameter is the variance. Given $\Theta=\theta$, each annual observation is independent and distributed as $N(\theta,900)$. Four observed annual costs are $1,000, $1,100, $1,200, and $1,200. Find the posterior mean and variance, and identify the credibility factor.

Method and assumptions. This is an equal-exposure normal-normal model with known prior and conditional variances. The prior mean is $\mu=1000$, the prior variance is $\tau^2=400$, the process variance is $\sigma^2=900$, and $n=4$.

First calculate the experience average:

\bar x=\frac{1000+1100+1200+1200}{4} =\frac{4500}{4}=1125.

The variance ratio is $K=900/400=2.25$. Therefore the factor is $Z_4=4/(4+2.25)=4/6.25=0.64$. The posterior mean is the prior-experience weighted average:

E[\Theta\mid x] =(1-0.64)(1000)+0.64(1125) =360+720=1080.

The posterior variance is the reciprocal of the sum of prior and data precisions:

\operatorname{Var}(\Theta\mid x) =\left(\frac{1}{400}+\frac{4}{900}\right)^{-1} =\left(\frac{9+16}{3600}\right)^{-1} =\frac{3600}{25}=144.

The posterior mean is $1,080 and the posterior variance is 144 dollars squared. As a check, the data average exceeds the prior mean, so a posterior mean of $1,080 lies between $1,000 and $1,125. The calculation avoids the common mistake of using the prior variance as the process variance in $K$; doing so would reverse the intended comparison of prior heterogeneity and observation noise.

Why the Same Factor Applies in the Buhlmann Model

The normal model gives an exact Bayesian derivation. The same factor also arises from a broader linear prediction argument that uses only the Buhlmann model's moment assumptions. In that setting, the observations need not be normally distributed. Assume that the $X_j$ are conditionally independent and identically distributed given $\Theta$, and that the relevant means and variances exist.

The law of total variance gives $\operatorname{Var}(X_j)=\operatorname{VHM}+\operatorname{EPV}$. Conditional independence gives $\operatorname{Cov}(X_j,X_k)=\operatorname{VHM}$ for distinct $j$ and $k$: once the latent type is fixed, the observations have no remaining covariance. Consequently, $\operatorname{Cov}(\bar X,\mu(\Theta))=\operatorname{VHM}$ and $\operatorname{Var}(\bar X)=\operatorname{VHM}+\operatorname{EPV}/n$.

Among predictors of the form $c+z\bar X$, the least-squares choice is the linear regression of $\mu(\Theta)$ on $\bar X$. Its slope is the covariance divided by the variance. Its intercept ensures that the predictor has mean $m$. These facts produce the Buhlmann premium:

z=\frac{\operatorname{Cov}(\mu(\Theta),\bar X)} {\operatorname{Var}(\bar X)} =\frac{\operatorname{VHM}}{\operatorname{VHM}+\operatorname{EPV}/n} =\frac{n}{n+\operatorname{EPV}/\operatorname{VHM}}, \qquad c=(1-z)m.

This derivation explains what the general result does and does not claim. The Buhlmann premium is the best linear predictor under squared-error loss. Without a distributional assumption that makes the posterior mean linear in the observations, it is not necessarily the full Bayesian posterior expectation.

Worked Example: A Buhlmann Premium Need Not Be the Posterior Mean

In a hypothetical book of risks, each latent type has equal prior probability of having mean cost 500 or 900. Given its type, each of three independent annual costs is normal with that type's mean and variance 40,000. For one risk, the observed costs are 700, 800, and 900. Calculate the Buhlmann linear premium and compare it with the exact posterior mean.

Method and assumptions. The latent type takes values $\Theta=500$ and $\Theta=900$, each with probability $1/2$. Conditional on its type, observations are independent $N(\Theta,40000)$. The grand mean is $m=(500+900)/2=700$. The expected process variance is $40{,}000$, while the variance of hypothetical means is

\operatorname{VHM} =\frac{(500-700)^2+(900-700)^2}{2} =\frac{40{,}000+40{,}000}{2} =40{,}000.

Thus $K=40{,}000/40{,}000=1$. The observed average is $\bar x=(700+800+900)/3=800$, so $Z_3=3/(3+1)=0.75$. The Buhlmann premium is

$$ 0.75(800)+0.25(700)=600+175=775. $$

To obtain the exact posterior, compare the likelihoods of the two types. The sums of squared deviations are $(200)^2+100^2+0^2=50{,}000$ for type 900 and $(200)^2+300^2+400^2=290{,}000$ for type 500. With equal prior probabilities, their posterior odds are therefore

\frac{P(\Theta=900\mid x)}{P(\Theta=500\mid x)} =\exp\left(\frac{290{,}000-50{,}000}{2(40{,}000)}\right) =e^3.

Hence $P(\Theta=900\mid x)=e^3/(1+e^3)$, and the exact posterior mean is

500+400\frac{e^3}{1+e^3} \approx 881.03.

The Buhlmann premium is $775, while the exact posterior mean is $881.03. The result is not a contradiction: the discrete prior does not make the posterior mean a linear function of the sample average. The Buhlmann value is the best linear predictor based on the model's variance components, not the exact posterior mean. This example avoids the mistake of treating every Buhlmann factor as proof that Bayesian updating must produce a linear posterior mean.

Worked Example: Applying Portfolio Variance Components

A hypothetical portfolio has estimated expected process variance of 18,000 dollars squared and variance of hypothetical means of 12,000 dollars squared. For a risk with four equal-exposure observations averaging $72,000, the portfolio grand mean is $60,000. Calculate $K$, the credibility factor, and the Buhlmann premium.

Method and assumptions. Treat the two variance components as applicable to this risk class, assume four conditionally independent and identically distributed observations, and use the Buhlmann linear premium.

The ratio is $K=18{,}000/12{,}000=1.5$. The factor is

Z_4=\frac{4}{4+1.5}=\frac{4}{5.5}=\frac{8}{11} \approx 0.7273.

The credibility premium is

\frac{8}{11}(72{,}000)+\frac{3}{11}(60{,}000) =\frac{576{,}000+180{,}000}{11} =\$68{,}727.27.

The check is that $68,727.27 lies between the experience average of $72,000 and the grand mean of $60,000. Because the experience weight is $8/11$, the premium is closer to $72,000. The calculation avoids rounding the factor too early; retaining $8/11$ gives the result to the nearest cent.

Interpreting the Factor and Avoiding Traps

The ratio $K=\operatorname{EPV}/\operatorname{VHM}$ is sometimes called the credibility parameter. It compares noise within a risk's experience with variation in underlying means across risks. A large EPV means observations are noisy and should receive less weight. A large VHM means risks differ more in their underlying means, so an individual risk's own experience is more informative relative to the class mean.

For positive $K$, the factor $Z_n=n/(n+K)$ is between zero and one, increases with $n$, and approaches one as $n$ grows. If VHM is zero, the model says all risks share the same hypothetical mean; the ratio $K$ is not finite, and individual experience does not improve prediction of that mean. If EPV is zero and VHM is positive, observations reveal the hypothetical mean without process noise, and the factor is one for $n>0$. The formula should not be used to claim that the posterior distribution is normal or that the posterior mean is exactly linear unless the model supports that conclusion.

Common exam trap: Reversing the ratio and using $K=\operatorname{VHM}/\operatorname{EPV}$ makes greater process noise increase credibility, which is backwards. For example, if EPV is 900, VHM is 400, and $n=3$, the correct $K$ is $900/400=2.25$ and $Z_3=3/(3+2.25)=4/7\approx0.5714$. The reversed ratio gives $K=4/9$ and the incorrect factor $3/(3+4/9)=27/31\approx0.8710$. Always place expected process variance in the numerator of $K$.

The Bayesian derivation and the linear prediction derivation answer complementary questions. The normal-normal model explains exactly how prior and likelihood precisions create a weighted posterior mean. The general Buhlmann result explains how the same weight emerges from conditional means and variances when the target is a best linear predictor. This distinction is aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus.

Check Your Understanding

For each problem, identify whether the requested premium is an exact normal-normal posterior mean or a Buhlmann linear premium, and retain full precision until the final answer.

A risk has prior $\Theta\sim N(1200,2500)$. Conditional on $\Theta$, five independent observations are $N(\Theta,625)$, and their average is $1,300. Find the credibility factor and posterior mean. (Answer: $Z=20/21\approx0.9524$; posterior mean $=\$1,295.24$.)
A class has EPV 900 and VHM 600. A risk has four equal-exposure observations averaging $62,000, and the grand mean is $50,000. Find the Buhlmann factor and premium. (Answer: $K=1.5$, $Z=8/11\approx0.7273$, premium $=\$58,727.27$.)
For a Buhlmann model with $K=2.5$, find the least integer $n$ for which $Z_n\geq0.8$. Verify the factor at that $n$. (Answer: $n=10$, for which $Z_{10}=10/12.5=0.8$.)
A risk has prior $\Theta\sim N(40,16)$, and conditional observations are independent $N(\Theta,64)$. Four observations have average 50. Find the posterior mean and posterior variance. (Answer: $Z=1/2$, posterior mean $=45$, posterior variance $=8$.)
A class has EPV 2,400 and VHM 800. One risk has six equal-exposure observations averaging $9,000, while the grand mean is $7,500. Find the Buhlmann premium and state whether it is necessarily an exact posterior mean. (Answer: $K=3$, $Z=2/3$, premium $=\$8,500$; it is not necessarily the exact posterior mean.)

What You'll Learn