Deriving the Bayesian Connection to Bühlmann Credibility

Why Bayesian Posterior Means Can Look Like Credibility Formulas

A credibility estimate balances a risk’s own experience against a collective value. In a Bayesian model, the corresponding balance is between the observed data and the prior distribution. These descriptions are not merely similar in some conjugate models: the posterior mean is exactly a credibility-weighted average, with the weight determined by the prior and sampling model.

The connection has assumptions. The observations must follow the stated conditional model, given a latent risk parameter; the prior must be the specified conjugate distribution; and the quantity being predicted must be the conditional mean or a quantity whose posterior expectation is that mean. If these conditions do not hold, a posterior mean need not have the same form as the Bühlmann estimate.

Definition: Let \(\Theta\) represent a latent risk characteristic, and let \(X_1,\ldots,X_n\) be observations conditional on \(\Theta\). The Bayesian posterior mean of the risk mean is \(E[m(\Theta)\mid X_1,\ldots,X_n]\), where \(m(\Theta)=E[X_j\mid\Theta]\). In the equal-exposure Bühlmann model, the credibility estimate is \(Z\bar X+(1-Z)\mu\), where \(\bar X\) is the individual’s sample mean and \(\mu=E[m(\Theta)]\).

For the classical Bühlmann model, conditional on \(\Theta\), the observations are identically distributed and independent, with conditional mean \(m(\Theta)\) and conditional variance \(v(\Theta)\). Define the expected process variance (EPV) and variance of hypothetical means (VHM) by

\text{EPV}=E[v(\Theta)],\qquad \text{VHM}=\operatorname{Var}(m(\Theta)),\qquad K=\frac{\text{EPV}}{\text{VHM}}.

When the quantities are finite and VHM is positive, the Bühlmann credibility factor for \(n\) equal-weight observations is \(Z=n/(n+K)\). If VHM is zero, there is no variation in the latent risk means and this ratio is not defined; that degenerate case does not provide a nontrivial credibility balance.

Poisson-Gamma: An Exact Bühlmann Form

Suppose the latent claim frequency per unit exposure is \(\Lambda\), with a Gamma prior parameterized by shape \(\alpha\) and rate \(\beta\). Thus \(E[\Lambda]=\alpha/\beta\) and \(\operatorname{Var}(\Lambda)=\alpha/\beta^2\). Given \(\Lambda\), claim counts in disjoint periods are independent Poisson variables: for exposure \(e_j\), \(N_j\mid\Lambda\sim\operatorname{Poisson}(e_j\Lambda)\). Exposures are known and measured in a common unit.

Write \(E=\sum_j e_j\) for total exposure and \(N=\sum_j N_j\) for the total observed count. The likelihood as a function of \(\Lambda\) is proportional to \(\Lambda^N e^{-E\Lambda}\). Multiplying by the Gamma prior density, proportional to \(\Lambda^{\alpha-1}e^{-\beta\Lambda}\), gives a posterior density proportional to

\Lambda^{\alpha+N-1}e^{-(\beta+E)\Lambda}.

This is a Gamma density with shape \(\alpha+N\) and rate \(\beta+E\). Therefore the posterior mean frequency is \((\alpha+N)/(\beta+E)\). The observed exposure-weighted frequency is \(N/E\), while the prior mean frequency is \(\alpha/\beta\). Rearranging gives

E[\Lambda\mid N_1,\ldots,N_j] =\frac{E}{E+\beta}\frac{N}{E} +\frac{\beta}{E+\beta}\frac{\alpha}{\beta}.

For a unit-exposure Poisson observation, the conditional mean is \(\Lambda\) and the conditional variance is also \(\Lambda\). Hence \(\text{EPV}=E[\Lambda]=\alpha/\beta\), \(\text{VHM}=\operatorname{Var}(\Lambda)=\alpha/\beta^2\), and \(K=\beta\). Thus the Bayesian weight \(E/(E+\beta)\) is precisely the Bühlmann-Straub weight \(E/(E+K)\) for total exposure \(E\). The weighted experience \(N/E\) is the exposure-weighted average of the period frequencies \(N_j/e_j\).

Worked Example: Poisson-Gamma Frequency With Unequal Exposures

A hypothetical book’s claim count is modeled with a Gamma prior for the latent frequency, using shape \(\alpha=4\) and rate \(\beta=2\) per exposure unit. One risk has 1 exposure unit with 1 claim and 3 exposure units with 5 claims. Find its posterior mean frequency and verify the Bühlmann-Straub representation. Interest and mortality assumptions are not applicable. Conditional counts are independent Poisson variables with means equal to exposure times the latent frequency.

Method and prior quantities. The prior mean is \(\alpha/\beta=4/2=2\) claims per exposure unit. For a unit-exposure Poisson count, EPV is 2 and VHM is \(\alpha/\beta^2=4/4=1\). Therefore \(K=\text{EPV}/\text{VHM}=2\).

Aggregate the experience. Total exposure is \(E=1+3=4\), and total claims are \(N=1+5=6\). The exposure-weighted observed frequency is \(N/E=6/4=1.5\). The period rates are \(1/1=1\) and \(5/3\); their exposure-weighted mean is \((1\cdot1+3\cdot(5/3))/4=6/4=1.5\).

Calculate the credibility estimate. The weight is \(Z=E/(E+K)=4/(4+2)=2/3\). The estimate is

\frac{2}{3}(1.5)+\frac{1}{3}(2)=1+\frac{2}{3}=\frac{5}{3}.

Check using the posterior. The posterior is Gamma with shape \(4+6=10\) and rate \(2+4=6\), so its mean is \(10/6=5/3\). The two methods agree exactly. The common mistake avoided here is taking the unweighted average of the two period rates, which would give \((1+5/3)/2=4/3\) and would incorrectly give the one-unit and three-unit periods equal influence. No rounding was needed; the answer is \(1.67\) claims per exposure unit to two decimals.

The Same Structure in Other Conjugate Models

The connection is not limited to counts. In each of the models below, the posterior mean is a weighted average of an individual statistic and the prior mean. The exact weight depends on the sampling distribution and prior parameters.

Worked Example: Beta-Binomial Claim Incidence

For a hypothetical risk, let the probability \(\Theta\) of a claim in a trial have prior distribution \(\operatorname{Beta}(2,6)\). Conditional on \(\Theta\), claim indicators are independent Bernoulli variables with probability \(\Theta\). Among 10 trials, there are 3 claims. Find the posterior mean claim probability and its credibility representation. Interest and mortality assumptions are not applicable.

Method and prior mean. The prior mean is \(2/(2+6)=1/4=0.25\). The observed claim proportion is \(3/10=0.30\). The Beta-Binomial update gives posterior parameters \(2+3=5\) and \(6+(10-3)=13\), so the posterior is \(\operatorname{Beta}(5,13)\).

Posterior mean and credibility weight. The posterior mean is \(5/(5+13)=5/18\). In general, with prior \(\operatorname{Beta}(a,b)\), \(s\) claims in \(n\) trials gives posterior mean \((a+s)/(a+b+n)\). This can be written as

\frac{n}{n+a+b}\frac{s}{n} +\frac{a+b}{n+a+b}\frac{a}{a+b}.

Here \(Z=10/(10+8)=5/9\), and \(1-Z=4/9\). The weighted estimate is \((5/9)(0.30)+(4/9)(0.25)=1/6+1/9=5/18=0.2778\). As a check on the Bühlmann constants for one Bernoulli trial, EPV is \(E[\Theta(1-\Theta)]=ab/[(a+b)(a+b+1)]=12/(8\cdot9)=1/6\), and VHM is \(\operatorname{Var}(\Theta)=ab/[(a+b)^2(a+b+1)]=12/(64\cdot9)=1/48\). Their ratio is \(K=8\), giving \(Z=n/(n+K)=10/18\). The trap avoided is using the observed proportion alone as the estimate; doing so ignores the prior information. The answer, rounded to four decimals, is 0.2778.

Worked Example: Normal-Normal Mean Estimation

A hypothetical risk has latent mean \(\Theta\sim N(100,400)\), where the second parameter is the variance. Given \(\Theta\), six independent observations have distribution \(N(\Theta,900)\), again with variance 900, and their sample mean is 115. Find the posterior mean of \(\Theta\) and identify its Bühlmann weight. Interest and mortality assumptions are not applicable.

Method and assumptions. The conditional observation variance is \(\sigma^2=900\), and the prior variance is \(\tau^2=400\). With a normal prior and conditionally normal observations of known variance, the posterior mean is

\frac{n\tau^2}{n\tau^2+\sigma^2}\bar X+ \frac{\sigma^2}{n\tau^2+\sigma^2}\mu_0.

Calculate the weight. EPV is 900 and VHM is 400, so \(K=900/400=2.25\). Equivalently, \(Z=n/(n+K)=6/(6+2.25)=6/8.25=8/11\). Thus \(1-Z=3/11\), and the posterior mean is

\frac{8}{11}(115)+\frac{3}{11}(100) =\frac{920+300}{11} =\frac{1220}{11} =110.9091.

Check. The posterior precision is \(1/400+6/900=1/400+1/150=11/1200\). The precision-weighted numerator is \(100/400+6(115)/900=1/4+23/30=61/60\); dividing gives \((61/60)(1200/11)=1220/11\), as above. The mistake to avoid is treating the prior variance as the credibility constant: the weight uses the ratio \(\sigma^2/\tau^2=2.25\). The posterior mean is 110.91 to two decimals.

What the Correspondence Does—and Does Not—Establish

The examples illustrate a useful derivation strategy: obtain the posterior mean from the conjugate update, express it as a weighted average, and then compare the weight with \(n/(n+K)\) or its exposure-weighted counterpart. Matching the coefficients identifies the implied credibility constant. In the Poisson-Gamma model, for example, the prior rate parameter \(\beta\) equals the Bühlmann-Straub constant \(K\); this follows from EPV/VHM, not from treating \(\beta\) as a sample size.

The match is exact for the posterior mean under the stated models. Bühlmann credibility, in general, is a linear prediction based on the first two conditional moments. A general Bayesian posterior mean can be nonlinear in the data or depend on features not summarized by the sample mean. Therefore, do not conclude that every Bayesian model has an exact Bühlmann form merely because both methods combine experience with a collective value.

A reliable exam approach is to check the target being estimated, the exposure or trial weights, and the parameterization of the prior before calculating. In particular, a Gamma distribution may be parameterized by a rate or a scale; the formulas here use the rate \(\beta\). Under this convention, its mean is \(\alpha/\beta\), and the Poisson-Gamma credibility constant is \(\beta\).

Specify the hierarchy.
State the prior for the latent parameter and the conditional distribution of observations, including independence and exposure assumptions.

Find the posterior mean.
Update the conjugate parameters and identify the quantity whose posterior expectation is required.

Rewrite as a weighted average.
Separate the observed individual statistic from the prior mean, and verify that the coefficients sum to one.

Check the Bühlmann constants.
Calculate EPV and VHM from the conditional model and prior, then verify that the resulting credibility weight matches.

Key takeaway: A conjugate Bayesian posterior mean is an exact Bühlmann credibility estimate when its algebra yields the same weighted-average form and the model’s EPV/VHM ratio gives the same credibility constant. The equality depends on the model assumptions and applies to the specified posterior mean, not automatically to every Bayesian predictor.

Check Your Understanding

For each problem, identify the model-based weight and verify the estimate against the conjugate posterior mean.

A hypothetical Poisson-Gamma risk has prior \(\operatorname{Gamma}(3,1.5)\), using shape-rate notation, and records 4 claims over 2.5 exposure units. Find the posterior mean frequency and credibility weight. (Answer: posterior mean \(7/4=1.75\); \(Z=0.625\).)
A hypothetical claim-incidence probability has prior \(\operatorname{Beta}(3,9)\). There are 5 claims in 12 independent trials. Find the posterior mean and the weight on the observed proportion. (Answer: posterior mean \(8/24=1/3=0.3333\); \(Z=12/24=0.5\).)
Suppose \(\Theta\sim N(50,100)\), and 8 conditionally independent observations have distribution \(N(\Theta,400)\) with observed mean 60. Find the posterior mean and the Bühlmann constant \(K\). (Answer: \(K=4\); posterior mean \(56.6667\).)
A hypothetical Poisson-Gamma model has prior \(\operatorname{Gamma}(2,1)\). A risk has 0 claims over 1 exposure unit and 6 claims over 3 exposure units. Find the exposure-weighted observed rate, \(Z\), and posterior mean frequency. (Answer: observed rate \(6/4=1.5\); \(Z=4/(4+1)=0.8\); posterior mean \(1.6\).)
For a unit-exposure Poisson-Gamma model with prior \(\operatorname{Gamma}(5,2)\), calculate EPV, VHM, and \(K\). If a risk has 12 claims over 3 exposure units, find its credibility estimate and confirm it from the posterior. (Answer: EPV \(=2.5\), VHM \(=1.25\), \(K=2\); \(Z=3/5\); estimate and posterior mean \(3.4\).)

What You'll Learn