Choosing the Complement in Bühlmann Credibility

Why the Complement Matters

A Bühlmann estimate does not use a class’s observed experience alone. It blends that experience with an overall mean, called the complement. The complement is the estimate’s benchmark: it pulls the result toward what is expected for the relevant collective when the class’s own data receive less weight.

That benchmark must represent the population to which the class belongs and the quantity being estimated. A mean drawn from a different mix of risks, a different measurement basis, or a different target population can systematically pull the estimate in the wrong direction. The credibility factor determines how strongly that mismatch affects the result.

Definition: In the ordinary Bühlmann model, let $\Theta$ denote a risk parameter and let $X_1,\ldots,X_n$ be a class’s observations. Conditional on $\Theta$, the observations are independent and identically distributed, with finite moments. Define the class’s conditional mean as $m(\Theta)=E[X_j\mid\Theta]$. The collective mean, or complement, is $\mu=E[m(\Theta)]=E[X_j]$, where the expectation is over the collective distribution of risks. It is not, in general, the class’s own observed mean.

Write the expected process variance as $\text{EPV}=E[\operatorname{Var}(X_j\mid\Theta)]$ and the variance of hypothetical means as $\text{VHM}=\operatorname{Var}(m(\Theta))$. When $\text{VHM}>0$, define $K=\text{EPV}/\text{VHM}$. Given $K$, the ordinary Bühlmann credibility factor for $n$ equally weighted observations is $Z=n/(n+K)$, and the credibility estimate of the class mean is

\widehat{m}=Z\bar{X}+(1-Z)\mu, \qquad \bar{X}=\frac{1}{n}\sum_{j=1}^{n}X_j.

Here $Z$ is between zero and one. If experience is relatively informative, $Z$ is larger and the estimate is closer to $\bar{X}$. If experience is relatively limited, the complement receives more weight. The usual formula assumes the stated equal-weight, conditionally independent model; it should not be applied unchanged when observations have unequal exposures or do not measure the same quantity.

How to Select a Representative Complement

The complement should be the expected mean for the target collective: the population of risks whose expected experience is an appropriate benchmark for the class being rated. Selection therefore requires more than finding a convenient average. First identify exactly what $X$ measures, then identify the target risks and the basis on which their experience is summarized. A mean of claim counts, for example, is not a complement for mean claim severity; a mean for one coverage or time period is not automatically suitable for another.

For a class drawn from a portfolio with differing risk groups, the collective mean reflects the target portfolio’s mix. If group $g$ has expected value $\mu_g$ and represents proportion $w_g$ of the target exposure, where $\sum_g w_g=1$, then the exposure-weighted complement is

\mu=\sum_g w_g\mu_g.

The weights should describe the target mix relevant to the estimate, not an unrelated mix that happens to be readily available. If the desired estimate concerns a particular class within that collective, the complement remains the collective benchmark; the class’s observed experience enters through $\bar{X}$ and its credibility weight. Substituting a class-specific observed mean as the complement would change the intended blend and can effectively use the class experience twice.

Specify the measured quantity.
State what each observation represents, including its unit and coverage or time basis.

Define the target collective.
Identify the population whose expected class means should provide the benchmark.

Match the mix and basis.
Use expected values and weights appropriate to the target risks and exposure mix.

Apply the credibility blend.
Combine the class mean with that complement using the model’s $Z$, and check the direction and size of the adjustment.

Why a Complement Error Moves the Estimate

The role of the complement follows directly from the linear projection underlying Bühlmann credibility. Let $M=m(\Theta)$. Under the model, the conditional expectation of the sample mean is $E[\bar{X}\mid\Theta]=M$, and its conditional variance is $\operatorname{Var}(\bar{X}\mid\Theta)=\operatorname{Var}(X_j\mid\Theta)/n$. Taking expectations and applying the variance decomposition gives

\operatorname{Cov}(M,\bar{X})=\operatorname{Var}(M)=\text{VHM}, \qquad \operatorname{Var}(\bar{X})=\text{VHM}+\frac{\text{EPV}}{n}.

The slope of the best linear predictor of $M$ from $\bar{X}$ is the covariance divided by the variance. Thus

\frac{\operatorname{Cov}(M,\bar{X})}{\operatorname{Var}(\bar{X})} = \frac{\text{VHM}}{\text{VHM}+\text{EPV}/n} = \frac{n}{n+\text{EPV}/\text{VHM}} = \frac{n}{n+K} = Z.

Both $M$ and $\bar{X}$ have collective mean $\mu$, so the best linear predictor has intercept $(1-Z)\mu$, giving $Z\bar{X}+(1-Z)\mu$. This derivation explains why the complement is the collective mean under the model: it centers the linear blend on the collective being modeled. The Bühlmann estimate is a best linear estimate under these assumptions; it need not equal the exact conditional expectation for every possible distribution.

Suppose the correct complement is $\mu$, but an analyst uses $\mu^\ast$ while keeping the same observed mean and credibility factor. The resulting estimate changes by

\widehat{m}^{\,\ast}-\widehat{m} = \bigl[Z\bar{X}+(1-Z)\mu^\ast\bigr] - \bigl[Z\bar{X}+(1-Z)\mu\bigr] = (1-Z)(\mu^\ast-\mu).

The sign of the change follows the complement error: a complement that is too high pulls the estimate upward, and one that is too low pulls it downward. The size is scaled by $1-Z$. This is a direct sensitivity calculation, not a claim that a poor complement changes the model’s underlying credibility factor.

Worked Example: Correcting a Mismatched Portfolio Benchmark

A hypothetical insurance class has an observed annual loss cost of $\$128$ per exposure from four observations. The supplied Bühlmann parameter is $K=6$. A target collective mean of $\$100$ per exposure is appropriate for this class. An analyst instead proposes a $\$120$ benchmark taken from a different risk mix. Calculate the estimate with the appropriate complement, then quantify the effect of using the mismatched one.

Method. Use the ordinary Bühlmann blend with $n=4$, the supplied $K$, and the target collective mean. The observations are being treated as equally weighted measurements on the same basis.

First calculate credibility:

Z=\frac{n}{n+K}=\frac{4}{4+6}=\frac{4}{10}=0.4, \qquad 1-Z=0.6.

Using the representative complement, the estimate is

\widehat{m} =0.4(128)+0.6(100) =51.2+60 =\$111.20.

Using the analyst’s $\$120$ benchmark instead gives

\widehat{m}^{\,\ast} =0.4(128)+0.6(120) =51.2+72 =\$123.20.

The mistaken benchmark raises the estimate by $\$12.00$. Checking with the sensitivity formula gives $0.6(120-100)=\$12.00$, the same result.

Common mistake avoided. It is wrong to treat any available portfolio average as an interchangeable complement. Here, the $\$120$ benchmark comes from a different risk mix, so its full mismatch is not justified by the target collective. The class mean and $Z$ were unchanged; the shift arose entirely from the complement error.

Worked Example: Weighting Subgroups to Match the Target Mix

A hypothetical property portfolio has two construction groups. Their expected annual loss costs are $\$80$ and $\$140$ per exposure. The class being estimated has a target mix of 60% and 40%, respectively. Its observed mean is $\$118$ from three equally weighted observations, and $K=3$. A broader portfolio mean based on a different mix is $\$95$. Find the appropriate estimate and the result of using that broader mean.

Method. First calculate the complement for the target mix, then apply the ordinary Bühlmann formula.

The target complement is the exposure-weighted subgroup mean:

\mu=0.60(80)+0.40(140)=48+56=\$104.

The credibility factor is

Z=\frac{3}{3+3}=\frac{1}{2}=0.5, \qquad 1-Z=0.5.

Therefore, the estimate using the target-mix complement is

\widehat{m}=0.5(118)+0.5(104)=59+52=\$111.

If the broader portfolio mean of $\$95$ is used instead, the estimate becomes

\widehat{m}^{\,\ast}=0.5(118)+0.5(95)=59+47.5=\$106.50.

The difference is $-\$4.50$, which agrees with $(1-Z)(95-104)=0.5(-9)=-\$4.50$. The fact that $\$95$ is an overall portfolio mean does not make it the appropriate complement: its mix differs from the target class’s mix.

Common mistake avoided. Simply averaging the subgroup means would give $(80+140)/2=\$110$, which assumes equal subgroup weights. The specified 60%/40% target mix instead gives $\$104$.

Worked Example: A Low-Credibility Class Pulled Toward the Wrong Mean

A hypothetical specialty class has an observed loss cost of $\$90$ per exposure. Its experience has $Z=0.25$, and the appropriate collective complement is $\$120$. A proposed substitute complement of $\$80$ comes from a lower-cost population. Calculate both estimates and explain the direction of the error.

Method. Use the stated credibility directly; no recalculation of $K$ is needed. The complement receives weight $1-Z=0.75$.

The estimate with the appropriate complement is

\widehat{m}=0.25(90)+0.75(120)=22.50+90=\$112.50.

With the lower, unrepresentative complement, the estimate is

\widehat{m}^{\,\ast}=0.25(90)+0.75(80)=22.50+60=\$82.50.

The estimate falls by $\$30.00$. The check is $(1-Z)(80-120)=0.75(-40)=-\$30.00$. Because the class has only 25% credibility, the complement has 75% weight, so the mismatch dominates the blend.

Common mistake avoided. Do not assume that a low observed class mean validates a low complement. The complement describes the target collective, not a second reading of the class’s own experience.

Practical Checks and Exam Traps

Before calculating, verify that the proposed complement and class mean refer to the same outcome, units, coverage, and time basis. Then ask whether the complement represents the target risk mix. If a portfolio mean combines groups with different expected values, use weights that match the population being estimated rather than defaulting to equal weights or borrowing a mean from a convenient but different book.

Confusing the complement with the class mean. The class mean is $\bar{X}$; the complement is the collective mean $\mu$.
Using a mismatched overall mean. A mean can be correctly calculated for one portfolio and still be unsuitable for a different target mix.
Reversing the weights. $Z$ multiplies the class experience, while $1-Z$ multiplies the complement.
Ignoring the direction of the shift. The estimate moves toward whichever complement is used; the change is $(1-Z)(\mu^\ast-\mu)$.
Changing $Z$ to compensate for a bad benchmark. A complement mismatch is not fixed by altering credibility while holding the model and experience constant.

The choice of complement matters most when $1-Z$ is large, because a larger share of the estimate comes from the collective benchmark. This does not make a high-credibility estimate immune to a bad complement; it makes the direct effect smaller. The appropriate benchmark remains a modeling choice grounded in the target population, not a tuning device selected to produce a desired estimate. This treatment of the collective complement is aligned with the CAS Exam 5 (Basic Ratemaking and Estimated Claim Liabilities) syllabus.

Key takeaway: Choose $\mu$ to represent the expected mean for the target collective and the same measurement basis as the class experience. If a substitute complement $\mu^\ast$ is used, the estimate shifts by exactly $(1-Z)(\mu^\ast-\mu)$ under the ordinary Bühlmann blend.

Check Your Understanding

For each problem, use the ordinary Bühlmann blend and the complement appropriate to the stated target population.

A hypothetical class has $\bar{X}=\$135$, $n=3$, and $K=6$. The appropriate collective mean is $\$105$. Calculate $Z$ and the credibility estimate. (Answer: $Z=1/3$; estimate $=\$115.00$.)
A target portfolio contains 25% of a subgroup with expected cost $\$60$ and 75% of a subgroup with expected cost $\$100$. The class has $\bar{X}=\$90$, $Z=0.4$, and an analyst mistakenly uses $\$90$ as the complement. Calculate the target-mix complement and both estimates. (Answer: target complement $=\$90$; both estimates $=\$90.00$, so the complement substitution causes no change.)
A hypothetical class has $\bar{X}=\$160$, $n=5$, and $K=5$. Its target collective mean is $\$120$, but a mismatched benchmark of $\$140$ is used. Calculate the correct estimate, the mistaken estimate, and the shift. (Answer: correct estimate $=\$140.00$; mistaken estimate $=\$150.00$; shift $=+\$10.00$.)
A class’s correct complement is $\$100$, but an alternative complement of $\$130$ is proposed. If $Z=0.7$, calculate the change in the estimate due solely to this substitution. (Answer: $0.3(130-100)=+\$9.00$.)
A target population has two subgroups with expected values $\$70$ and $\$150$, and the target exposure mix is 40% and 60%. A class has $\bar{X}=\$125$, $n=4$, and $K=4$. Find the target complement and credibility estimate. (Answer: complement $=\$118$; $Z=0.5$; estimate $=\$121.50$.)

What You'll Learn