Bühlmann and Limited-Fluctuation Weights

Two Credibility Weights, Two Objectives

Bühlmann credibility and limited-fluctuation credibility both produce a weight between zero and one, but the weights answer different questions. The Bühlmann factor balances an individual risk’s observations against a collective mean to minimize mean squared error under a risk model. Limited fluctuation asks how much exposure is needed for observed experience to fall within a specified accuracy band with a specified probability.

That difference matters in practice. A weight from one method is not automatically a substitute for a weight from the other, even when both are applied to the same experience. Bühlmann’s factor depends on risk-model variance components; a limited-fluctuation standard depends on the chosen tolerance, confidence level, and process distribution.

Definition: In the Bühlmann model, let $\Theta$ denote a risk’s latent risk parameter, $X_1,\ldots,X_n$ its conditionally independent observations, and $m=E[\mu(\Theta)]$ the collective mean, where $\mu(\Theta)=E[X_j\mid\Theta]$. The expected process variance is $\text{EPV}=E[\operatorname{Var}(X_j\mid\Theta)]$, and the variance of hypothetical means is $\text{VHM}=\operatorname{Var}(\mu(\Theta))$. When these quantities are finite and $\text{VHM}>0$, the Bühlmann factor for the average $\bar X$ is $Z_B=n/(n+K)$, where $K=\text{EPV}/\text{VHM}$. The credibility-weighted estimate is $Z_B\bar X+(1-Z_B)m$.

The factor gives more weight to experience as $n$ increases, and less weight when process variation is large relative to variation among risk means. The ratio $K$ is often called the Bühlmann parameter. If $\text{VHM}=0$, the model has no variation in risk means, so this formula for $K$ is undefined; the situation does not call for the usual Bühlmann weighting calculation.

Why the Bühlmann Weight Has This Form

The result follows by choosing the weight that minimizes the unconditional mean squared error of a linear credibility estimate for the risk’s own mean. Assume the conditional independence and common conditional moments stated above, and treat the model quantities as known. For a proposed weight $Z$, write the estimate as $P=Z\bar X+(1-Z)m$. Its error separates into a process-fluctuation term and a risk-mean term:

P-\mu(\Theta)=Z\bigl(\bar X-\mu(\Theta)\bigr)+(1-Z)\bigl(m-\mu(\Theta)\bigr).

Conditional on $\Theta$, the first term has mean zero, so its cross-product with the second term has expectation zero. Conditional independence gives $E[\operatorname{Var}(\bar X\mid\Theta)]=\text{EPV}/n$. Therefore,

E[(P-\mu(\Theta))^2]=Z^2\frac{\text{EPV}}{n}+(1-Z)^2\text{VHM}.

Differentiating with respect to $Z$, setting the derivative to zero, and solving gives

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

The second derivative is $2\text{EPV}/n+2\text{VHM}$, which is positive when the variance components are nonnegative and at least one is positive. Thus the stationary point is a minimum. This derivation explains the objective behind the Bühlmann weight; it does not establish a probability guarantee that the estimate will lie within a chosen tolerance.

Limited-Fluctuation Credibility

A limited-fluctuation standard starts with an accuracy target. For $n$ independent observations with mean $\mu>0$ and variance $\sigma^2$, consider the requirement that the sample mean be within a relative tolerance $k$ of its target:

P\left(\left|\bar X-\mu\right|\leq k\mu\right)\geq p,

where $0<k<1$ is the allowed relative error and $p$ is the desired probability. If $\bar X$ is normal, or a normal approximation is appropriate, set $z=\Phi^{-1}((1+p)/2)$, the standard-normal quantile for a central probability $p$. Since the standard deviation of $\bar X$ is $\sigma/\sqrt n$, the approximate full-credibility requirement is

z\frac{\sigma}{\sqrt n}\leq k\mu \quad\Longrightarrow\quad n\geq n_0=\left(\frac{z\sigma}{k\mu}\right)^2.

Here $n_0$ is the full-credibility standard in the same units as $n$. A common partial-credibility convention assigns

Z_L=\min\left(1,\sqrt{\frac{n}{n_0}}\right).

The square-root rule reflects the fact that sampling standard deviation decreases in proportion to $1/\sqrt n$. It is a convention for scaling partial credibility from the full standard; by itself it is not an exact probability guarantee for every exposure below $n_0$. The normal approximation can also be poor for small samples, highly skewed observations, or distributions with unusual tails. The selected tolerance and probability are design choices, not quantities determined by the Bühlmann variance components.

Key distinction: $Z_B$ is derived from a mean-squared-error objective under a risk model. $Z_L$ is based on a probability-and-accuracy standard, usually with a normal approximation and a square-root partial-credibility convention. The two weights can differ without either calculation being incorrect.

Worked Comparisons

Worked Example: Different Objectives, Different Weights

Hypothetical scenario. A book of property risks is summarized by a model with $\text{EPV}=3{,}600$ and $\text{VHM}=1{,}200$, in squared dollars per observation. For a benchmark risk with expected claim cost $\mu=\$100$, the conditional claim-cost standard deviation is $\sigma=\$60$. The limited-fluctuation target is 90% probability within 10% of the benchmark mean. There are $n=25$ observations and their average is $\bar X=\$130$. Use $z=1.645$ for the central 90% normal approximation.

For specificity, suppose the risk mean $\Theta$ takes $100-\sqrt{2{,}400}$, $100$, and $100+\sqrt{2{,}400}$ dollars with probabilities $0.25$, $0.50$, and $0.25$, respectively. Its mean is $\$100$, and its variance is $0.5(2{,}400)=1{,}200$. Conditional on $\Theta=\theta$, let claim costs be independent gamma variables with shape $\theta^2/3{,}600$ and scale $3{,}600/\theta$. Their conditional mean is $\theta$ and variance is $3{,}600$. Thus the stated EPV and VHM agree with the model. The limited-fluctuation calculation below uses the benchmark risk with mean $\$100$ and process standard deviation $\$60$, and a normal approximation for its average. Interest and mortality assumptions do not apply.

Bühlmann method. First calculate $K=3{,}600/1{,}200=3$. Then

Z_B=\frac{25}{25+3}=\frac{25}{28}=0.8929.

The weighted estimate is $0.892857(\$130)+0.107143(\$100)=\$116.0714+\$10.7143=\$126.7857$, or \$126.79 to the nearest cent.

Limited-fluctuation method. The allowed absolute deviation is $0.10(\$100)=\$10$. The full standard is

n_0=\left(\frac{1.645(\$60)}{\$10}\right)^2 =(9.87)^2=97.4169.

Since $25<97.4169$, the partial weight is $Z_L=\sqrt{25/97.4169}=\sqrt{0.25663}=0.5066$. If this weight is used to blend the same average with the same collective mean, the result is $0.5066(\$130)+0.4934(\$100)=\$115.198$, or \$115.20 to the nearest cent.

The methods give different weights because the Bühlmann calculation accounts for both within-risk process variation and between-risk variation, while this limited-fluctuation target concerns process variation around the benchmark mean. Common mistake: treating $Z_L$ as if it were the MSE-optimal Bühlmann factor. The calculations above show that the same data can yield $0.8929$ and $0.5066$ under distinct objectives.

Worked Example: Matching Weights at a Particular Exposure

Hypothetical scenario. A model has $\text{EPV}=2{,}500$ and $\text{VHM}=625$, so $K=4$. For a benchmark risk with mean $\$100$, assume independent gamma observations with conditional mean $\$100$ and variance $2{,}500$, giving standard deviation $\$50$. A limited-fluctuation target is a central probability of $0.6827$ within 10% of the mean; use $z=1$. Find the full standard and compare weights at $n=16$. Interest and mortality do not apply.

The Bühlmann parameter is $K=2{,}500/625=4$. The full limited-fluctuation standard is

n_0=\left(\frac{1(\$50)}{0.10(\$100)}\right)^2 =\left(\frac{50}{10}\right)^2=25.

At $n=16$, $Z_B=16/(16+4)=16/20=0.8$. The limited-fluctuation weight is $Z_L=\sqrt{16/25}=4/5=0.8$. They happen to match. For example, if $\bar X=\$112$, either blended estimate is $0.8(\$112)+0.2(\$100)=\$109.60$.

This equality is specific to these inputs and this exposure; it does not make the formulas interchangeable. Common mistake: inferring from one matching pair that the weights will match at other exposures. The two rules have different functional forms and different inputs.

Worked Example: Converting Frequency Exposure Correctly

Hypothetical scenario. A portfolio has a Poisson claim count with frequency $\lambda=0.05$ claims per policy-year over $800$ policy-years. The limited-fluctuation target is a central 90% probability that the count is within 20% of its expected value; use $z=1.645$. Find the expected count, full standard in expected claims, and partial-credibility weight. Interest and mortality do not apply.

The expected number of claims is $0.05(800)=40$. For a Poisson count with expected count $M$, the variance is $M$, so its standard deviation is $\sqrt M$. The normal-approximation requirement $z\sqrt M\leq0.20M$ gives the full standard

M_0=\left(\frac{1.645}{0.20}\right)^2 =(8.225)^2=67.650625\text{ expected claims}.

The partial weight is $Z_L=\sqrt{40/67.650625}=\sqrt{0.59127}=0.7689$. The frequency standard has been stated in expected claims, not policy-years. Common mistake: comparing 800 policy-years directly with 67.650625 claims. That mixes units and would incorrectly imply full credibility.

Choosing and Interpreting a Weight

Before calculating either weight, identify the quantity being weighted and keep its units consistent. For Bühlmann, $n$ counts observations of the unit used to define EPV and VHM. For limited fluctuation, $n$ or $M$ must match the unit in which the full standard was derived: observations, expected claims, or another explicitly defined exposure basis.

The methods also react differently to changes in assumptions. Increasing VHM relative to EPV lowers $K$, increasing Bühlmann credibility at a fixed volume. Tightening a limited-fluctuation tolerance $k$ increases $n_0$ in inverse proportion to $k^2$; demanding a higher probability generally increases $z$ and therefore $n_0$. Neither weight should be reported without its target and assumptions.

Use Bühlmann when the objective is an MSE-based blend under a risk model with defensible EPV and VHM.
Use limited fluctuation when a specified probability of meeting an accuracy tolerance is the relevant standard.
State whether a limited-fluctuation result is a full standard or a square-root partial weight.
Do not interpret a normal-approximation standard as an exact guarantee for arbitrary distributions.

The comparison is aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus because it requires distinguishing credibility objectives, identifying model inputs, and keeping exposure and probability assumptions explicit.

Check Your Understanding

For each hypothetical problem, calculate the requested weight or standard and check the units and stated approximation.

A risk model has EPV $=500$, VHM $=100$, and $n=12$ observations. Calculate the Bühlmann factor. (Answer: $K=5$, $Z_B=12/17=0.7059$.)
For a benchmark mean of $\$80$, process standard deviation $\$24$, and a 90% central normal-approximation target within 15%, use $z=1.645$. Calculate $n_0$ and $Z_L$ at $n=9$. (Answer: $n_0=(1.645(24)/12)^2=10.8241$; $Z_L=\sqrt{9/10.8241}=0.9117$.)
A Poisson frequency is $0.04$ claims per policy-year over $1{,}000$ policy-years. For a 95% central target within 25%, use $z=1.96$. Calculate the expected count, full standard in expected claims, and partial weight. (Answer: expected count $40$; $M_0=(1.96/0.25)^2=61.4656$ expected claims; $Z_L=\sqrt{40/61.4656}=0.8067$.)
A model has EPV $=800$, VHM $=400$, and $n=18$. A separately specified limited-fluctuation standard is $n_0=36$. Calculate both weights. (Answer: $K=2$, $Z_B=18/20=0.9000$, and $Z_L=\sqrt{18/36}=0.7071$.)
A Bühlmann model has EPV $=1{,}600$ and VHM $=400$. A limited-fluctuation standard is $n_0=100$. At $n=25$, compare the weights and state whether either reaches full credibility. (Answer: $K=4$, $Z_B=25/29=0.8621$, $Z_L=\sqrt{25/100}=0.5000$; neither reaches full credibility.)

What You'll Learn