Model-Based and Limited-Fluctuation Credibility

Why Two Credibility Methods?

Credibility methods address how much weight to give observed experience, but they answer that question in different ways. Bühlmann credibility uses a model of variation within and between risks to blend an individual experience mean with a collective mean. Limited-fluctuation credibility starts with a probability standard: how much exposure is needed for an aggregate result to fall within a specified relative error of its expected value with a specified probability?

The methods can produce weights that look similar, but their interpretations differ. A Bühlmann factor is determined by variance components and the amount of experience. A limited-fluctuation threshold is determined by a tolerance and a probability target. Neither method automatically validates the other’s assumptions.

Definition: In the classical Bühlmann model, each risk has an unobserved risk parameter $\Theta$. Conditional on $\Theta$, observations $X_1,\ldots,X_n$ are independent and identically distributed, with conditional mean $\mu(\Theta)=E[X_j\mid\Theta]$ and conditional variance $v(\Theta)=\operatorname{Var}(X_j\mid\Theta)$. The expected process variance is $\mathrm{EPV}=E[v(\Theta)]$, the variance of hypothetical means is $\mathrm{VHM}=\operatorname{Var}(\mu(\Theta))$, and the collective mean is $m=E[\mu(\Theta)]$. These quantities are assumed finite, with $\mathrm{VHM}>0$.

Bühlmann Credibility: A Model-Based Blend

For $n$ equally weighted observations, let $\bar X$ be their arithmetic mean. The Bühlmann linear credibility estimate of the risk’s expected value is

\widehat{\mu}=Z\bar X+(1-Z)m,\qquad Z=\frac{n}{n+K},\qquad K=\frac{\mathrm{EPV}}{\mathrm{VHM}}.

The factor $Z$ increases with the number of observations and with the degree of variation between risks. It decreases as process variation increases relative to variation between risks. If $\mathrm{VHM}=0$, risks have the same expected value under the model, and there is no between-risk variation for individual experience to identify; the displayed $K$ formula is then not used as a finite ratio.

Derivation: Consider linear estimates of the form $a\bar X+(1-a)m$. The mean squared error, averaged over the risk parameter and observations, is $(1-a)^2\mathrm{VHM}+a^2\mathrm{EPV}/n$: the first term reflects shrinkage away from a risk’s true mean, and the second is the sampling variance of $\bar X$. Differentiating with respect to $a$ and setting the derivative to zero gives

-2(1-a)\mathrm{VHM}+\frac{2a\,\mathrm{EPV}}{n}=0, \qquad a=\frac{\mathrm{VHM}}{\mathrm{VHM}+\mathrm{EPV}/n} =\frac{n}{n+\mathrm{EPV}/\mathrm{VHM}}.

Thus $a=Z$. This derivation establishes the best linear credibility factor under the model’s assumptions. It does not mean that the resulting estimate satisfies a specified probability of being within a fixed percentage of the risk’s expected value.

Worked Example: Blending Six Periods of Experience

A hypothetical rating class has six equally exposed annual observations of loss cost: $900, 1{,}100, 1{,}300, 1{,}400, 1{,}500,$ and $1{,}600$ dollars. Suppose the applicable Bühlmann model has $\mathrm{EPV}=900$, $\mathrm{VHM}=300$, and collective mean $m=\$1{,}200$. Estimate the class mean.

Method. This is a classical Bühlmann estimate: first calculate the experience mean and the variance ratio $K$, then blend the experience with the collective mean.

The observations sum to $900+1{,}100+1{,}300+1{,}400+1{,}500+1{,}600=7{,}800$, so $\bar X=7{,}800/6=\$1{,}300$. The parameter is $K=900/300=3$, and hence $Z=6/(6+3)=2/3$.

\widehat{\mu} =\frac{2}{3}(1{,}300)+\frac{1}{3}(1{,}200) =866.67+400 =\$1{,}266.67.

The estimate lies between the observed mean and the collective mean, as a credibility blend should. The common mistake avoided here is to use $Z=\mathrm{VHM}/\mathrm{EPV}$ directly; that ratio is $1/3$, while the credibility factor for six observations is $n/(n+K)=2/3$.

Limited-Fluctuation Credibility: A Probability Standard

Let $Y_j$ be the aggregate loss for one independent, homogeneous exposure unit, with mean $\mu_Y>0$ and standard deviation $\sigma_Y$. For $N$ such units, total loss is $S_N=\sum_{j=1}^{N}Y_j$, with mean $N\mu_Y$ and variance $N\sigma_Y^2$. A full-credibility standard specifies a relative tolerance $k>0$ and a target probability $p$, and requires

P\left(\left|S_N-N\mu_Y\right|\leq kN\mu_Y\right)\geq p.

Using a normal approximation for the aggregate total, let $z_{(1+p)/2}$ be the standard normal quantile with central probability $p$. The standard requires the margin $z_{(1+p)/2}\sigma_Y\sqrt{N}$ to be no greater than the permitted deviation $kN\mu_Y$. Solving for $N$ gives the continuous full-credibility threshold

N_0= \left(\frac{z_{(1+p)/2}\sigma_Y}{k\mu_Y}\right)^2.

This approximation is most defensible when the aggregate total is sufficiently well approximated by a normal distribution. The calculation also presumes independent, comparable exposure units and a stable mean and variance per unit. If the standard is applied to an integer number of units, round $N_0$ upward. For an exposure amount $N<N_0$, the conventional limited-fluctuation partial-credibility factor is

Z_{\mathrm{LF}}=\min\left(1,\sqrt{\frac{N}{N_0}}\right).

The square-root relationship follows because the standard deviation of the total grows in proportion to $\sqrt{N}$, while its mean grows in proportion to $N$. This partial factor is a conventional extension of the full-credibility rule; it is not the Bühlmann factor.

Worked Example: Applying the Relative-Error Standard

Suppose a hypothetical exposure unit has aggregate loss $Y$ with a Gamma distribution using shape $0.25$ and scale $1$, so that $E[Y]=0.25$ and $\operatorname{Var}(Y)=0.25$. Use a normal approximation to find the exposure needed for a 90% central probability that total loss is within 20% of its expected value. Also find the partial factor for 100 exposure units.

Method. Use the limited-fluctuation standard with $\mu_Y=0.25$, $\sigma_Y=\sqrt{0.25}=0.5$, $k=0.20$, and $z_{0.95}=1.644854$. The shape-scale convention gives the stated mean and variance: shape times scale is $0.25$, and shape times scale squared is also $0.25$.

Substitute into the threshold formula. The permitted deviation per unit is $k\mu_Y=0.20(0.25)=0.05$, so

N_0 =\left(\frac{1.644854(0.5)}{0.20(0.25)}\right)^2 =(16.44854)^2 \approx 270.5545.

The minimum integer exposure is therefore 271 units. For 100 units, the partial factor is

Z_{\mathrm{LF}}=\sqrt{\frac{100}{270.5545}}\approx 0.6080.

The probability standard is met at approximately 271 units under the normal approximation; 100 units receive about 60.80% partial credibility under the conventional square-root rule.

Common mistake and correction. A tempting but wrong calculation is to omit the mean per unit and use $(1.644854(0.5)/0.20)^2=(4.112135)^2\approx16.9097$. The target deviation is not merely the relative tolerance $0.20$; it is $k$ times the expected aggregate loss per unit, $0.20(0.25)=0.05$. Using that denominator gives the correct threshold, approximately 270.5545, not 16.9097.

Frequency-Severity Form of the Standard

A common actuarial application models the number of claims per exposure unit as Poisson with mean $\lambda$, and assumes claim severities are independent and identically distributed, independent of frequency, with severity $X$. If $Y$ is aggregate loss for one unit, then

E[Y]=\lambda E[X],\qquad \operatorname{Var}(Y)=\lambda E[X^2].

The variance uses the severity second moment $E[X^2]$, not just the severity variance: random claim counts contribute additional variability. Substitution in the limited-fluctuation threshold gives

N_0= \left( \frac{z_{(1+p)/2}\sqrt{\lambda E[X^2]}} {k\lambda E[X]} \right)^2 = \frac{z_{(1+p)/2}^2E[X^2]} {k^2\lambda(E[X])^2}.

This form requires finite severity second moment, positive frequency and mean severity, and the stated Poisson and independence assumptions. It should not be applied unchanged when frequency is overdispersed, severities depend on frequency, or the exposure units are not comparable.

Worked Example: A Poisson Frequency and Exponential Severity Model

For a hypothetical coverage, assume claim frequency per exposure unit is Poisson with mean $\lambda=2$. Conditional claim amounts are independent exponential random variables with mean $\$1{,}000$, independent of frequency. Find the full-credibility exposure for a 95% central probability and a 10% relative tolerance.

Method. Use the compound Poisson frequency-severity formula. For an exponential severity with mean $1{,}000$, $E[X]=1{,}000$ and $E[X^2]=2(1{,}000)^2=2{,}000{,}000$. The aggregate mean per exposure unit is $\lambda E[X]=2(1{,}000)=\$2{,}000$. Its variance is $\lambda E[X^2]=2(2{,}000{,}000)=4{,}000{,}000$, so its standard deviation is $\$2{,}000$.

For $p=0.95$, use $z_{0.975}=1.959964$, and set $k=0.10$. Then

N_0 =\left(\frac{1.959964(2{,}000)}{0.10(2{,}000)}\right)^2 =(19.59964)^2 \approx384.146.

Rounding upward gives a full-credibility standard of 385 exposure units. The standard deviation here is $\$2{,}000$ per unit because the count is random; using only $\lambda\operatorname{Var}(X)$ would omit the frequency contribution and understate variability.

How to Interpret the Difference

Feature	Bühlmann credibility	Limited-fluctuation credibility
Starting point	Risk-parameter model with within-risk and between-risk variation	Probability, relative-error tolerance, and aggregate variability
Main result	A linear estimate $Z\bar X+(1-Z)m$	A full-credibility exposure threshold and, conventionally, a partial factor
Meaning of the weight	Weight assigned to observed experience in the model-based estimate	Exposure-based factor derived from a probability standard
Key assumptions	Conditional observations follow the Bühlmann structure and have finite variance components	Independent comparable exposure units, finite variance, and an adequate aggregate approximation

For example, a limited-fluctuation result of $Z_{\mathrm{LF}}=0.6080$ does not imply that a Bühlmann estimate should use $Z=0.6080$. The former comes from the target probability and tolerance; the latter comes from $n$, EPV, and VHM. A comparison is meaningful only after stating each method’s inputs and purpose.

The distinction matters in ratemaking because a probability standard can answer whether a volume of experience is sufficiently stable under a specified criterion, while a model-based estimate can blend that experience with a collective expectation. A calculation can be numerically correct yet answer the wrong question if a candidate treats the two weights as interchangeable.

Key takeaway: Bühlmann credibility is a model-based linear blend, with $Z=n/(n+\mathrm{EPV}/\mathrm{VHM})$. Limited-fluctuation credibility translates a probability and relative-error target into $N_0=(z\sigma_Y/(k\mu_Y))^2$. Always identify which quantity is being sought before choosing the formula.

Check Your Understanding

For each problem, identify the method and apply its assumptions before calculating.

A hypothetical portfolio has $\mathrm{EPV}=640$, $\mathrm{VHM}=160$, collective mean $m=\$800$, and eight observations with mean $\$950$. Calculate the Bühlmann factor and estimate. (Answer: $K=4$, $Z=2/3$, estimate $=\$900$.)
An aggregate exposure unit has mean $\$0.40$ and standard deviation $\$0.80$. Under a normal approximation, find the 90% full-credibility threshold for a 25% relative tolerance, rounding up to a whole unit. (Answer: $N_0\approx173.15$, so 174 units.)
Claim counts are Poisson with mean 1.5 per exposure unit, and the limited-fluctuation criterion is a 90% central probability within 10% of expected count. Find the full-credibility threshold, rounded up. (Answer: $N_0\approx180.37$, so 181 exposure units.)
A class has $n=3$, experience mean $\$900$, collective mean $\$600$, $\mathrm{EPV}=900$, and $\mathrm{VHM}=300$. Calculate the Bühlmann estimate and explain whether its factor is a probability guarantee. (Answer: $K=3$, $Z=0.5$, estimate $=\$750$; the factor is not a probability guarantee.)
Frequency is Poisson with mean 1 per unit; severity is exponential with mean $\$500$, independent of frequency. Under a normal approximation, calculate the 95% full-credibility threshold for a 20% relative tolerance, rounding up. (Answer: $N_0\approx192.1$, so 193 exposure units.)

What You'll Learn