Limiting Behavior of the Bühlmann Factor

Why the Limits Matter

The Bühlmann credibility factor describes how much weight an experience estimate receives relative to the collective mean. Its value depends on the observation volume $n$ and on $K$, the ratio of expected process variance to variance of hypothetical means. Studying what happens at very small and very large volumes helps check whether a calculated factor is reasonable: little experience should generally receive little weight, while a large amount should receive more.

The limits are conditional on the variability in the model. In particular, the familiar result that credibility approaches zero as volume vanishes and one as volume grows requires both positive expected process variance (EPV) and positive variance of hypothetical means (VHM). If either component is zero, the limiting behavior changes.

Definition: In the Bühlmann model, let $\Theta$ denote a risk parameter and $X_j$ the observation for period $j$. The EPV is $v=E[\operatorname{Var}(X_j\mid\Theta)]$, and the VHM is $a=\operatorname{Var}(E[X_j\mid\Theta])$. When $v>0$ and $a>0$, define $K=v/a$. For $n$ observations, the Bühlmann factor is $Z_n=n/(n+K)$. In a continuous-exposure formulation, $n$ denotes exposure volume in the specified unit; its variance scaling must be consistent with that unit.

Deriving the Factor and Its Limits

Assume observations are conditionally independent and identically distributed given $\Theta$, with finite second moments. For the sample mean $\bar X$, conditional variance is $\operatorname{Var}(\bar X\mid\Theta)=\operatorname{Var}(X_j\mid\Theta)/n$. Taking expectations gives the within-risk contribution $v/n$. The variability of the conditional mean across risks is $a$. The optimal linear credibility weight therefore compares the between-risk variance $a$ with the total variance $a+v/n$:

Z_n=\frac{a}{a+v/n} =\frac{na}{na+v} =\frac{n}{n+v/a} =\frac{n}{n+K}.

This also explains the limits. When $v>0$ and $a>0$, $K$ is finite and positive. As $n$ decreases toward zero, the numerator of $n/(n+K)$ vanishes while its denominator approaches $K$. As $n$ grows without bound, divide numerator and denominator by $n$: the resulting expression approaches one.

\lim_{n\to 0^+} Z_n=0, \qquad \lim_{n\to\infty} Z_n =\lim_{n\to\infty}\frac{1}{1+K/n}=1.

The $n\to0^+$ statement uses a continuous-exposure interpretation. In the classical model, $n$ is an integer count, so there is no positive integer sequence converging to zero. Moreover, with zero observations there is no observed sample mean to use. The limit describes how the credibility weight behaves as a positive volume becomes arbitrarily small; it does not create an experience estimate from no data.

Two equivalent expressions make the rate of convergence easier to interpret:

1-Z_n=\frac{K}{n+K}, \qquad \frac{dZ_n}{dn}=\frac{K}{(n+K)^2}>0.

Thus credibility increases with volume, but the marginal increase diminishes as $n$ grows. For small $n$ relative to $K$, $Z_n$ is approximately $n/K$. For large $n$ relative to $K$, the remaining gap from full credibility is approximately $K/n$. These approximations describe behavior, not exact values; use the exact factor for a requested calculation.

Worked Example: Checking Both Limits at Two Volumes

A hypothetical portfolio has EPV $v=\$1{,}200$ and VHM $a=\$200$, with both quantities measured in squared units of the same observation basis. The exposure-scaled Bühlmann model applies, and both variance components are positive. Find the factor at $n=3$ and $n=60$, then use the results to assess the limiting behavior.

Method. Calculate $K=v/a$, then evaluate $Z_n=n/(n+K)$. This comparison illustrates the low- and high-volume behavior without treating either finite volume as an actual limit.

First calculate the structural ratio:

K=\frac{1{,}200}{200}=6.

At $n=3$,

Z_3=\frac{3}{3+6}=\frac{3}{9}=\frac{1}{3}=0.3333.

At $n=60$,

Z_{60}=\frac{60}{60+6}=\frac{60}{66} =\frac{10}{11}=0.9091.

The corresponding gaps from one are $1-Z_3=0.6667$ and $1-Z_{60}=0.0909$. The factor rises as volume grows, and the larger volume is substantially closer to one. The check is consistent with both limits: small positive volume has a relatively small factor, and much larger volume has a factor nearer one. Neither $0.3333$ nor $0.9091$ is itself a limiting value.

Common mistake avoided: Do not say that a finite, large volume gives $Z=1$. The formula gives $Z<1$ for every finite $n$ when $K>0$; one is approached only in the limit.

Solving for a Target Credibility

The limiting result is also useful when a problem asks how much exposure is needed to reach a stated credibility level. For $0<q<1$ and $K>0$, solving $Z_n\ge q$ gives:

\frac{n}{n+K}\ge q \quad\Longleftrightarrow\quad n\ge q(n+K) \quad\Longleftrightarrow\quad n(1-q)\ge qK \quad\Longleftrightarrow\quad n\ge\frac{qK}{1-q}.

For a continuous exposure volume, the result can be a non-integer threshold. If $n$ must be a whole number of observations, round the threshold up to the next integer. The units of $n$ must remain those used to calculate $K$.

Worked Example: Finding the 95% Credibility Threshold

For a hypothetical risk class, the fitted Bühlmann parameter is $K=1.5$ observation units. Assume positive EPV and VHM and use the same unit for $n$. Find the minimum continuous volume for $Z_n\ge0.95$, then find the minimum whole-number volume if observations must be counted in integers.

Method. Rearrange the inequality before substituting. This avoids confusing the proportion $0.95$ with its complement $0.05$.

Starting with the factor,

\frac{n}{n+1.5}\ge0.95.

Multiply by the positive denominator $n+1.5$, then collect the terms containing $n$:

n\ge0.95n+1.425, \qquad 0.05n\ge1.425, \qquad n\ge\frac{1.425}{0.05}=28.5.

Therefore, the minimum continuous exposure is $28.5$ units. If $n$ must be an integer, the minimum is $29$ observations. Check the boundary: at $n=28.5$, $Z=28.5/(28.5+1.5)=28.5/30=0.95$. At $n=29$, $Z=29/30.5=0.9508$ to four decimal places. The next smaller integer gives $28/29.5=0.9492$, below the target.

Common mistake avoided: The threshold is $qK/(1-q)$, not $K(1-q)/q$. The latter reverses the target and its complement and produces the wrong exposure requirement.

How EPV and VHM Affect the Limits

The usual limits require both variance components to be strictly positive. Their roles differ: EPV represents variation within a risk at a fixed risk parameter, while VHM represents variation in conditional means among risks. Their ratio $K=v/a$ sets the volume scale over which credibility changes. Holding EPV fixed, a larger VHM makes $K$ smaller, so a given volume receives more credibility. Holding VHM fixed, a larger EPV makes $K$ larger, so the factor is lower.

The following boundary cases clarify why positivity matters:

Positive EPV and positive VHM: $K$ is finite and positive, so $Z_n\to0$ as $n\to0^+$ and $Z_n\to1$ as $n\to\infty$.
Zero VHM and positive EPV: conditional means do not vary across risks. The direct expression $Z_n=na/(na+v)$ gives zero for every finite $n$. There is no between-risk signal for experience to identify, so the factor does not tend to one.
Zero EPV and positive VHM: there is no within-risk process variation in the model. For every positive $n$, the direct expression gives $Z_n=1$. At $n=0$, however, there is still no observation; the endpoint is not an observed estimate.
Both components zero: the variance ratio is $0/0$, so the variance-based credibility factor is not identified. Since there is no modeled variation in either component, a weight has no usual variance-minimization interpretation.

These cases are not obtained by casually substituting zero into $K=v/a$: when $a=0$, that ratio is undefined, and when both components are zero it is indeterminate. Use the direct variance expression $na/(na+v)$ to inspect a boundary case, and state whether its denominator is positive.

Worked Example: Comparing Portfolios With Different VHM

Two hypothetical portfolios use the same exposure unit and have $n=10$. Both have EPV $v=900$. Portfolio A has VHM $a_A=300$, while Portfolio B has VHM $a_B=100$. Assume each portfolio satisfies the Bühlmann model and both variance components are positive. Calculate each $K$ and factor, then explain the difference.

Method. Calculate each EPV-to-VHM ratio separately, then apply the factor formula. This isolates the effect of changing VHM while holding EPV and volume fixed.

For Portfolio A,

K_A=\frac{900}{300}=3, \qquad Z_{10,A}=\frac{10}{10+3}=\frac{10}{13}=0.7692.

For Portfolio B,

K_B=\frac{900}{100}=9, \qquad Z_{10,B}=\frac{10}{10+9}=\frac{10}{19}=0.5263.

Portfolio A has the larger VHM, so its risks have more variation in their conditional means. With the same within-risk variation and exposure, experience is more informative about which conditional mean applies; accordingly, $Z_{10,A}>Z_{10,B}$. The calculated difference is $0.7692-0.5263=0.2429$, to four decimal places.

Common mistake avoided: Do not interpret a larger $K$ as more credibility. Since $K=\text{EPV}/\text{VHM}$ appears in the denominator of $n/(n+K)$, a larger $K$ means lower credibility at fixed $n$.

Common Exam Traps and Interpretation

A limiting statement is a model result, not a guarantee that a practical portfolio will reach the limit. The factor can approach one slowly if $K$ is large, and an exposure volume that seems large in ordinary terms may still be modest relative to $K$. Conversely, a small $K$ can produce substantial credibility at a relatively low volume. Always compare volume with the model’s $K$, rather than labeling volume “large” or “small” without reference to the parameter.

Several checks help prevent errors:

For $K>0$ and $n>0$, verify $0<Z_n<1$.
At fixed $K$, increasing $n$ must increase $Z_n$; the derivative $K/(n+K)^2$ is positive.
At fixed $n$, increasing $K$ must decrease $Z_n$.
When solving for a target $q$, use $n\ge qK/(1-q)$, and round upward only if the exposure count must be an integer.
Keep exposure units consistent. A value of $K$ calibrated in one observation unit cannot be paired with $n$ measured in another without converting the variance basis.

The factor’s convergence explains its qualitative meaning: with vanishing positive exposure, experience contributes almost no weight; with increasing exposure, process noise is averaged down relative to between-risk variation. This conclusion concerns the credibility factor itself. It does not, by itself, specify a premium or replace the need to define the collective estimate and experience quantity in an application.

Key takeaway: If EPV and VHM are both positive, $Z_n=n/(n+K)$ increases from zero toward one as exposure volume grows. The value of $K=\text{EPV}/\text{VHM}$ controls how quickly it does so; zero variance components require separate boundary analysis.

Check Your Understanding

For each problem, use the stated observation unit and give the requested factor or threshold.

A hypothetical model has $K=8$ and $n=2$. Calculate $Z_n$. (Answer: $0.2000$)
A risk class has $K=3$. What is the minimum continuous volume for credibility of at least $0.80$? (Answer: $n\ge12$ units)
A model has EPV $450$, VHM $150$, and $n=9$. Calculate $K$ and $Z_n$. (Answer: $K=3$, $Z_n=0.7500$)
Two hypothetical groups have the same $n=6$ and EPV $600$. Group R has VHM $200$; Group S has VHM $100$. Calculate both factors and identify the higher one. (Answer: $Z_R=0.6667$, $Z_S=0.5000$; Group R is higher)
In a model with positive EPV but zero VHM, what does the direct variance expression imply for the factor at every positive finite volume, and what does that say about the usual large-volume limit? (Answer: $Z_n=0$ for every positive finite $n$; it does not approach one)

What You'll Learn