How Experience Volume Changes Bühlmann Credibility

Why Experience Volume Changes Credibility

A class with more observations generally provides more information about its underlying expected result. In Bühlmann credibility, this effect appears through the credibility factor: with more observations, the experience mean receives more weight relative to the collective mean. The key comparison in this tutorial holds the model’s variance components fixed and changes only the number of observations.

The observations here are assumed to be equally weighted and to follow the basic Bühlmann model. This is not a model for observations with different exposure weights; the quantity being varied is the count of observations, each with the same role in the sample mean. The model does not require a particular probability distribution, so the examples specify the moment assumptions needed for the calculation rather than imposing a distributional form. There is no interest or mortality assumption in these examples.

Definition: Let $\Theta$ be a risk’s hypothetical expected result, with collective mean $m=E[\Theta]$ and variance of hypothetical means $\tau^2=\operatorname{Var}(\Theta)$, also called the variance of hypothetical means (VHM). Let $X_1,\ldots,X_n$ be conditionally independent observations with $E[X_j\mid\Theta]=\Theta$ and $\operatorname{Var}(X_j\mid\Theta)=\phi$, where $\phi$ is the expected value of process variance (EPV). For $n$ equally weighted observations, the Bühlmann credibility factor and estimate are

Z_n=\frac{n}{n+K}, \qquad K=\frac{\phi}{\tau^2}, \qquad \widehat{\Theta}_n=Z_n\overline{X}+(1-Z_n)m, \qquad \overline{X}=\frac{1}{n}\sum_{j=1}^n X_j.

The formula presumes that $\phi$ and $\tau^2$ are positive and finite. If $\tau^2=0$, there is no variation among hypothetical means, so this expression for $K$ is not defined; the model gives no between-risk variation for experience to distinguish. The factor $K$ is often called the Bühlmann credibility parameter. It is fixed when the variance components are fixed, while $Z_n$ changes with $n$.

Why the Factor Increases with the Observation Count

The sample mean contains both process variation and variation among the underlying hypothetical means. Under the stated conditional independence and moment assumptions, its variance is $\tau^2+\phi/n$. The process-variance contribution becomes smaller as $n$ grows, while the between-risk variation remains. This is the reason that a larger sample mean deserves more weight.

Find the sample-mean variance.
By conditional independence, the conditional variance of $\overline{X}$ is $\phi/n$. The law of total variance then gives $\operatorname{Var}(\overline{X})=\tau^2+\phi/n$.

Find its covariance with the hypothetical mean.
Because $E[\overline{X}\mid\Theta]=\Theta$, $\operatorname{Cov}(\Theta,\overline{X})=\operatorname{Var}(\Theta)=\tau^2$.

Obtain the linear-projection coefficient.
The coefficient of $\overline{X}$ in the best linear projection of $\Theta$ on $\overline{X}$ is $\operatorname{Cov}(\Theta,\overline{X})/\operatorname{Var}(\overline{X})$. Substitution gives $\tau^2/(\tau^2+\phi/n)=n/(n+\phi/\tau^2)=Z_n$. The corresponding intercept makes the estimate equal to $m$ when $\overline{X}=m$, yielding $Z_n\overline{X}+(1-Z_n)m$.

This derivation explains the Bühlmann estimate as a linear projection under the stated moment model. It does not require a fully specified distribution, and the estimate is not, in general, an exact Bayesian posterior mean. The ratio $K=\phi/\tau^2$ compares within-risk process variation with between-risk variation. Large $K$ means that many observations are needed before the sample mean receives substantial credibility.

How the Factor Changes as Volume Grows

For fixed $K>0$, $Z_n=n/(n+K)$ is increasing in $n$, but it does not increase by the same amount for every added observation. If $n$ is treated as continuous, its first derivative is $K/(n+K)^2>0$, and its second derivative is $-2K/(n+K)^3<0$. Thus the factor rises at a decreasing rate. For integer observation counts, the exact change from $n$ to $n+1$ is:

Z_{n+1}-Z_n =\frac{n+1}{n+1+K}-\frac{n}{n+K} =\frac{K}{(n+K)(n+1+K)}.

This increment is positive and declines as $n$ increases. Also, $Z_n$ approaches 1 as $n$ grows without bound; for any finite $n$ and positive $K$, it remains below 1. The estimate therefore moves toward the observed mean as volume increases, but does not assign it full weight at any finite count.

Situation, with parameters fixed	Effect on the credibility factor	Effect on the estimate
Increase $n$	$Z_n$ increases toward 1	More weight goes to $\overline{X}$
Increase $\phi$	$K$ increases, so $Z_n$ decreases at a given $n$	More weight remains on $m$
Increase $\tau^2$	$K$ decreases, so $Z_n$ increases at a given $n$	Experience more readily differentiates risks

Worked Calculations

Worked Example: Comparing Three Experience Volumes

A hypothetical rating class has collective mean $m=\$100$, EPV $\phi=40{,}000$ dollars squared, and VHM $\tau^2=10{,}000$ dollars squared. Suppose its observed mean is $\overline{X}=\$140$. Compare the Bühlmann factor and estimate for $n=1,4,$ and $16$. Assume the equal-weight Bühlmann moment conditions stated above; no parametric distribution, interest rate, or mortality assumption is required.

Method and parameter check. The variance components stay fixed, so first calculate $K=40{,}000/10{,}000=4$. The experience mean is $\$40$ above the collective mean, so each estimate must lie between $\$100$ and $\$140$.

One observation. $Z_1=1/(1+4)=0.2$. Thus $\widehat{\Theta}_1=0.2(\$140)+0.8(\$100)=\$28+\$80=\$108$.

Four observations. $Z_4=4/(4+4)=0.5$. Thus $\widehat{\Theta}_4=0.5(\$140)+0.5(\$100)=\$70+\$50=\$120$.

Sixteen observations. $Z_{16}=16/(16+4)=0.8$. Thus $\widehat{\Theta}_{16}=0.8(\$140)+0.2(\$100)=\$112+\$20=\$132$.

The check is consistent: the factors $0.2,0.5,0.8$ rise with volume, and the estimates $\$108,\$120,\$132$ move toward $\$140$. No intermediate result required rounding.

Common mistake avoided: $n$ is the number of observations, not the number of dollars of variance or the observed mean. The ratio of the variance components determines $K=4$; the observation count is then substituted in $n/(n+K)$.

Worked Example: Separating the Factor Change From the Mean Change

A hypothetical class has $m=\$100$, $\phi=600$ dollars squared, and $\tau^2=100$ dollars squared. Its experience mean is $\$130$. Calculate the estimates for $n=2,6,$ and $18$, and state how much of the difference between the first and last estimates is attributable to the changing credibility factor. Assume the equal-weight moment model, with no specified parametric distribution, interest rate, or mortality assumption.

Method and parameter check. $K=600/100=6$. Across all three calculations, $\overline{X}-m=\$30$; therefore $\widehat{\Theta}_n=m+Z_n(\overline{X}-m)=\$100+\$30Z_n$.

For $n=2$. $Z_2=2/(2+6)=0.25$, so $\widehat{\Theta}_2=\$100+0.25(\$30)=\$107.50$.

For $n=6$. $Z_6=6/(6+6)=0.5$, so $\widehat{\Theta}_6=\$100+0.5(\$30)=\$115.00$.

For $n=18$. $Z_{18}=18/(18+6)=0.75$, so $\widehat{\Theta}_{18}=\$100+0.75(\$30)=\$122.50$.

The first-to-last estimate change is $\$122.50-\$107.50=\$15.00$. Equivalently, the factor rises by $0.75-0.25=0.50$, and that change multiplied by the fixed $\$30$ gap gives $0.50(\$30)=\$15.00$. The check confirms that the difference comes only from the changed factor: $m$, $\overline{X}$, and both variance components were held fixed.

Common mistake avoided: A rise in credibility of 0.50 is not a 50% increase in the estimate. The estimate changes by the credibility-factor change multiplied by the difference between experience and complement.

Worked Example: Finding the Count Needed for a Target Factor

A hypothetical class has EPV $\phi=800$ dollars squared and VHM $\tau^2=200$ dollars squared. Its collective mean is $\$100$, and its observed mean is $\$150$. Find the smallest integer $n$ for which $Z_n$ is at least $0.80$, then calculate the estimate at that count and at the count immediately below it. Assume equal-weight observations under the Bühlmann moment model; no parametric distribution, interest rate, or mortality assumption is specified.

Method and parameter check. $K=800/200=4$. To find the count, solve the target inequality rather than rounding a credibility factor calculated at an arbitrary count:

\frac{n}{n+4}\geq 0.80 \quad\Longleftrightarrow\quad n\geq 0.80(n+4) \quad\Longleftrightarrow\quad 0.20n\geq 3.20 \quad\Longleftrightarrow\quad n\geq 16.

So the smallest integer count is $16$. At $n=16$, $Z_{16}=16/20=0.80$, and $\widehat{\Theta}_{16}=0.80(\$150)+0.20(\$100)=\$120+\$20=\$140$.

At $n=15$, $Z_{15}=15/19\approx0.7895$, which is below the target. Its estimate is $(15/19)(\$150)+(4/19)(\$100)=\$2{,}650/19\approx\$139.47$, rounded to the nearest cent. This confirms that $16$, not $15$, is the smallest count satisfying the requirement.

Common mistake avoided: Solving $n/(n+K)=0.80$ and reporting a noninteger count would not answer a question about the smallest number of observations. The inequality and integer check establish the threshold.

Interpreting Volume Changes Carefully

A larger $n$ increases the experience weight only when $K$ is held fixed. If EPV or VHM changes, then $K$ changes too, and the volume comparison alone no longer describes the full effect. In practice, estimated variance components may also be uncertain; the calculations here deliberately treat them as fixed so the effect of observation count is isolated.

The Bühlmann factor describes the weight assigned to the experience mean, not the proportion by which the class estimate rises or falls. Rearranging the estimate makes this distinction explicit:

\widehat{\Theta}_n-m=Z_n(\overline{X}-m).

The estimate’s displacement from the collective mean equals the observed displacement multiplied by credibility. If $\overline{X}=m$, changing $n$ changes the factor but not the estimate: it remains $m$. If $\overline{X}<m$, greater credibility moves the estimate downward toward $\overline{X}$, just as it moves it upward when $\overline{X}>m$.

A useful final check is therefore to compare the estimate with both means. For positive $K$ and finite $n$, $0<Z_n<1$, so the estimate must be between $m$ and $\overline{X}$, inclusive. An answer outside that interval usually signals a complement-weight error, an incorrect factor, or arithmetic inconsistent with the linear blend.

Key takeaway: With EPV and VHM fixed, additional equally weighted observations increase $Z_n=n/(n+K)$ toward 1, with smaller gains at higher counts. The class estimate moves toward the experience mean by exactly the credibility-weighted portion of its difference from the collective mean.

Check Your Understanding

Use the Bühlmann factor and estimate for each hypothetical scenario. Assume the equal-weight moment model; a parametric distribution, interest rate, and mortality assumption are not specified or needed.

A class has EPV $500$ dollars squared, VHM $100$ dollars squared, collective mean $\$100$, experience mean $\$80$, and $n=5$. Calculate $K$, $Z_n$, and the estimate. (Answer: $K=5$, $Z_5=0.5000$, estimate $=\$90.00$)
For a class with EPV $1{,}200$ dollars squared, VHM $300$ dollars squared, $m=\$90$, and $\overline{X}=\$110$, calculate the estimate at $n=12$. (Answer: $K=4$, $Z_{12}=0.7500$, estimate $=\$105.00$)
A model has EPV $600$ dollars squared and VHM $100$ dollars squared. What is the smallest integer $n$ that gives $Z_n\geq0.90$? (Answer: $K=6$; $n/(n+6)\geq0.90$ requires $n\geq54$, so the smallest count is $54$)
With $m=\$110$, $\overline{X}=\$140$, EPV $900$ dollars squared, and VHM $100$ dollars squared, calculate the estimates for $n=9$ and $n=27$, then find their difference. (Answer: $K=9$, $Z_9=0.5000$, estimate $=\$125.00$; $Z_{27}=0.7500$, estimate $=\$132.50$; difference $=\$7.50$)
A class has $n=12$, $m=\$90$, and $\overline{X}=\$150$. Initially EPV is $600$ dollars squared and VHM is $200$ dollars squared. Recalculate the factor and estimate if EPV doubles while VHM stays fixed. (Answer: Initially $K=3$, $Z_{12}=0.8000$, estimate $=\$138.00$. After EPV doubles, $K=6$, $Z_{12}=12/18=0.6667$, estimate $=\$130.00$.)

Situation, with parameters fixed	Effect on the credibility factor	Effect on the estimate
Increase \(n\)	\(Z_n\) increases toward 1	More weight goes to \(\overline{X}\)
Increase \(\phi\)	\(K\) increases, so \(Z_n\) decreases at a given \(n\)	More weight remains on \(m\)
Increase \(\tau^2\)	\(K\) decreases, so \(Z_n\) increases at a given \(n\)	Experience more readily differentiates risks

What You'll Learn