Why Class Variability Matters
Two classes can have the same number of observations but very different levels of noise. For example, losses in one class may be relatively stable from risk to risk, while losses in another class fluctuate substantially. If the class means are otherwise modeled on the same basis, the noisier class provides less precise evidence about its underlying expected loss cost. Applying an identical credibility factor to both classes can therefore give too much weight to the noisier class and too little to the more stable one.
The key distinction is between process variance, which describes variation in observations around a class’s own expected cost, and between-class variance, which describes variation in the underlying expected costs among classes. Credibility depends on their relative sizes, as well as on the amount of exposure behind each class mean.
Model and Class-Specific Credibility
Suppose each class has \(n_j\) independent observations, each with conditional mean \(\Theta_j\) and conditional variance \(\sigma_j^2\). Assume the classes share a collective mean \(\mu=E[\Theta_j]\) and a common between-class variance \(\tau^2=\operatorname{Var}(\Theta_j)\). For this comparison, the process variance may differ by class, but the collective mean and between-class variance are common. The class sample mean is \(\bar X_j\).
Under these assumptions, the credibility-weighted indication is \(\widehat{\Theta}_j=\mu+Z_j(\bar X_j-\mu)\), where the class-specific credibility factor is:
This is the Bühlmann best linear credibility factor for the class mean under the stated model. If observations and class means are normally distributed, the weighted indication is also the exact posterior mean. Without normality, it remains the best linear predictor under the stated moment assumptions; it need not equal the full Bayesian posterior mean.
The denominator compares between-class variance with the total variance of a class sample mean. That sample-mean variance includes both the variation in underlying class means and the class’s process noise:
Thus, higher process variance or lower exposure reduces credibility, while higher between-class variance increases it. The formula does not apply unchanged if the classes have different collective means or different between-class variances; those differences require a model that reflects them rather than treating them as process variance alone.
Why the Weight Has This Form
The factor can be justified by finding the linear estimate closest to the true class mean on average. Consider an estimate of the form \(\mu+Z(\bar X_j-\mu)\). Its mean squared prediction error is minimized when the coefficient \(Z\) is the covariance of the target and the data-based deviation divided by the variance of that deviation.
The target deviation is \(\Theta_j-\mu\), and the observed deviation is \(\bar X_j-\mu\). Both have mean zero.
Conditional on \(\Theta_j\), the sample mean has expectation \(\Theta_j\). Therefore, \(\operatorname{Cov}(\Theta_j,\bar X_j)=\operatorname{Cov}(\Theta_j,E[\bar X_j\mid\Theta_j])=\operatorname{Var}(\Theta_j)=\tau^2\).
The law of total variance gives \(\operatorname{Var}(\bar X_j)=\tau^2+\sigma_j^2/n_j\), since the conditional variance of the mean of \(n_j\) independent observations is \(\sigma_j^2/n_j\).
For a linear predictor, the minimizing coefficient is the covariance divided by the variance. Hence \(Z_j=\tau^2/(\tau^2+\sigma_j^2/n_j)\), as stated.
The derivation shows why a single factor is not automatically appropriate just because classes use the same collective complement. With common \(\tau^2\), classes have the same factor precisely when their process variance per observation divided by exposure, \(\sigma_j^2/n_j\), is the same. In practical terms, their sample means must have equal process-noise variance relative to the common between-class variance.
Worked Class Indications
Worked Example: Equal Exposure, Different Process Variance
A hypothetical insurer uses a collective mean of \(\mu=\$1{,}000\) and a between-class variance of \(\tau^2=40{,}000\) dollars squared. Two classes each have four independent observations. Class A has process variance \(\sigma_A^2=160{,}000\) and observed mean \(\bar X_A=\$1{,}300\); Class B has \(\sigma_B^2=640{,}000\) and \(\bar X_B=\$700\). Calculate the class-specific indications. Identify the mistake avoided.
Method. Apply the class-specific Bühlmann factor, using each class’s own process variance. For Class A:
Its indication is \(\widehat{\Theta}_A=1{,}000+0.5(1{,}300-1{,}000)=\$1{,}150\).
For Class B:
Its indication is \(\widehat{\Theta}_B=1{,}000+0.2(700-1{,}000)=\$940\). The results are reasonable checks: Class A is moved halfway from its observed mean to the collective mean; the noisier Class B is moved 80% of the way toward the collective mean.
Common mistake avoided. Using one factor based on an average process variance, \((160{,}000+640{,}000)/2=400{,}000\), gives \(Z=160{,}000/(160{,}000+400{,}000)=0.2857\). That would produce about \(\$1{,}085.71\) for A and \(\$914.29\) for B. This calculation silently imposes a common process variance and does not use either class’s specified variance correctly. Money is rounded to the nearest cent only at the final step.
Worked Example: Different Variances but a Common Factor
A hypothetical program has collective mean \(\mu=\$800\) and between-class variance \(\tau^2=30{,}000\) dollars squared. Class A has \(n_A=2\), process variance \(\sigma_A^2=90{,}000\), and observed mean \(\$1{,}100\). Class B has \(n_B=8\), process variance \(\sigma_B^2=360{,}000\), and observed mean \(\$650\). Determine whether a common factor is justified under the model and calculate both indications. Identify the trap.
Method. Compare the process variance of each sample mean. For A it is \(90{,}000/2=45{,}000\); for B it is \(360{,}000/8=45{,}000\). Since both classes have the same \(\tau^2\), their factors agree:
The indication for A is \(800+0.4(1{,}100-800)=\$920\). The indication for B is \(800+0.4(650-800)=\$740\). Although their observation-level process variances differ, their sample means have equal process variance because B has four times the exposure. Under the stated assumptions, the common factor is appropriate.
Common mistake avoided. Looking only at \(\sigma_A^2\) and \(\sigma_B^2\) would suggest different factors. The relevant comparison is \(\sigma_j^2/n_j\), not process variance alone. Money is rounded to the nearest cent only at the final step.
Worked Example: Variance Assumptions Can Change an Indication Ranking
A hypothetical classification has collective mean \(\mu=\$1{,}200\) and between-class variance \(\tau^2=25{,}000\) dollars squared. Both classes have five observations. Class A has process variance \(100{,}000\) and observed mean \(\$1{,}400\); Class B has process variance \(400{,}000\) and observed mean \(\$1{,}500\). Calculate the indications using the stated variances, then compare them with indications from an unjustified common factor based on the average process variance. State the mistake this comparison exposes.
Method. For A, \(n_A\tau^2=5(25{,}000)=125{,}000\), so \(Z_A=125{,}000/(125{,}000+100{,}000)=5/9\). Its indication is \(1{,}200+(5/9)(200)=\$1{,}311.11\).
For B, \(Z_B=125{,}000/(125{,}000+400{,}000)=5/21\). Its indication is \(1{,}200+(5/21)(300)=\$1{,}271.43\). Thus A’s indicated cost exceeds B’s, even though B’s raw mean is higher, because B’s experience is much less credible under the specified variance assumptions.
If the variances were incorrectly replaced by their average, \(250{,}000\), both factors would be \(125{,}000/(125{,}000+250{,}000)=1/3\). That would give A an indication of \(1{,}200+200/3=\$1{,}266.67\) and B an indication of \(1{,}200+300/3=\$1{,}300\), reversing the ranking. The error is not that a common factor always reverses a ranking; it is that this common factor ignores the stated class-level difference in process precision. Money is rounded to the nearest cent only at the final step.
Interpreting and Applying the Comparison
A common factor may still be defensible when process variances differ, provided exposure offsets the difference so that the sample means have equal process-noise variance, as in the second example. Conversely, equal exposure does not justify a common factor when process variances differ. The decision should follow the model’s variance assumptions, not a preference for simpler calculations.
| Situation under the stated model | Implication for credibility |
|---|---|
| Same \(\tau^2\), same \(n_j\), different \(\sigma_j^2\) | Different factors; the class with lower process variance has higher credibility. |
| Same \(\tau^2\), different \(n_j\) and \(\sigma_j^2\), but equal \(\sigma_j^2/n_j\) | Same factors; the sample means have equal process-noise variance. |
| Same process variance and same exposure, but different \(\tau^2\) | The common between-class variance assumption fails; the stated common-factor comparison is not sufficient. |
In practice, process variance may be estimated rather than known. Such estimates can be uncertain, especially with limited data. A calculated difference between class factors should therefore be interpreted in light of the credibility model and the basis of the variance estimates; it is not, by itself, proof that every observed difference is a permanent characteristic of the class. The formula also assumes conditional independence of observations and a meaningful exposure count \(n_j\). If observations are dependent or have unequal variances within a class, the variance of the actual class estimator must be modeled instead of mechanically using \(\sigma_j^2/n_j\).
- Do not confuse process variance with between-class variance. Increasing \(\sigma_j^2\) makes that class’s observed mean less precise; increasing \(\tau^2\) makes class deviations from the collective mean more important.
- Do not average class variances without a reason. A pooled variance may be suitable only if the model and estimation method justify it; it is not a substitute for given class-specific assumptions.
- Check the unit of the variance. If the loss cost is in dollars, variances are in dollars squared. The ratio in the credibility formula is dimensionless.
- Check the direction of the adjustment. Since \(0\le Z_j\le1\), an indication lies between \(\mu\) and \(\bar X_j\); the lower-credibility class lies closer to \(\mu\).
Check Your Understanding
For each problem, use the stated common collective mean and between-class variance. Round money to the nearest cent where needed.
- A class has \(n=4\), \(\mu=\$900\), \(\tau^2=36{,}000\), \(\sigma^2=108{,}000\), and \(\bar X=\$1{,}200\). Find its factor and indication. (Answer: \(Z=4/7\); indication \(=\$1{,}071.43\))
- Two classes have \(n=3\) and \(\tau^2=18{,}000\), with common \(\mu=\$1{,}100\). Class A has \(\sigma_A^2=54{,}000\) and \(\bar X_A=\$1{,}000\); Class B has \(\sigma_B^2=162{,}000\) and \(\bar X_B=\$1{,}300\). Find both factors and indications. (Answer: \(Z_A=0.5000\), indication \(=\$1{,}050\); \(Z_B=0.2500\), indication \(=\$1{,}150\))
- Class A has \(n_A=2,\sigma_A^2=80{,}000,\bar X_A=\$900\); Class B has \(n_B=8,\sigma_B^2=320{,}000,\bar X_B=\$600\). If \(\mu=\$700\) and \(\tau^2=40{,}000\), determine whether a common factor applies and calculate both indications. (Answer: Yes; both factors are \(0.5000\); indications are \(\$800\) and \(\$650\))
- Two classes each have \(n=5\), with \(\mu=\$1{,}000\) and \(\tau^2=50{,}000\). Class A has \(\sigma_A^2=100{,}000,\bar X_A=\$1{,}200\); Class B has \(\sigma_B^2=200{,}000,\bar X_B=\$900\). Calculate the factors and indications, and state whether the factors match. (Answer: \(Z_A=0.7143\), indication \(=\$1{,}142.86\); \(Z_B=0.5556\), indication \(=\$944.44\); the factors do not match.)
- A class system has \(\mu=\$1{,}500,\tau^2=10{,}000\), and each class has four observations. Class A has process variance \(40{,}000\) and mean \(\$1{,}800\); Class B has process variance \(160{,}000\) and mean \(\$2{,}000\). Find the factors and indications, and explain why the higher raw mean need not produce the higher indication. (Answer: \(Z_A=0.5000\), indication \(=\$1{,}650\); \(Z_B=0.2000\), indication \(=\$1{,}600\); B’s noisier experience receives less weight.)