Why Blend Class Experience?
A classification loss cost should reflect the experience of the class without allowing a short or unusually favorable history to dominate the indication. Bühlmann-Straub credibility provides a structured way to do this: it gives more weight to a class’s own experience as its exposure grows, while retaining a pooled collective estimate as a stabilizing complement.
The result is a class-level indication, not a prediction for an individual insured. It is useful when classes have different amounts of experience and risk periods have unequal exposures. The calculation is meaningful only when the observations are expressed on a comparable basis—for example, the same coverage, development status, and trend level.
Model and Assumptions
The formula rests on the Bühlmann-Straub hierarchical model. Each class has an underlying mean \(\Theta_i\), which varies across classes. Conditional on that mean, the period observations are unbiased for it, with process variance inversely proportional to exposure:
The standard model also assumes that, conditional on \(\Theta_i\), the period observations are uncorrelated, and that the process variance parameter is common across periods. Exposures must measure comparable amounts of risk; a period with twice the exposure should contribute twice the weight to the class mean. The class means are treated as draws from a common collective structure with mean \(m\) and variance \(\tau^2\).
In an application, \(m\) may be supplied or estimated from the portfolio. A common pooled estimate, when classes share the same exposure basis, is the exposure-weighted mean across the classes:
This is a collective estimate, not a separate complement chosen for each class. The usual plug-in calculation treats the estimated collective mean and structural constant as given; it does not separately allow for uncertainty in their estimation. If experience is not on a common basis, or if the model’s variance assumptions are unsuitable, the formula should not be applied mechanically.
Why the Credibility Factor Has This Form
The class mean contains two sources of variation: differences among underlying class means and random variation in observed experience. Under the model, the conditional variance of the exposure-weighted class mean is \(\sigma^2/T_i\). To see how these components determine credibility, consider a linear estimate that blends the class mean with the collective mean:
Write \(\bar X_i=\Theta_i+\varepsilon_i\), where \(E[\varepsilon_i\mid\Theta_i]=0\) and \(\operatorname{Var}(\varepsilon_i\mid\Theta_i)=\sigma^2/T_i\). Its estimation error is
The two terms have zero covariance under the model. Thus the mean squared error is
Differentiating with respect to \(c\), setting the derivative to zero, and solving gives
This explains the weighting: more exposure reduces the process variance of the class mean, so the class earns more credibility. The structural constant \(K\) is the exposure amount at which the credibility factor is one-half. When \(T_i=K\), \(Z_i=0.5\); as exposure grows, \(Z_i\) approaches 1.
Calculating a Class Indication
Confirm the class-period loss costs use the same exposure unit and are adjusted to the same coverage, development, and trend basis.
Multiply each period’s loss cost by its exposure, sum the products, and divide by the class’s total exposure.
Use the supplied collective mean or calculate the exposure-weighted pooled estimate across the relevant classes.
Find \(Z_i=T_i/(T_i+K)\), then blend the class mean and collective mean using weights \(Z_i\) and \(1-Z_i\).
Verify that the two weights sum to one and that the indicated loss cost lies between the class experience and the collective mean.
Worked Example: Three Classes With Unequal Period Exposures
A hypothetical portfolio has three classes with loss costs in dollars per exposure unit. The experience is on a common basis, and the structural constant is \(K=100\) exposure units. Calculate the pooled collective estimate and each credibility-weighted class indication. Name the method and identify the weighting error to avoid.
| Class | Period 1: exposure, loss cost | Period 2: exposure, loss cost |
|---|---|---|
| A | 100, $120 | 200, $100 |
| B | 50, $160 | 150, $140 |
| C | 100, $80 | 100, $90 |
Method. Use the Bühlmann-Straub class indication, with exposure-weighted means and the pooled exposure-weighted mean as the collective estimate.
Pool the experience. Class A has loss amounts \(100(120)=12{,}000\) and \(200(100)=20{,}000\), so its total is \(32{,}000\) over \(300\) exposure units. Class B has \(50(160)+150(140)=8{,}000+21{,}000=29{,}000\) over \(200\) units. Class C has \(100(80)+100(90)=8{,}000+9{,}000=17{,}000\) over \(200\) units. Therefore, total loss is \(32{,}000+29{,}000+17{,}000=78{,}000\), and total exposure is \(300+200+200=700\). The pooled mean is \(78{,}000/700=\$111.4286\) per exposure unit.
Class A. Its weighted mean is \(32{,}000/300=\$106.6667\), and \(Z_A=300/(300+100)=0.75\). The indication is \(0.75(106.6667)+0.25(111.4286)=80+27.8571=\$107.8571\), or $107.86.
Class B. Its weighted mean is \(29{,}000/200=\$145.00\), and \(Z_B=200/(200+100)=2/3\). The indication is \((2/3)(145)+(1/3)(111.4286)=96.6667+37.1429=\$133.8096\), or $133.81.
Class C. Its weighted mean is \(17{,}000/200=\$85.00\), and \(Z_C=200/(200+100)=2/3\). The indication is \((2/3)(85)+(1/3)(111.4286)=56.6667+37.1429=\$93.8096\), or $93.81.
Check. Each result lies between that class’s experience and the pooled mean. The common mistake is to give each period equal weight: that would ignore the different period exposures and produce incorrect class means. Round only the final indications to the nearest cent.
Worked Example: Unequal Exposures Within One Class
A hypothetical equipment class has two periods: 40 exposure units with a loss cost of $190 per unit, and 60 units with a loss cost of $110 per unit. The collective mean is $130, and \(K=150\) exposure units. Find the class indication. Explain why the simple average of the period loss costs is not the class experience.
Method. Calculate the exposure-weighted class mean first, then apply Bühlmann-Straub credibility.
The total loss amount represented by the experience is \(40(190)+60(110)=7{,}600+6{,}600=\$14{,}200\), and total exposure is \(40+60=100\). Thus \(\bar X=14{,}200/100=\$142\). The credibility factor is \(Z=100/(100+150)=0.4\). The indicated loss cost is \(0.4(142)+0.6(130)=56.8+78=\mathbf{\$134.80}\) per exposure unit.
Check. The indication is between $130 and $142. A simple average would give \((190+110)/2=\$150\), which is not the exposure-weighted experience: the $110 observation represents 60 units, not the same exposure as the $190 observation. The wrong calculation therefore overstates the class mean and does not follow the model’s exposure weighting.
Worked Example: Obtain the Structural Constant From Variances
For a hypothetical class model, the process variance per exposure unit is \(\sigma^2=3{,}600\) and the variance of underlying class means is \(\tau^2=900\). A particular class has period exposures of 1, 2, and 3 units, with loss costs of $180, $120, and $150, respectively. The collective mean is $130. Calculate the class indication.
Method. First calculate \(K=\sigma^2/\tau^2\), then calculate the exposure-weighted class mean and credibility-weighted estimate.
The structural constant is \(K=3{,}600/900=4\) exposure units. Total class exposure is \(T=1+2+3=6\). Total weighted loss is \(1(180)+2(120)+3(150)=180+240+450=870\), so the class mean is \(870/6=\$145\). The credibility factor is \(Z=6/(6+4)=0.6\). Therefore, the indication is \(0.6(145)+0.4(130)=87+52=\mathbf{\$139}\) per exposure unit.
Check. The result is between $130 and $145, as it must be for a blend with weights 0.6 and 0.4. The simple average of the three period loss costs is $150, but it is not the model’s class mean because the periods have unequal exposures.
Interpretation and Common Exam Traps
The indicated class loss cost is a weighted average, not an arbitrary adjustment. A class with little exposure is pulled more strongly toward the collective level; a class with substantial exposure is driven more by its own experience. This makes the indication less sensitive to random variation in small class histories while preserving differences supported by larger amounts of experience.
- Using the number of periods instead of total exposure. \(T_i\) is the sum of the period exposures, not the number of observations.
- Taking an unweighted average of period loss costs. The class mean is \(\sum_j t_{ij}X_{ij}/T_i\); unequal exposures must receive unequal weights.
- Reversing the credibility weights. The class experience receives weight \(Z_i\), while the collective mean receives \(1-Z_i\).
- Mixing incompatible experience. Different coverage, trend, development, or exposure bases can make the pooled estimate and class mean incomparable.
- Treating the pooled mean as a guaranteed standard. It is an estimate of the collective level, and its suitability depends on the classes and experience included.
These checks are especially useful when a class table contains both very different volumes and very different loss costs. A calculation can be arithmetically correct but still misleading if the exposure basis is inconsistent or the chosen collective estimate does not represent the intended portfolio.
Check Your Understanding
For each problem, calculate using the stated exposure units and round money to the nearest cent when needed.
- A class has total exposure of 80 units, an exposure-weighted loss cost of $125, a collective mean of $100, and \(K=120\) units. Find its credibility factor and indicated loss cost. (Answer: \(Z=0.4000\); indicated loss cost = $110.00 per unit.)
- A hypothetical class has two periods: 20 units at $150 per unit and 30 units at $90 per unit. The collective mean is $100 and \(K=50\) units. Find the weighted class mean and indication. (Answer: weighted class mean = $114.00; \(Z=0.5000\); indication = $107.00 per unit.)
- A class has total exposure \(T=240\), experience of $92 per unit, a collective mean of $104, and \(K=160\). Calculate its indicated loss cost and state which estimate receives more weight. (Answer: \(Z=0.6000\); indication = $96.80 per unit; class experience receives more weight.)
- Two classes have exposure-weighted means of $80 and $120, with total exposures 100 and 300 units, respectively. Use the pooled exposure-weighted mean as the collective estimate and \(K=100\). Find both indications. (Answer: pooled mean = $110.00; Class 1 indication = $95.00; Class 2 indication = $117.50 per unit.)
- For a class model, \(\sigma^2=2{,}400\), \(\tau^2=600\), total class exposure is 12 units, the class experience is $75 per unit, and the collective mean is $90. Calculate \(K\), \(Z\), and the indicated loss cost. (Answer: \(K=4\) exposure units; \(Z=0.7500\); indication = $78.75 per unit.)