The Role of Credibility in Class Rating

Why Credibility Matters in Class Rating

A class rate is intended to reflect the expected cost of the risks assigned to that class. Its own observed loss cost is relevant, but it may be volatile, particularly when the class has limited experience. A collective estimate provides another source of information: the expected cost for a broader group judged suitable as a complement. Credibility blends these two estimates, so the class indication reflects its experience without treating that experience as perfectly reliable.

The distinction between the level of analysis and the unit being rated is essential. In class rating, the experience being summarized belongs to a class, such as a hypothetical category of small commercial workshops. An individual workshop’s own claims history is a separate experience-rating question. A class indication may inform the rate applied to members of that class, but it is not automatically an estimate of any one member’s expected cost.

For equal-sized, comparable observations under the Bühlmann model, let \(X_j\) be the loss cost for observation \(j\), and let \(\bar X\) be the class’s average loss cost across \(n\) observations. Let \(\mu=E[\Theta]\) be the collective mean of the hypothetical class parameter \(\Theta\). The Bühlmann estimate is

Here, \(\mathrm{EPV}=E[\operatorname{Var}(X_j\mid\Theta)]\) is the expected process variance, and \(\mathrm{VHM}=\operatorname{Var}(\Theta)\) is the variance of hypothetical means. This form assumes observations are conditionally independent and identically distributed given \(\Theta\), with a common conditional mean \(\Theta\) and finite variances. It also assumes the observations are comparable in exposure and measurement basis. If observations have unequal exposure, this equal-observation formula should not be applied without adjustment.

What the Blend Means for a Class

The blend is an indicated class loss cost, not necessarily the final premium charged to a policyholder. A rating plan may use other components, adjustments, or constraints to determine a rate. Nor does the blend, by itself, turn class costs into relativities. Those are distinct steps.

The role of the collective estimate is not to declare that all classes have the same expected cost. Rather, it supplies a starting point that can stabilize a class estimate when its observed data do not warrant full reliance. When class experience is more informative under the model, its credibility weight is larger; the collective estimate still supplies the remaining share.

Why the Formula Has This Form

The blend is not an arbitrary compromise. Under the Bühlmann assumptions, it is the best linear predictor of \(\Theta\) based on the class average \(\bar X\), where “best” means minimizing mean squared prediction error among predictors of the form \(a+b\bar X\).

First, conditional independence and a common conditional mean imply \(E[\bar X\mid\Theta]=\Theta\). Therefore \(E[\bar X]=\mu\). By the law of total variance,

Also, using \(E[\bar X\mid\Theta]=\Theta\),

For a linear predictor \(a+b\bar X\), minimizing \(E[(\Theta-a-b\bar X)^2]\) gives the regression slope \(b=\operatorname{Cov}(\Theta,\bar X)/\operatorname{Var}(\bar X)\), and the minimizing intercept is \(a=\mu-b\mu\). Substitution gives

Thus the predictor is \(Z\bar X+(1-Z)\mu\). The result is a best linear predictor under these moment assumptions; it need not equal the exact conditional expectation \(E[\Theta\mid X_1,\ldots,X_n]\) for every possible probability model.

Class Experience Is Not Individual Experience

A class-level calculation summarizes experience across the class observations and estimates the class’s expected loss cost. Individual risk rating asks a different question: how should the experience of one insured affect that insured’s estimate or premium? The observed data, credibility assumptions, and rating rules may differ. In particular, an individual’s claims history should not be inserted in place of a class average just because both are loss-cost observations.

Interpreting a Class Indication Carefully

The credibility estimate should be interpreted as a model-based indicated loss cost. It is not a guarantee that future class losses will equal that amount, nor does it establish that every member of the class has the same expected cost. It summarizes the class experience and its complement under stated assumptions.

The estimate is also sensitive to the meaning of its inputs. A loss cost must be measured consistently across observations, the collective mean must correspond to a suitable complement, and the variance components must represent the population being modeled. The arithmetic can be correct while the indication is not useful if the observations or complement are not appropriate.

Common Exam Traps

• Confusing the target. A class estimate describes the class, not an individual member’s expected cost or premium.
• Reversing the weights. \(Z\) applies to observed class experience; \(1-Z\) applies to the collective estimate.
• Using the wrong average. The data used for \(\bar X\) must be the class observations, not one insured’s history or an unrelated portfolio average.
• Calling the result a final premium. A credibility-weighted loss-cost indication is an input to rating, not necessarily the final charged rate.
• Ignoring assumptions. The displayed \(n/(n+K)\) form is for equal, comparable observations under the stated Bühlmann assumptions. Unequal exposure requires an appropriate treatment rather than automatic use of this formula.

For a class indication, a useful calculation sequence is: identify the class-level observations and collective complement; calculate the observed class average; determine the applicable credibility weight; blend each estimate with its proper weight; and interpret the result at the class level. The distinction between the target and the data is as important as the arithmetic. This treatment is aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus.