Credibility As a Weight on Subject Experience

Why Give a Risk’s Experience a Weight?

A risk’s own claim history is relevant to its expected future claims, but it may not tell the whole story. A short history can be unusually high or low because of random variation. A collective estimate, based on a broader group of comparable risks, can provide a useful reference. Credibility describes how much weight an estimate gives to the subject risk’s own experience when combining it with that collective view.

The key distinction is between what happened to the subject and what is expected for the subject going forward. The observed record is evidence, not automatically the final estimate. A credibility-weighted estimate uses the subject’s experience while also accounting for the collective benchmark. The two inputs must describe the same quantity, on the same basis and time period, before they are combined.

The value \(Z\) is a weight, not a claim count, an exposure measure, or the probability that a particular future claim will occur. When \(Z=0\), the estimate uses only the collective benchmark. When \(Z=1\), it uses only the subject experience. Between those endpoints, the estimate lies between the two inputs. The credibility value may be supplied by a rating method or by an assumption in a problem; the weighted calculation itself does not determine how \(Z\) was selected.

The estimate also needs a clear unit. If the quantity is claim frequency, both inputs must be claims per exposure. If it is loss cost, both must be dollars per exposure. Combining a subject’s total dollars with a collective dollars-per-exposure estimate would be meaningless. A useful check is to write the unit beside each input before calculating.

What the Weighted Estimate Means

The formula is a weighted average, so it does not claim that the subject’s observed experience will repeat exactly. Instead, it blends that experience with the collective view. For example, if a subject’s loss cost is higher than the collective estimate, a positive credibility weight moves the estimate above the collective amount, but it does not move it all the way to the subject amount unless \(Z=1\).

This interpretation can be justified directly. Suppose an estimate \(m\) is judged by its weighted squared distance from the two available measures: the subject experience \(E\) and collective estimate \(C\). Assign relative weights \(Z\) and \(1-Z\) to those squared distances. The quantity to minimize is

Differentiate with respect to \(m\), set the derivative equal to zero, and solve:

The second derivative is \(2Z+2(1-Z)=2>0\), so this stationary point minimizes \(Q(m)\). This is a justification for the weighted-average form, not a rule for choosing \(Z\): the problem or a credibility method must supply or determine the weight. The interpretation also presumes that \(E\) and \(C\) measure the same thing; otherwise their squared distances and weighted average are not meaningful.

Applying Credibility to Claim Measures

For frequency, the subject experience is typically the number of claims divided by the relevant exposure. For loss cost, it is typically aggregate losses divided by exposure. The collective estimate must use the corresponding measure and exposure basis. These calculations produce a rate-like estimate per exposure; multiplying by future exposure converts it to an expected total for that exposure volume.

Combining Experience Across Periods

Before applying credibility, experience gathered over multiple periods may need to be summarized. When periods have different exposure volumes, dividing each period’s loss by its exposure and then taking a simple average gives equal influence to periods of unequal size. If the intended subject measure is loss per exposure across all periods, combine the losses and exposures first. This creates an exposure-weighted experience measure.

Common Mistakes and Practical Checks

Credibility is simple to calculate once the inputs are prepared, but errors often occur before the multiplication. Use a consistent sequence: identify the quantity being estimated, calculate subject experience in the right units, confirm that the collective estimate has matching units, apply the stated credibility, and interpret the result as an estimate rather than as observed claims.

• Using the wrong weight. In the weighted estimate, \(Z\) multiplies subject experience. A larger \(Z\) therefore pulls the result closer to the subject’s measure.
• Skipping exposure normalization. A total loss amount is not a loss cost. Divide by the relevant exposure before blending with a per-exposure collective estimate.
• Averaging unequal periods equally. If the target is loss per exposure across periods, pool losses and exposures rather than taking a simple average of period rates.
• Mixing incompatible bases. Frequency cannot be blended directly with loss cost, and amounts from different coverage or valuation bases should not be combined without adjustment.
• Calling the estimate observed experience. The weighted result is a prospective estimate that uses observed experience; it is not the raw historical record.

A useful reasonableness check is that for \(0\leq Z\leq1\), the weighted estimate must fall between its two input values. If it falls outside that interval, check the arithmetic, the weights, and whether the inputs were stated on a consistent basis. The calculations here are aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus.