Why Blend Experience with a Complement?
A class’s observed frequency, loss ratio, or pure premium can be informative, but its reliability depends in part on how much experience produced it. A small class may have an unusually high or low observed result simply because of random variation. A credibility-weighted estimate balances that class result against a complement, such as a broader portfolio result or an actuarially selected expected value.
The blend provides a direct way to express how much weight is assigned to each source. It does not, by itself, determine whether the selected complement is appropriate or how the credibility weight should be calculated. Those choices depend on the measure, data, and credibility procedure being used.
For example, if \(Z=0.30\), the estimate assigns 30% weight to the observed statistic and 70% to the complement. A credibility weight is not a probability that the class experience is correct. It is a weight in the specified estimation rule.
Conditions for a Meaningful Blend
The two inputs must measure the same quantity in compatible units. A class frequency per policy-year cannot be blended directly with a complement stated as claims per 1,000 vehicle-years unless one is first converted to the other basis. Likewise, a loss ratio based on one coverage or period should not be blended with a complement based on a different coverage or period without a justified adjustment.
The linear blend also presumes that the chosen weight is appropriate for the method and purpose. It is an estimation rule, not a guarantee that the result is unbiased or more accurate than either input. If the complement is poorly selected or the experience is not comparable, a mathematically correct calculation can still produce a poor estimate.
| Weight | Interpretation | Result |
|---|---|---|
| \(Z=0\) | No weight on observed experience | \(\hat{T}=C\) |
| \(0<Z<1\) | Weight on both inputs | \(\hat{T}\) lies between \(T\) and \(C\) |
| \(Z=1\) | All weight on observed experience | \(\hat{T}=T\) |
When \(T\) and \(C\) are on the same scale and \(Z\) is between zero and one, the estimate cannot lie outside the interval between them. This is a useful arithmetic check. If the observed statistic exceeds the complement, the estimate is between the two; if it is lower, the estimate is also between the two.
Worked Example: Blending an Observed Claim Frequency
A hypothetical group of small commercial equipment policies has an observed claim frequency of 0.084 claims per policy-year. An actuarial complement is 0.060 claims per policy-year, and the assigned credibility weight is \(Z=0.40\). Calculate the credibility-weighted frequency. State the method and check the result.
Method and inputs. Use the linear credibility blend. Both inputs are frequencies per policy-year, so they are on the same basis. The complement receives weight \(1-Z=1-0.40=0.60\).
Substitute.
The observed-experience contribution is \(0.40(0.084)=0.0336\). The complement contribution is \(0.60(0.060)=0.0360\). Therefore,
The estimated frequency is 0.0696 claims per policy-year. The result is between 0.060 and 0.084, as a blend with a weight between zero and one must be. A common mistake is to apply \(Z\) to the complement instead of the observed statistic. That would give \(0.40(0.060)+0.60(0.084)=0.0744\), reversing the assigned shares; the stated rule gives 0.0696.
How the Blend Responds to the Inputs
The blend can be rewritten to show how far the estimate moves from the complement:
This form follows by expanding the definition: \(ZT+(1-Z)C=ZT+C-ZC=C+Z(T-C)\). The quantity \(T-C\) is the difference between the observed statistic and the complement. The weight \(Z\) takes that difference in full when \(Z=1\), in part when \(0<Z<1\), and not at all when \(Z=0\).
The direction of movement depends on which input is larger. If \(T>C\), increasing \(Z\) raises the estimate toward \(T\). If \(T<C\), increasing \(Z\) lowers the estimate toward \(T\). If \(T=C\), the blend equals that common value for every possible weight.
This statement holds with the two input values fixed. In actual experience, an increase in volume may also change the observed statistic \(T\). Therefore, a higher credibility weight does not guarantee that estimates calculated at successive dates will rise or fall: the observed experience itself may have changed.
Connecting the Weight to Experience Volume
In limited fluctuation credibility, a common partial-credibility rule assigns a weight based on the ratio of observed volume to the volume required for full credibility. For an exposure-based illustration, let \(n\) be the class’s exposure volume and \(N_0\) the full-credibility standard, both measured in the same exposure units. The capped weight is
This rule is used under the assumptions underlying the relevant limited fluctuation standard; it is not a universal weight formula for every credibility model. Before full credibility, it gives partial weight, and once \(n\) reaches or exceeds \(N_0\), the cap sets \(Z=1\). The particular standard must be suitable for the measure being estimated.
The square-root relationship means that multiplying volume by four doubles the uncapped weight, rather than quadrupling it. For instance, if \(n/N_0=0.25\), then \(Z=\sqrt{0.25}=0.50\). If volume rises to \(n/N_0=1\), the weight becomes 1. The blend then uses the observed statistic alone, subject to the credibility procedure and assumptions that established the full-credibility standard.
Specify whether the class statistic is a frequency, loss ratio, pure premium, or another measure.
Check coverage, time period, exposure unit, and other relevant definitions for \(T\) and \(C\).
Use the stated credibility procedure and respect any cap at full credibility.
Calculate \(ZT+(1-Z)C\), verify the weights sum to one, and confirm the result lies between the inputs when \(0 \leq Z \leq 1\).
Worked Example: How More Exposure Changes a Pure Premium Estimate
A hypothetical specialty repair class has an observed pure premium of \$520 per exposure unit and a complement of \$460 per exposure unit. For this illustration, assume the applicable limited fluctuation standard is \(N_0=1{,}600\) exposure units, and use \(Z=\min(1,\sqrt{n/N_0})\). Calculate the estimate at 400, 900, and 1,600 exposure units, holding the observed pure premium and complement fixed to isolate the effect of the weight.
Method and assumption. Apply the stated capped square-root rule, then blend the resulting weight with \(T=\$520\) and \(C=\$460\). The exposure units, observed pure premium, and complement are assumed comparable.
At \(n=400\). The uncapped weight is \(\sqrt{400/1{,}600}=\sqrt{0.25}=0.50\), below the cap. The complement weight is \(1-0.50=0.50\). Thus,
At \(n=900\). The weight is \(\sqrt{900/1{,}600}=\sqrt{0.5625}=0.75\). The complement weight is \(0.25\). Therefore,
At \(n=1{,}600\). The uncapped weight is \(\sqrt{1{,}600/1{,}600}=1\), so the cap leaves \(Z=1\). The complement weight is zero, and
The estimates, in order of increasing volume, are \$490, \$505, and \$520 per exposure unit. They move toward the observed value because that value exceeds the complement and the credibility weight increases. A common mistake is to use \(n/N_0\) as the weight: at 400 exposures that gives 0.25 rather than the correct square-root weight of 0.50 under the stated rule. The final results are between the two inputs and reach the observed value at the full-credibility standard.
Worked Blends for a Loss Ratio
The same algebra applies to other measures, provided the observed statistic and complement are comparable. Percentages must be used consistently: 72% may be entered as 0.72 or as 72 percentage points, but the same representation must be used for both inputs.
Worked Example: Blending a Class Loss Ratio
A hypothetical inland marine class has an observed loss ratio of 72%. A selected complement loss ratio is 60%, and the credibility assigned to the class experience is \(Z=0.25\). Calculate the estimate and explain why it is closer to the complement.
Method and inputs. Use the linear blend with both loss ratios expressed as percentages. The complement receives \(1-0.25=0.75\) of the weight.
The observed contribution is \(18\) percentage points, and the complement contribution is \(45\) percentage points. Their sum is \(63\) percentage points:
The estimated loss ratio is 63%, which is closer to the 60% complement because only 25% of the blend weight is assigned to the observed result. As a check, the estimate lies between 60% and 72%. A common mistake is to report \(0.63\%\) after mixing decimal and percentage notation; the calculation represents 0.63 as a ratio, or 63% when stated as a percentage.
Common Errors and Interpretation
- Reversing the weights. In the defined convention, \(Z\) multiplies the observed statistic \(T\), and \(1-Z\) multiplies the complement \(C\).
- Using incompatible quantities. Confirm that both inputs use the same units, period, coverage, and definition before blending.
- Treating \(Z\) as a probability. It is the assigned share of the estimate, not the chance that the class’s observed result is correct.
- Forgetting the complement weight. The two weights must add to one: \(Z+(1-Z)=1\).
- Assuming volume alone determines the direction of change. With fixed \(T\) and \(C\), the estimate moves toward \(T\) as \(Z\) rises. If \(T\) changes as new experience arrives, the resulting estimate can move either way.
- Ignoring the full-credibility cap. Under a capped rule, a calculated weight above one is set to one; it is not used to extrapolate beyond the observed statistic.
Check Your Understanding
For each problem, calculate the requested blend and check its relationship to the observed statistic and complement.
- A hypothetical transit-property class has an observed frequency of 0.075 claims per exposure-year and a complement of 0.055 claims per exposure-year. With \(Z=0.60\), calculate the blended frequency. (Answer: 0.067 claims per exposure-year)
- A hypothetical crop-equipment segment has an observed pure premium of \$840 per policy and a complement of \$780 per policy. If \(Z=0.35\), calculate the estimate and identify which input receives the larger weight. (Answer: \$801 per policy; the complement receives the larger weight, 0.65)
- For a hypothetical class, \(T=\$430\), \(C=\$390\), and the full-credibility standard is 2,500 exposure units. Use \(Z=\min(1,\sqrt{n/N_0})\) at \(n=1{,}000\). Calculate the weight and blended estimate, rounding the estimate to cents. (Answer: \(Z\approx0.6325\); estimate \(\$415.30\))
- A hypothetical class has fixed observed frequency \(T=0.12\) and complement \(C=0.09\). Its full-credibility standard is 900 exposure units. Under the capped square-root rule, calculate the estimate at 100, 400, and 900 exposure units. (Answer: \(Z=1/3, 2/3, 1\); estimates \(0.10, 0.11, 0.12\) claims per exposure unit)
- A hypothetical class’s estimate is first calculated with \(T=0.14\), \(C=0.10\), and \(Z=0.50\). At a later review, the values are \(T=0.11\), \(C=0.10\), and \(Z=0.80\). Calculate both estimates and explain why the later estimate is lower despite the higher weight. (Answer: first estimate \(0.12\); later estimate \(0.108\); the observed statistic fell enough to move the blend downward.)