Selecting A Relevant Credibility Complement

Why the Complement Matters

A credibility-weighted estimate combines experience from a segment with a complementary estimate. The segment experience reflects the risks actually observed, but may fluctuate substantially when exposure or claim counts are limited. A complement can stabilize the estimate—but only if it represents a sensible benchmark for the segment. A large data set from a poorly matched group can be less useful than a smaller, more comparable one.

Selecting a complement is therefore a question of comparability, not simply size. Consider whether the candidate represents similar insureds and hazards, provides the same coverage and claim measure, uses a consistent exposure base, and reflects a period that can be aligned with the target. These checks support a defensible complement; they do not, by themselves, determine the credibility weight.

Definition: A credibility complement is the benchmark estimate combined with a segment’s own experience in a credibility procedure. A candidate complement is relevant to the extent that its underlying risks, coverage, measurement basis, and experience period are appropriate for estimating the target segment. Relevance does not mean the candidate is identical to the segment, nor does a larger volume of data prove that it is suitable.

For examples in this tutorial, use the following credibility-weighted estimate when a credibility weight $Z$ is supplied:

\widehat{\theta}=Z\widehat{\theta}_{\text{segment}}+(1-Z)\widehat{\theta}_{\text{complement}}, \qquad 0\leq Z\leq 1.

Here, both estimates must refer to the same quantity and be expressed on a comparable basis. The weight is treated as given; choosing a relevant complement is a separate task from determining $Z$. If the complement cannot reasonably be aligned to the target, applying the blend mechanically does not fix the mismatch.

A Practical Comparability Review

Start by defining the target estimate precisely. For example, a claim frequency per earned vehicle-year is not the same quantity as claim frequency per policy, and an average paid severity is not interchangeable with an incurred loss cost. The earlier material on defining claim measures and aligning experience periods is useful here: a candidate must measure the intended quantity over a period that can be compared with the target.

Then review each candidate along several dimensions. The dimensions are related, but a mismatch in one may be more consequential than a mismatch in another.

Dimension	Questions to ask	Possible response
Risk characteristics	Are the insureds, hazards, and important rating characteristics similar?	Use a closer subgroup or standardize for material mix differences.
Coverage and terms	Are the covered causes of loss, limits, deductibles, and claim rules comparable?	Adjust only when a defensible adjustment is available; otherwise seek a better match.
Claim measure	Do both estimates use the same measure, such as frequency, severity, or loss cost?	Reconstruct the measure on a consistent basis before comparison.
Exposure basis	Does each rate use a meaningful, consistent exposure unit?	Convert or recalculate the rate using the target’s exposure definition.
Measurement period	Do the periods reflect comparable conditions and development?	Align periods or trend the measure when appropriate.
Geography and operations	Are hazard conditions, legal environments, and operating practices similar?	Prefer a comparable region or account for known differences.

The review is not a checklist in which every difference can be corrected by a simple multiplier. Some differences may be material but not measurable with available information. In that case, the analyst should recognize the limitation rather than imply that the candidate is fully comparable.

Define the target.
Specify the segment, estimate, claim measure, exposure base, coverage, and period to be represented.

Screen candidates for substantive fit.
Compare risk characteristics, coverage terms, geography, and operations before focusing on the candidate’s volume.

Check measurement alignment.
Confirm that claim definitions, exposure units, and periods can be made comparable.

Adjust what can be supported.
Standardize for a measurable mix difference or align the period using an appropriate stated assumption.

Use the complement transparently.
Apply the supplied credibility weight to estimates on a common basis and state any remaining limitations.

Why Standardizing Risk Mix Works

Suppose a benchmark contains two risk groups, indexed by $j$, with rates $\theta_j$. If the target has proportions $w_j$ in those groups, its expected rate under the benchmark’s group-specific rates is the target-weighted average. The weights must describe the target mix, not the benchmark’s original mix.

\theta_{\text{target mix}}=\sum_j w_j\theta_j, \qquad \sum_j w_j=1.

To see why, imagine $E$ units of target exposure. Group $j$ contributes $Ew_j$ exposure and therefore has expected claims $Ew_j\theta_j$, when $\theta_j$ is a rate per exposure unit. Summing expected claims gives $E\sum_jw_j\theta_j$. Dividing by the total exposure $E$ yields $\sum_jw_j\theta_j$. This standardization is appropriate only when the group definitions and group-specific rates are comparable to the target. It does not correct unmeasured differences within groups.

Worked Examples

Worked Example: Choosing Among Three Fleet Complements

A hypothetical insurer is estimating claim frequency for a segment of small electric-delivery fleets in 2025. The segment has 600 earned vehicle-years and 42 claims. Three candidate complements are available: (A) the insurer’s entire commercial-auto portfolio, with 150 claims over 3,000 vehicle-years, including long-haul trucks; (B) a broader territory’s small electric-delivery fleets, with 72 claims over 1,200 vehicle-years; and (C) small electric-delivery fleets in a neighboring territory, with 54 claims over 900 vehicle-years. All candidates use the same claim definition, coverage terms, exposure unit, and 2025 period. The neighboring territory is known to have similar operating conditions. Select a complement and calculate the blended frequency if $Z=0.40$.

Method. First compare the candidate rates and their fit to the target. The segment’s observed frequency is:

\widehat{\theta}_{\text{segment}}=\frac{42}{600}=0.0700 \quad\text{claims per vehicle-year.}

Candidate A has frequency $150/3{,}000=0.0500$. Its data volume is greatest, but its inclusion of long-haul trucks makes its risk mix less comparable. Candidate B has frequency $72/1{,}200=0.0600$. Candidate C has frequency $54/900=0.0600$ and represents the same fleet type in a neighboring territory with similar operating conditions.

B and C have the same observed rate, but C has the closer stated geographic and operating match. Choose C as the complement. The blend is:

\begin{aligned} \widehat{\theta} &=0.40(0.0700)+0.60(0.0600)\\ &=0.0280+0.0360\\ &=0.0640\text{ claims per vehicle-year.} \end{aligned}

Check. Since the result is a weighted average of 0.0700 and 0.0600, it must lie between them; 0.0640 does. The wrong approach is to select A merely because it has the most data. Its broader portfolio risk mix is less relevant, and its lower frequency would pull the estimate down for a reason unrelated to the target’s own experience. The calculation avoids that mistake by selecting on comparability first.

Worked Example: Aligning a Complement to the Target Period

A hypothetical specialty-equipment segment has a 2024 loss cost of $1,100 per exposure unit. A candidate complement consists of similar equipment and coverage, but its measured 2022 loss cost is $1,000 per exposure unit. Assume the loss cost increases by 3% per year, compounded annually, and that this trend applies to the complement’s measure. If $Z=0.30$, calculate the estimate on a 2024 basis.

Method. The complement matches the target in risk and coverage, but its period is two years earlier. Apply two annual trend factors to convert the 2022 complement to 2024:

\widehat{\theta}_{\text{complement, 2024}} =1{,}000(1.03)^2 =1{,}000(1.0609) =\$1{,}060.90.

Both estimates now refer to 2024. Blend them using the supplied weight:

\begin{aligned} \widehat{\theta} &=0.30(\$1{,}100)+0.70(\$1{,}060.90)\\ &=\$330.00+\$742.63\\ &=\$1{,}072.63\text{ per exposure unit.} \end{aligned}

Check. The result lies between $1,060.90 and $1,100, as required for a weighted average. Blending the untrended $1,000 directly would produce $0.30(1{,}100)+0.70(1{,}000)=\$1{,}030$, which compares different periods and understates the estimate under the stated upward-trend assumption. The trend adjustment is justified here because the problem explicitly supplies a common annual trend; in practice, the appropriateness of the trend assumption must be assessed.

Worked Example: Standardizing a Territory Complement to the Target Mix

A hypothetical target segment contains two risk groups, standard and elevated hazard. Its own frequency is 0.060 claims per exposure unit. A territory complement reports group-specific frequencies of 0.020 for standard risks and 0.080 for elevated risks. The territory’s exposure mix is 80% standard and 20% elevated, while the target’s mix is 50% in each group. All figures use the same period, claim definition, coverage, and exposure unit. Calculate the mix-adjusted complement and the blend for $Z=0.25$.

Method. The territory’s unadjusted aggregate frequency is:

$$ 0.80(0.020)+0.20(0.080)=0.016+0.016=0.032. $$

That aggregate reflects the territory’s 80%/20% mix, not the target’s 50%/50% mix. Standardize using the target weights:

\widehat{\theta}_{\text{complement, target mix}} =0.50(0.020)+0.50(0.080) =0.010+0.040 =0.050.

Now blend this target-mix complement with the segment frequency:

\begin{aligned} \widehat{\theta} &=0.25(0.060)+0.75(0.050)\\ &=0.015+0.0375\\ &=0.0525\text{ claims per exposure unit.} \end{aligned}

Check. The standardized complement is higher than the territory’s raw 0.032 because the target has a larger share of elevated risks. The final estimate, 0.0525, lies between 0.050 and 0.060. A common mistake is to use 0.032 as the complement and report $0.25(0.060)+0.75(0.032)=0.039$. That calculation blends on the wrong exposure mix and therefore answers a different question. Standardization is justified here because the group-specific rates and target group weights are supplied.

Common Traps and Practical Limits

A candidate complement can look attractive for the wrong reason. Keep the distinction between the quality of the benchmark and the quantity of data clear: volume may affect the stability of a candidate’s estimate, but it does not make its risks more like the target. Similarly, a matching rate value does not prove that two populations are comparable; different risk mixes can produce the same aggregate rate.

Choosing the largest source automatically. Start with risk and measurement fit. Consider size only after identifying candidates that are relevant.
Blending rates from different periods. Align periods first when a supported adjustment is available. Do not silently assume that an older rate is current.
Comparing different measures. A frequency, severity, and loss cost are not interchangeable complements. Confirm the claim definition and exposure denominator.
Using the complement’s mix instead of the target’s mix. When standardizing group rates to the target, weight them using the target proportions.
Assuming every difference can be adjusted. If an important difference cannot be measured credibly, state the limitation and consider another candidate.

Key takeaway: Select a complement because it represents the target on relevant risk and measurement dimensions, not merely because it is available or large. Align periods and exposure mix when the necessary data and assumptions support an adjustment; then blend estimates expressed on the same basis.

Check Your Understanding

For each problem, identify the relevant complement or adjustment before calculating.

A hypothetical target segment has 36 claims over 600 exposure-years. Candidate A has 90 claims over 1,500 exposure-years but includes several unrelated vehicle classes. Candidate B has 48 claims over 800 exposure-years for the same vehicle class and period. Choose the more relevant candidate and calculate the blend for $Z=0.50$. (Answer: Candidate B; $0.060$ claims per exposure-year.)
A target’s 2024 loss cost is $950. A comparable complement’s 2022 loss cost is $800, and the applicable annual compounded trend is 4%. Calculate the adjusted complement and blend for $Z=0.25$. (Answer: Adjusted complement = $865.28; blend = $886.46 per exposure unit.)
A target is 40% low hazard and 60% high hazard. A comparable complement has group frequencies of 0.010 and 0.050, respectively, but its own mix is 75% low hazard and 25% high hazard. Calculate its raw aggregate and its target-mix standardized frequency. (Answer: Raw aggregate = 0.020; standardized frequency = 0.034 claims per exposure unit.)
A segment estimate is 0.070 claims per policy-year. A candidate complement reports 0.060 claims per vehicle-year for similar risks, in the same period and with similar coverage. Can these rates be blended directly? State the needed step. (Answer: No; first express both estimates on a consistent, defensible exposure basis, or use a complement already measured per policy-year.)
A target has a 2025 frequency of 0.050 per exposure unit. Two otherwise comparable complements each report 0.040: one is from 2025 and the other from 2023. Assume an applicable compounded frequency trend of 5% per year for the older complement. Calculate the 2023 complement on a 2025 basis and state which is directly time-aligned. (Answer: Older complement adjusted to 0.0441; the 2025 complement is directly time-aligned.)

What You'll Learn