Credibility Theory: SOA FAM
Limited-fluctuation credibility gives more weight to observed experience as the exposure approaches a stated full-credibility standard.
- If the standard is specified as 3,600 claims and the subject has 900, square-root partial credibility is 0.5.
- Blending experience cost $120 with a base cost $100 gives $110.
On this page13 sections
- The October syllabus limits the topic
- Square-root partial credibility
- Interpret the blend
- Why the base estimate matters
- Original calculation and common errors
- Compare different exposure levels
- Find the exposure needed for a chosen weight
- Apply the full-credibility cap
- Recover a missing experience estimate
- Keep exposure and estimate units consistent
- Combine groups in the right order
- Prepare comparable inputs before blending
- Solve an original practice question
The October syllabus limits the topic
The October 2026 FAM Topic 4 is Introduction to Credibility, weighted 2.5-5%. Its stated calculation outcome is limited-fluctuation, also called classical, credibility. Do not treat every credibility method found in an older text or broad problem bank as a current required FAM outcome. The SOA research report provides background, but the current syllabus controls examinable scope.
At a high level, credibility blends a subject’s observed experience with a collective or base estimate. The credibility weight Z lies between zero and one: Z=0 gives full weight to the base, Z=1 gives full weight to experience, and intermediate Z values produce a blend. The current learning outcome requires calculation using limited fluctuation; keep the assumptions attached to the example.
Square-root partial credibility
For the specified classical rule, partial credibility can scale with the square root of observed exposure relative to a supplied full-credibility standard: Z=min(1,√(n/n₀)). Here n is exposure count and n₀ is the full-credibility count given under the example’s stated standard. Do not claim that one n₀ is universal. It depends on the confidence, tolerance and model assumptions used to set the standard.
If n₀=3,600 and a subject has n=900 claims, Z=√(900/3,600)=√0.25=0.5. If experience cost is $120 and the base is $100, the blended estimate is 0.5($120)+0.5($100)=$110. This example takes the full-credibility standard as given rather than deriving it from an unstated probability model.
Interpret the blend
A credibility-weighted estimate is Z times the subject experience plus (1−Z) times the base estimate. It moves toward the subject as credibility rises. Check units: both estimates must be comparable costs or rates, not one frequency and one severity. Check Z lies between zero and one. If Z=0.5, the output should be halfway between the two estimates.
Classical limited-fluctuation credibility derives a full-credibility criterion from a chosen tolerance and confidence under a specified process. Since those inputs matter, an illustrative count such as 3,600 cannot be presented as a standard for every insurer, class or exam. Use an explicitly supplied standard or derive only where the problem gives the required assumptions and asks for it.
Why the base estimate matters
The collective base is not an arbitrary number; it represents broader experience or an appropriate complement for the subject. A high individual estimate is blended toward that base when the exposure is limited. If exposure grows, the partial credibility weight rises toward one under the square-root rule. This reduces volatility without pretending the subject data are irrelevant.
In a pricing setting, the final blend is a modeled estimate, not automatically a customer premium. Expenses, profit, trend, rate level and regulatory constraints may enter other stages. Keep credibility estimation separate from the later pricing decision unless the problem explicitly connects them. Topic 4 and Topic 5 are related but distinct in the FAM outline.
Original calculation and common errors
Given n₀=3,600, n=900, experience=120 and base=100, compute Z=0.5 and estimate 110. A common error is using n/n₀ directly as the weight, which would give 0.25. Another is weighting the base by Z instead of 1−Z. Write the blend before substituting: Z·experience+(1−Z)·base.
A third error is deriving n₀ from 3,600 as if that number were universal. Here the problem supplies it. To calculate a threshold, the candidate would need the specified confidence, relative error tolerance and distributional assumptions. The example avoids inventing those inputs. This is the careful scope required by the October 2026 syllabus.
Compare different exposure levels
Keep the same supplied full-credibility standard of 3,600 claims and the same experience and base estimates. At 144 claims, the exposure ratio is 144/3,600=.04, so Z=.20. The blend is.20(120)+.80(100)=104. At 576 claims, the ratio is.16, Z=.40 and the blend is 108. At 1,296 claims, the ratio is.36, Z=.60 and the blend is 112. These are original numerical illustrations, not published insurer rates.
The sequence shows why exposure and credibility do not grow proportionally. Increasing exposure from 144 to 576 multiplies the exposure by four but doubles the weight from.20 to.40. Increasing exposure from 576 to 1,296 multiplies exposure by 2.25 and the weight by 1.5. Work through the square root before blending; treating the exposure ratio as Z would underweight experience in both examples.
Notice another distinction: the blended estimate rises here because the experience estimate of 120 is above the base of 100. More exposure does not inherently mean a higher estimated cost. If subject experience were 80 instead, the same three weights would produce estimates of 96, 92 and 88. Exposure changes the reliance on experience; the difference between experience and the complement determines the direction of the estimate.
Find the exposure needed for a chosen weight
A question can reverse the usual calculation. Suppose the full-credibility standard is supplied as 3,600 claims and the desired partial-credibility factor is.80. Starting from Z=sqrt(n/N), square both sides to obtain Z²=n/N. Therefore n=NZ²=3,600(.80)²=2,304 claims. This algebra applies to the uncapped partial-credibility region. It does not establish a new full-credibility standard.
A useful check is to substitute back: sqrt(2,304/3,600)=sqrt(.64)=.80. A common wrong answer is 2,880 claims, found by multiplying 3,600 by.80 rather than by.64. That exposure would actually produce sqrt(.80), about.8944 credibility. The requested weight and the fraction of the full-credibility exposure are different quantities.
The same rearrangement helps with planning comparisons. Moving from Z=.50 to Z=.75 under an unchanged standard requires exposures proportional to.25 and.5625. The latter is 2.25 times the former. If the initial exposure is 900 claims, the new exposure is 2,025 claims. This describes the specified mathematical rule; it is not a promise that an insurer can obtain those claims or that all added observations are comparable.
Apply the full-credibility cap
Suppose the supplied standard is still 3,600 claims but observed exposure is 4,000 claims. The uncapped square root is sqrt(4,000/3,600), approximately 1.0541. The rule sets Z=min(1,1.0541)=1. With experience cost 120 and a base of 100, the blend is 120. Leaving Z above one would give a negative weight to the complement and an estimate above 120, contrary to this rule.
Full credibility does not mean the subject experience becomes perfectly accurate, free of random variation or immune to future changes. It means the experience meets the selected criterion for full weight under the specified framework. A threshold based on particular assumptions cannot guarantee that the next period will match the estimate. That distinction matters when translating an exam calculation into a business explanation.
At zero exposure, the numerical rule gives Z=0, so the estimate relies on the complement. In an applied setting, an experience estimate based on zero observations may not even be calculable. Do not manufacture an observed average to fill the formula. Explain that the example uses the complement when there is no usable subject experience, subject to the data and modeling instructions supplied.
Recover a missing experience estimate
Another original question gives a credibility-weighted estimate of 118, a complement of 100 and Z=.60, then asks for the subject experience E. Write 118=.60E+.40(100). Subtract 40 from both sides to obtain 78=.60E, then divide by.60: E=130. The unknown is the experience estimate, not the credibility factor.
Check the reconstructed value in the original equation:.60(130)+.40(100)=78+40=118. The blend is between 100 and 130, and it is closer to 130 because the experience receives the larger weight. These two checks help catch both algebra and interpretation errors. Keep the units attached: if all supplied estimates are dollars per claim, the recovered value is also dollars per claim.
The rearranged expression is E=[C-(1-Z)B]/Z, where C is the blend and B is the complement. This expression requires Z to be nonzero. At Z=0, the blend equals B regardless of E, so the blend cannot identify the experience estimate. That is a genuine limitation of the information supplied, rather than an invitation to divide by zero or guess a value.
Keep exposure and estimate units consistent
Suppose a problem defines full credibility as 10,000 policy-years and supplies 2,500 policy-years of subject exposure. The ratio is.25 and Z=.50. If the subject estimate is 12 claims per 100 policy-years and the complement is 10 claims per 100 policy-years, the blend is 11 claims per 100 policy-years. It is a frequency estimate, not $11 and not a claim severity estimate.
Now imagine the subject frequency is instead expressed as.12 claims per policy-year while the complement is still 10 per 100 policy-years. Convert the complement to.10 per policy-year before combining them. The blend becomes.11 per policy-year, equivalent to 11 per 100 policy-years. Mixing.12 with 10 without conversion produces a numerical result with no coherent interpretation.
Do not replace policy-years with a claim count simply because both are counts in a table. They measure different things. A full-credibility criterion based on number of claims needs a compatible observed claim count; one expressed in policy-years needs that exposure basis. If the question supplies a conversion or an alternative standard, follow it. Without that information, the units are a reason to pause the calculation, not to select the larger number.
Combine groups in the right order
Consider two invented groups, each with a supplied full-credibility standard of 3,600 claims and a complement estimate of 100. Group A has 900 claims and experience cost 120. Its factor is.50 and its estimate is 110. Group B has 3,600 claims and experience cost 90. Its factor is 1 and its estimate is 90. Each group has a different credibility calculation even though the complement is the same.
If a task specifically requests a claim-count-weighted average of these two already calculated estimates, use (900×110+3,600×90)/(900+3,600)=423,000/4,500=94. An unweighted average of 110 and 90 would be 100 and would give the smaller group the same contribution as the larger one. The problem must specify a suitable aggregation basis; claim-count weighting is an assumption of this illustration.
This does not mean pooling all observations first and calculating one credibility factor is always equivalent. Pooling changes the subject being estimated and may require a different full-credibility standard, complement or model. Keep the task sequence visible: calculate each specified group, then combine if requested using the stated weights. Do not average credibility factors as a shortcut unless the method explicitly calls for it.
Prepare comparable inputs before blending
Credibility cannot repair a mismatch in what two estimates measure. Suppose subject experience is 120 dollars per claim from an earlier period, while the base estimate is 110 dollars per claim on the target period basis. If an original exercise explicitly instructs you to apply a 5% trend adjustment to subject experience first, the adjusted experience is 126. At Z=.50, the comparable blend is.50(126)+.50(110)=118.
Blending 120 and 110 first gives 115. Multiplying that entire result by 1.05 then gives 120.75, which unnecessarily trends the complement that was already on the target basis. The difference comes from applying an adjustment to the wrong input. The 5% figure here is an invented problem assumption, not a forecast or a recommended actuarial trend.
A similar problem arises when experience reflects one coverage limit and the complement another. Decide whether the question provides a required adjustment, already normalized estimates or an explicit simplification. Do not invent a correction factor. The useful habit is to identify the target quantity and basis before assigning weights. A technically correct square root can still produce an unsuitable estimate if the input costs are not comparable.
Solve an original practice question
An original question supplies a full-credibility standard of 2,500 claims, actual exposure of 400 claims, subject experience cost of 150 dollars per claim and a complement of 125 dollars per claim. It asks for the classical square-root partial credibility estimate. First calculate the exposure ratio: 400/2,500=.16. Then take the square root to obtain Z=.40. The complement receives.60 weight.
The answer is.40(150)+.60(125)=60+75=135 dollars per claim. It lies between 125 and 150. The difference between the two inputs is 25, and the estimate is.40×25=10 above the complement, providing an independent arithmetic check. An answer of 129 would result from using.16 as the weight rather than its square root.
To explain the answer in words, the specified method assigns 40% weight to this subject’s experience and 60% to the complement because exposure remains below full credibility. Do not describe 40% as the probability that 150 is correct. Do not call 135 the final charged premium: expense provisions, other adjustments and pricing decisions have not been supplied. The calculation produces the estimate requested by this example.
Common questions
Which credibility method is in current FAM scope?
Topic 4 names limited-fluctuation or classical credibility.
Is 3,600 a universal full-credibility count?
No. It is supplied in the example; a standard depends on assumptions and tolerance/confidence.
How do I blend estimates?
Use Z times subject experience plus (1−Z) times the base estimate.