Why Expected Claims Are the Relevant Exposure Measure
Exposure measures how much insurance is in force: for example, policy-years, vehicle-years, or insured-unit months. But a unit of exposure does not always represent the same expected claim activity. If one portfolio is expected to produce 0.2 claims per policy-year and another 1.0 claim per policy-year, equal numbers of policy-years do not imply equal expected claim counts. For limited-fluctuation credibility, the relevant quantity is the expected number of claims.
The conversion is straightforward once the units are aligned: multiply exposure by the expected claim frequency per exposure unit. The resulting expected count can then be compared with a full-credibility standard, expressed in expected claims. This distinction matters whenever the expected frequency is not one claim per exposure unit, or when the portfolio contains groups with different frequencies.
From Expected Claims to Credibility
Suppose a full-credibility standard requires the observed claim count to estimate its expected value within a specified relative error with a specified probability. Under the usual limited-fluctuation claim-count model, claim counts are Poisson: if the expected count is \(\mu\), then \(N\sim\operatorname{Poisson}(\mu)\), so \(E[N]=\operatorname{Var}(N)=\mu\). The full-credibility standard can be expressed as a threshold \(N_0\) in expected claims. A portfolio with \(\mu\geq N_0\) meets that standard; one with fewer expected claims receives partial credibility.
When the standard is specified by a relative error bound \(k\) and probability \(p\), a normal approximation to the Poisson count gives the familiar threshold \(N_0=(z/k)^2\), where \(z\) is the standard normal quantile associated with the probability requirement. For a central probability \(p\), \(z=\Phi^{-1}((1+p)/2)\). If a problem supplies \(N_0\) directly, use that standard rather than recalculating it.
Here \(Z\) is the limited-fluctuation credibility assigned to the experience. The square-root rule reflects how random error changes with the expected count; it does not mean that the exposure itself is square-rooted. When groups have different frequencies, calculate \(\mu\) by summing their expected claims before applying the rule.
Why the Square-Root Relationship Appears
The link between expected claims and credibility follows from the variation of a Poisson count. For \(N\sim\operatorname{Poisson}(\mu)\), the standard deviation is \(\sqrt{\mu}\), so the count's standard deviation relative to its mean is \(1/\sqrt{\mu}\). Increasing the expected count reduces relative random fluctuation.
For a sufficiently large Poisson mean, \(N\) is approximately normal with mean \(\mu\) and standard deviation \(\sqrt{\mu}\).
The event \(|N-\mu|\leq k\mu\) is, after standardizing, approximately \(|(N-\mu)/\sqrt{\mu}|\leq k\sqrt{\mu}\).
For a central probability \(p\), the standardized bound must be at least \(z=\Phi^{-1}((1+p)/2)\). Thus \(k\sqrt{\mu}\geq z\), or \(\mu\geq(z/k)^2=N_0\).
For partial credibility, the same standard deviation scaling gives a useful comparison. Relative fluctuation is proportional to \(1/\sqrt{\mu}\), while the full standard corresponds to \(1/\sqrt{N_0}\). The ratio of the full-standard fluctuation to the portfolio's fluctuation is therefore \(\sqrt{\mu/N_0}\). This is the partial credibility factor when \(\mu<N_0\); once the portfolio reaches the threshold, credibility is capped at 1. The derivation uses the Poisson variance assumption and a normal approximation for the full standard, so it is not a substitute for a different variance model when counts are overdispersed.
Worked Applications
Worked Example: Converting Policy-Years to Expected Claims
A hypothetical insurer has 240 policies, each in force for six months. The expected claim frequency is 0.80 claims per policy-year. The full-credibility standard is \(N_0=(1.645/0.10)^2\) expected claims. Find the exposure, expected claim count, and partial credibility.
Model and assumptions. Assume independent Poisson claim counts with a common frequency of 0.80 claims per policy-year. The exposure unit is one policy-year. No interest or mortality assumption applies to this claim-count calculation.
Method: convert to policy-years, then use expected claims. Six months is \(6/12=0.5\) policy-year per policy. Therefore, total exposure is \(E=240(0.5)=120\) policy-years.
The expected number of claims is \(\mu=E\lambda=120(0.80)=96\) claims.
Calculate the threshold: \(N_0=(1.645/0.10)^2=16.45^2=270.6025\) expected claims. Since \(96<270.6025\), the experience is below the full-credibility standard. Its credibility is \(Z=\sqrt{96/270.6025}=\sqrt{0.3548}=0.5956\), rounded to four decimals.
Check. The portfolio has fewer expected claims than the full standard, so the calculated credibility must be below 1. The value 0.5956 satisfies that condition. The common mistake avoided here is treating the 240 policies as 240 policy-years and overlooking their six-month duration.
Worked Example: Combining Groups With Different Frequencies
A hypothetical coverage program has 80 policy-years in a low-frequency class with an expected frequency of 0.50 claims per policy-year, and 150 policy-years in a higher-frequency class with an expected frequency of 1.20 claims per policy-year. The full-credibility standard is 270.6025 expected claims. Find the overall partial credibility.
Model and assumptions. Assume independent Poisson claim counts in the two classes, with the stated constant frequency within each class. Both exposures are measured in policy-years. No interest or mortality assumption applies.
Method: calculate expected claims separately, then add. The first class contributes \(\mu_1=80(0.50)=40\) expected claims. The second contributes \(\mu_2=150(1.20)=180\) expected claims. Hence the total is \(\mu=\mu_1+\mu_2=40+180=220\) expected claims.
Because \(220<270.6025\), credibility is partial: \(Z=\sqrt{220/270.6025}=\sqrt{0.8130}=0.9017\), rounded to four decimals.
Check. The total expected count lies below the standard, and the resulting credibility is less than 1. The common mistake avoided is multiplying total exposure, \(80+150=230\) policy-years, by one arbitrarily selected frequency. The two groups have different expected claim rates, so their expected counts must be calculated separately.
Worked Example: Matching Vehicle-Months to an Annual Frequency
A hypothetical fleet has 400 insured vehicles, each covered for 18 months. Its expected frequency is 0.12 claims per vehicle-year. Use a full-credibility standard of 270.6025 expected claims to determine the exposure, expected count, and credibility.
Model and assumptions. Assume independent Poisson claim counts with a common frequency of 0.12 claims per vehicle-year. No interest or mortality assumption applies.
Method: convert the duration before multiplying by frequency. Eighteen months is \(18/12=1.5\) years per vehicle. Thus, \(E=400(1.5)=600\) vehicle-years.
The annual frequency is already stated per vehicle-year, so \(\mu=600(0.12)=72\) expected claims.
Since \(72<270.6025\), the credibility factor is \(Z=\sqrt{72/270.6025}=\sqrt{0.2661}=0.5158\), rounded to four decimals.
Check. A frequency of 0.12 claims per year across 600 vehicle-years implies 72 expected claims, which is below the full standard; a partial factor is therefore appropriate. The common mistake avoided is multiplying 400 vehicles by the annual rate while ignoring the 18-month coverage period.
Common Mistakes and Practical Checks
Before calculating credibility, write down the exposure unit and the frequency unit beside the numbers. Multiplying policy-months directly by claims per policy-year is not dimensionally consistent; convert policy-months to policy-years first. Likewise, a frequency stated per vehicle-year cannot be multiplied by exposure measured in policy-years unless those units represent the same insured exposure.
- Using exposure as if it were a claim count. Exposure of 200 policy-years does not imply 200 expected claims. The expected frequency must also be applied.
- Averaging frequencies without appropriate weights. With different groups, use \(\sum_j E_j\lambda_j\). A simple average of group frequencies generally gives the wrong expected count.
- Applying the square root to the wrong quantity. The partial factor is \(\sqrt{\mu/N_0}\), where \(\mu\) is expected claims—not \(\sqrt{E/N_0}\) unless the frequency happens to be one per exposure unit.
- Forgetting the cap. If \(\mu>N_0\), the uncapped square root exceeds 1, but limited-fluctuation credibility is capped at 1.
- Using the Poisson rule when its variance assumption is unsuitable. The simple expected-count standard relies on a Poisson claim-count variance equal to the mean. If a problem specifies a different count model or variance, use the corresponding standard rather than assuming the Poisson formula applies.
Check Your Understanding
For each problem, use the stated full-credibility standard of 270.6025 expected claims unless another standard is given. Round credibility factors to four decimals.
- A hypothetical book has 150 policy-years and expected frequency 0.60 claims per policy-year. Find expected claims and partial credibility. (Answer: 90 expected claims; \(Z=0.5767\).)
- A hypothetical insurer covers 200 policies for nine months each. Expected frequency is 0.80 claims per policy-year. Find the expected claim count and credibility. (Answer: 120 expected claims; \(Z=0.6660\).)
- A hypothetical portfolio has 50 policy-years at frequency 0.40 and 100 policy-years at frequency 1.00, with independent Poisson counts. Find total expected claims and credibility. (Answer: 120 expected claims; \(Z=0.6660\).)
- A hypothetical fleet has 900 vehicle-years of exposure and expected frequency 0.30 claims per vehicle-year. Does it meet the full standard? If not, find its credibility. (Answer: 270 expected claims; it does not meet the standard; \(Z=0.9989\).)
- A hypothetical program has 300 policy-years at frequency 0.50 and 200 policy-years at frequency 0.80. Find the total expected claim count and determine whether credibility is partial or full. (Answer: 310 expected claims; full credibility, \(Z=1\).)