Scaling Credibility Standards Across Risk Segments

Why Credibility Standards Scale Differently Across Segments

Two insurance segments can be subject to the same confidence probability and relative tolerance but need different amounts of exposure to meet a limited-fluctuation standard. A segment with fewer claims per policy-year generally needs more policy-years. For aggregate losses, variation in individual claim sizes also matters: more variable severities increase the required exposure, even when claim frequency is unchanged.

The useful comparison is not just “which segment is larger?” It is how each segment’s frequency and severity variability combine to affect the standard. This tutorial develops that comparison while keeping the confidence probability and relative tolerance fixed. It is aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus.

Definition: A risk segment is a group of risks treated as sharing the frequency and, where relevant, severity assumptions used in a credibility calculation. An exposure unit is the unit over which claim frequency is measured, such as one policy-year. A relative tolerance $k$ is the allowed deviation as a fraction of the expected value; for example, $k=0.20$ allows a deviation of 20% of that value.

Standards for Comparing Segments

Let $N$ denote the number of exposure units. For a claim-count standard, assume claims follow a Poisson process with frequency $\lambda$ claims per exposure unit. Then the total claim count $C$ over $N$ units has mean and variance $N\lambda$. Under the limited-fluctuation normal approximation, a central confidence probability $p$ and relative tolerance $k$ give the threshold

N_{\text{count}}=\frac{z^2}{k^2\lambda},

where $z$ is the standard normal quantile satisfying $P(-z\leq Z\leq z)=p$ for a standard normal random variable $Z$. The standard assumes homogeneous frequency across the segment and independent claims. The normal approximation is a planning standard, not an exact tail probability for every finite Poisson count.

For aggregate losses, suppose claim counts are Poisson with frequency $\lambda$, and individual severities $Y$ are independent and identically distributed, independent of the count, with mean $\mu>0$ and variance $\sigma^2<\infty$. Write $S$ for total loss over $N$ exposure units and define the severity squared coefficient of variation as $CV^2=\sigma^2/\mu^2$. The corresponding standard is

N_{\text{agg}}=\frac{z^2(1+CV^2)}{k^2\lambda}.

Both standards hold $z$ and $k$ fixed when comparing segments. The aggregate-loss standard has an additional factor $1+CV^2$: frequency alone is not enough to compare aggregate-loss credibility.

Why Severity Variability Changes the Threshold

For a single exposure unit, let the claim count be $M\sim\operatorname{Poisson}(\lambda)$, and let $S_1=\sum_{j=1}^{M}Y_j$. Conditional on $M$, the expected loss is $M\mu$, and the conditional variance is $M\sigma^2$. Applying the conditional variance formula gives

\begin{aligned} E[S_1]&=E[E[S_1\mid M]]=E[M\mu]=\lambda\mu,\\ \operatorname{Var}(S_1) &=E[\operatorname{Var}(S_1\mid M)] +\operatorname{Var}(E[S_1\mid M])\\ &=E[M\sigma^2]+\operatorname{Var}(M\mu)\\ &=\lambda\sigma^2+\lambda\mu^2 =\lambda(\sigma^2+\mu^2). \end{aligned}

For $N$ independent exposure units, the mean and variance both multiply by $N$. Therefore the squared coefficient of variation of total loss is

\frac{\operatorname{Var}(S)}{E[S]^2} =\frac{N\lambda(\sigma^2+\mu^2)}{(N\lambda\mu)^2} =\frac{1+CV^2}{N\lambda}.

The limited-fluctuation approximation requires $z$ times the relative standard deviation to be no greater than $k$. Squaring that inequality and rearranging yields the aggregate-loss threshold. The severity mean cancels because the tolerance is relative to expected loss; the severity variance remains through $CV^2$. If the standard instead used a fixed dollar tolerance, that cancellation would not generally occur.

Worked Segment Comparisons

Worked Example: Aggregate-Loss Exposure for a Gamma Severity Segment

A hypothetical portfolio has Poisson claim frequency of $0.20$ claims per policy-year. Claim severity follows a Gamma distribution with shape $4$ and scale $\$300$, so its mean is $\$1{,}200$ and its variance is $\$360{,}000$. Find the exposure threshold for central 95% confidence, using $z=1.96$, and relative tolerance $k=0.25$. Assume independent policy-years, no discounting, and no mortality model; the exposure unit is one policy-year.

Method and severity variation. For a Gamma distribution with shape $\alpha$ and scale $\theta$, the mean is $\alpha\theta$ and the variance is $\alpha\theta^2$. Here, $E[Y]=4(300)=\$1{,}200$, and $\operatorname{Var}(Y)=4(300^2)=\$360{,}000$. Thus $CV^2=360{,}000/(1{,}200^2)=0.25$.

Substitute in the aggregate standard.

N=\frac{1.96^2(1+0.25)}{0.25^2(0.20)} =\frac{3.8416(1.25)}{0.0625(0.20)} =\frac{4.802}{0.0125} =384.16\text{ policy-years}.

The calculated threshold is $384.16$, or 384.2 policy-years to one decimal place. This is the exposure-equivalent threshold. If exposure must be accumulated only in complete policy-years and the standard must be met rather than approximated, the operational minimum is the next whole unit, 385. The distinction is between reporting the calculated threshold and rounding up to a complete-unit minimum.

Check. At the unrounded threshold, $z\sqrt{(1+CV^2)/(N\lambda)}=1.96\sqrt{1.25/(384.16(0.20))}=0.25$, matching the specified tolerance.

Worked Example: Claim-Count Exposure at a Different Frequency

A hypothetical segment has Poisson frequency $0.06$ claims per policy-year. Find the exposure threshold for central 90% confidence with $z=1.645$ and relative tolerance $k=0.15$. Assume independent, homogeneous policy-years. There is no severity model in this claim-count calculation, and no discounting or mortality assumption is involved.

Method. Use the claim-count standard; adding a severity coefficient of variation would be inappropriate because the quantity being tested is claim count, not aggregate loss.

N=\frac{1.645^2}{0.15^2(0.06)} =\frac{2.706025}{0.0225(0.06)} =\frac{2.706025}{0.00135} =2004.46296\text{ policy-years}.

The unrounded threshold is $2004.46296$ policy-years. To the nearest whole policy-year it is 2004 policy-years; this is a rounded estimate of the threshold, not a sufficient whole-unit minimum. If the requirement is to meet or exceed the threshold using complete policy-years, the minimum is 2005. Rounding to the nearest integer and rounding up to a whole-unit minimum answer different questions.

Check. At the unrounded threshold, $1.645/\sqrt{N(0.06)}=0.15$. The result is therefore consistent with the stated confidence approximation and tolerance.

Worked Example: Comparing Two Aggregate-Loss Segments

Two hypothetical segments have the same target: central 95% confidence with $z=1.96$ and relative tolerance $k=0.20$. Segment A has Poisson frequency $0.12$ claims per policy-year and Gamma severity with shape $2$, scale $\$250$ (so $CV^2=1/2$). Segment B has frequency $0.30$ and Gamma severity with shape $0.5$, scale $\$1{,}000$ (so $CV^2=2$). Find each exposure threshold and compare them. Assume independent policy-years and no discounting or mortality model.

Segment A. Its severity mean is $2(250)=\$500$, variance is $2(250^2)=\$125{,}000$, and $CV^2=125{,}000/500^2=0.5$. Therefore,

N_A=\frac{1.96^2(1+0.5)}{0.20^2(0.12)} =\frac{3.8416(1.5)}{0.0048} =\frac{5.7624}{0.0048} =1200.5\text{ policy-years}.

Segment B. Its severity mean is $0.5(1{,}000)=\$500$, variance is $0.5(1{,}000^2)=\$500{,}000$, and $CV^2=500{,}000/500^2=2$. Therefore,

N_B=\frac{1.96^2(1+2)}{0.20^2(0.30)} =\frac{3.8416(3)}{0.012} =\frac{11.5248}{0.012} =960.4\text{ policy-years}.

The exposure ratio is $N_A/N_B=1200.5/960.4=1.25$. Segment A needs 25% more exposure, even though B has the higher severity $CV^2$, because B’s claim frequency is high enough to offset its greater severity variability. The same result follows directly from the ratio of $(1+CV^2)/\lambda$: for A it is $1.5/0.12=12.5$, while for B it is $3/0.30=10$, and $12.5/10=1.25$.

Check. The common multiplier $z^2/k^2$ is $3.8416/0.04=96.04$. Multiplying gives $96.04(12.5)=1200.5$ and $96.04(10)=960.4$, confirming both thresholds and their ratio.

Scaling Rules and Common Traps

With confidence and tolerance fixed, the factors $z^2/k^2$ are common to all segments. For claim counts, exposure varies inversely with frequency. For aggregate losses, compare $(1+CV^2)/\lambda$. The severity mean does not need to be equal across segments for this relative-tolerance comparison, provided each segment uses its own severity coefficient of variation and the model assumptions hold.

Identify the measured quantity.
Use the claim-count standard for counts and the aggregate-loss standard when random claim severities are included.

Put frequency on a common exposure basis.
For example, convert claims per month to claims per policy-year before comparing segments.

Include severity variation when required.
For aggregate losses, compute $CV^2=\sigma^2/\mu^2$ for each segment.

Compare thresholds consistently.
Use the same confidence quantile and tolerance in each segment, and state whether the reported exposure is fractional, rounded, or a whole-unit minimum.

A common mistake is to compare aggregate-loss standards using frequency alone. For the two segments above, using only the frequency ratio would give $N_A/N_B=0.30/0.12=2.5$, incorrectly suggesting A needs 2.5 times as much exposure. The correct ratio is

\frac{N_A}{N_B} =\frac{(1+CV_A^2)/\lambda_A}{(1+CV_B^2)/\lambda_B} =\frac{1.5/0.12}{3/0.30} =1.25.

The error omits the severity-variability factor, which materially changes the comparison. Another trap is treating a rounded threshold as a guaranteed minimum. A result of $2004.46296$ policy-years rounds to 2004 to the nearest whole number, but 2005 complete policy-years are required to meet or exceed that threshold. State which convention the question asks for.

Key takeaway: For a fixed confidence probability and relative tolerance, claim-count exposure scales as $1/\lambda$; aggregate-loss exposure scales as $(1+CV^2)/\lambda$. Compare segments using consistent frequency units, include severity variation for aggregate losses, and keep threshold rounding distinct from a whole-unit minimum.

Check Your Understanding

Use the limited-fluctuation standards above. Treat each exposure unit as one policy-year and report fractional thresholds to one decimal place unless the question asks for a whole-unit minimum.

A hypothetical segment has Poisson frequency $0.08$ claims per policy-year. Find the claim-count threshold for central 95% confidence, $z=1.96$, and relative tolerance $0.20$. Give both the calculated threshold and the minimum complete policy-years. (Answer: 1200.5 policy-years; 1201 complete policy-years)
A hypothetical segment has Poisson frequency $0.40$ claims per policy-year and Gamma severity with shape $5$, scale $\$200$. Find the aggregate-loss threshold for central 90% confidence, $z=1.645$, and relative tolerance $0.10$. (Answer: 811.8 policy-years; 812 complete policy-years)
Two segments share the same confidence and tolerance. Segment A has frequency $0.10$ and $CV_A^2=0.25$; segment B has frequency $0.25$ and $CV_B^2=1$. Find the ratio of their aggregate-loss exposure thresholds, $N_A/N_B$. (Answer: 1.5625)
A reference segment requires 500 exposure units and has frequency $0.20$ with $CV^2=0.5$. A second segment has frequency $0.30$ and $CV^2=1.5$, with the same confidence and tolerance. Find the second segment’s exposure threshold. (Answer: 555.6 exposure units)
Segment A has frequency $0.15$ and $CV_A^2=0.4$; segment B has frequency $0.30$ and $CV_B^2=1.1$. With identical confidence and tolerance, which needs more aggregate-loss exposure, and by what ratio? (Answer: Segment A; $N_A/N_B=1.3333$)

What You'll Learn