Full Credibility Standards for Territory Data

Why Full-Credibility Standards Are Needed

Territory experience is subject to random fluctuation. A territory may have a claim count or loss total above or below its expected value simply because claims occur unpredictably. A full-credibility standard states how much experience is needed so that this random fluctuation is likely to remain within a specified tolerance. It is a benchmark for the volume of data; it does not, by itself, say how to adjust an indicated territory factor when the experience falls short.

A standard must specify both a confidence level and an error tolerance. For example, “95% confidence within 5%” means that the probability the observed measure is within 5% of its expected value is at least 95%, under the model being used. Changing either part changes the volume required. The resulting threshold is a model-based approximation, not a guarantee that a particular territory’s experience will be stable.

Definition: A limited-fluctuation full-credibility standard specifies a target probability $p$ and a relative error tolerance $k$. For a random experience measure $T$ with mean $\mu_T>0$, the objective is $$ P\left(\left|T-\mu_T\right|\leq k\mu_T\right)\geq p. $$ The tolerance $k$ is expressed as a proportion: a 5% tolerance is $k=0.05$. A standard is meaningful only after the measure, probability model, confidence level, and tolerance have been identified.

For the usual claim-count standard, the model is $N\sim\operatorname{Poisson}(\lambda)$, where $N$ is the number of claims in a specified territory and period, and $\lambda$ is its expected count. This model assumes claims occur independently at a constant rate over the exposure represented by the data. Its mean and variance are both $\lambda$. Using the normal approximation to a Poisson count, the required expected count is

\lambda_{\text{full}}\approx \left(\frac{z_{(1+p)/2}}{k}\right)^2,

where $z_{(1+p)/2}$ is the standard normal quantile with cumulative probability $(1+p)/2$. For instance, a central 95% confidence level uses $z_{0.975}\approx1.96$. The calculated threshold is an expected count; when reporting a minimum whole-number claim-count benchmark, round it up.

Deriving the Claim-Count Standard

The claim-count formula follows directly from the normal approximation. Under that approximation, $N$ has mean $\lambda$ and standard deviation $\sqrt{\lambda}$. The required interval around the mean has half-width $k\lambda$. Expressing this half-width in standard deviations gives the standardized boundary:

\frac{k\lambda}{\sqrt{\lambda}}=k\sqrt{\lambda}.

For the central probability to be at least $p$, this boundary must be at least $z_{(1+p)/2}$. Thus $k\sqrt{\lambda}\geq z_{(1+p)/2}$; squaring both nonnegative sides and rearranging gives $\lambda\geq z_{(1+p)/2}^2/k^2$. The derivation also explains the directions of change: a higher confidence level raises $z$, and a tighter tolerance lowers $k$, so either requires more expected claims.

The formula uses a normal approximation and omits continuity corrections. It is a conventional limited-fluctuation standard, not an exact Poisson probability calculation. If an exact probability is required, the relevant Poisson probabilities should be calculated directly. The approximation is also not a remedy for a count process whose variance materially exceeds its mean or whose claims are dependent.

Confidence and tolerance	Normal quantile	Approximate expected-count threshold
90% within 5%	1.645	1,082.41
95% within 5%	1.960	1,536.64
95% within 10%	1.960	384.16
99% within 5%	2.576	2,654.31

These values describe expected counts, so they need not be integers. For example, the usual whole-number benchmark for 95% confidence within 5% is 1,537 expected claims. The table is not a set of universal requirements: the selected confidence and tolerance should reflect the intended use of the territorial measure.

Worked Example: Checking a Claim-Count Standard

A hypothetical territory has an expected count of 1,500 claims for the period being studied. Determine whether it meets a standard of 95% confidence within 5% under a Poisson count model. The count is modeled as $N\sim\operatorname{Poisson}(\lambda)$, with $\lambda=1{,}500$. No interest or mortality assumptions apply to this count calculation.

Step 1: Identify the model inputs. For 95% central confidence, $z_{0.975}=1.96$. The relative error tolerance is $k=0.05$.

Step 2: Calculate the expected-count threshold.

\lambda_{\text{full}}\approx\left(\frac{1.96}{0.05}\right)^2 =(39.2)^2=1{,}536.64.

The whole-number benchmark is $\lceil1{,}536.64\rceil=1{,}537$ expected claims.

Step 3: Compare the territory with the threshold. Its expected count, 1,500, is below 1,536.64, so it does not meet this stated full-credibility standard. As a check, the normal-approximation boundary for the territory is $0.05\sqrt{1{,}500}\approx1.9365$, which is less than 1.96.

Method and mistake avoided. This is the Poisson limited-fluctuation count standard. A common mistake is to round 1,536.64 down to 1,536; doing so would report a benchmark below the calculated requirement. The whole-number benchmark must be rounded up.

Converting Claim Counts to Exposure Requirements

A count threshold can be translated into an exposure requirement if the expected claim frequency per exposure is known and appropriate for the territory and period. If $e$ is the exposure volume and $f$ is expected claims per exposure, the Poisson mean is $\lambda=ef$. Substituting into the count standard gives the exposure threshold:

e_{\text{full}}\approx \frac{z_{(1+p)/2}^{\,2}}{k^2 f}.

Use a frequency with the same exposure and time units as $e$. For example, a frequency expressed per policy-year must be multiplied by policy-years, not by policies-in-force measured at a single date. The conversion assumes the expected frequency is applicable to the exposure being assessed. If mix, coverage, or period differences make it inapplicable, the calculation does not establish a valid territory threshold merely because the units divide correctly.

Worked Example: Translating a Count Standard into Policy-Years

A hypothetical territory has an expected claim frequency of 0.06 claims per policy-year. Find the exposure needed for 90% confidence within 5%, and determine whether 10,000 policy-years meet the standard. Model the claim count as $N\sim\operatorname{Poisson}(\lambda)$, where $\lambda=0.06e$ for $e$ policy-years. Interest and mortality assumptions do not apply.

Step 1: Calculate the expected-count requirement. For central 90% confidence, use $z_{0.95}=1.645$; the tolerance is $k=0.05$.

\lambda_{\text{full}}\approx\left(\frac{1.645}{0.05}\right)^2 =(32.9)^2=1{,}082.41.

Step 2: Convert the requirement to policy-years. Since the frequency is 0.06 claims per policy-year,

e_{\text{full}}\approx\frac{1{,}082.41}{0.06} =18{,}040.1667\text{ policy-years}.

If exposure must be reported in whole policy-years, round up to 18,041. At that volume the expected count is $18{,}041(0.06)=1{,}082.46$, which exceeds 1,082.41.

Step 3: Check the available experience. At 10,000 policy-years, the expected count is $10{,}000(0.06)=600$, below 1,082.41. Equivalently, 10,000 is below the 18,040.1667 policy-year threshold.

Method and mistake avoided. The method first finds the count threshold and then divides by frequency. Using 0.06 claims per policy-year as though it were 6 claims per policy-year would understate the required volume by a factor of 100. Keep the frequency units explicit throughout.

Full Credibility for Aggregate Losses

A count standard does not automatically provide a full-credibility standard for a loss total. Loss amounts vary as well as claim counts. For aggregate losses, use a compound Poisson model: $N\sim\operatorname{Poisson}(\lambda)$, individual severities $X_1,X_2,\ldots$ are independent and identically distributed, and severities are independent of $N$. Let $S=\sum_{j=1}^{N}X_j$, with $S=0$ if $N=0$. Assume $0<E[X]$ and finite $E[X^2]$.

Conditional on $N$, the expected aggregate is $N E[X]$ and its conditional variance is $N\operatorname{Var}(X)$. The law of total variance therefore gives

E[S]=\lambda E[X], \qquad \operatorname{Var}(S) =E[N]\operatorname{Var}(X)+\operatorname{Var}(N)(E[X])^2 =\lambda E[X^2].

The last equality uses $\operatorname{Var}(N)=E[N]=\lambda$ and $E[X^2]=\operatorname{Var}(X)+(E[X])^2$. Using a normal approximation for the aggregate loss, its coefficient of variation is $\sqrt{E[X^2]}/(\sqrt{\lambda}E[X])$. Requiring the aggregate loss to be within relative tolerance $k$ of its mean with central confidence $p$ gives:

\lambda_{\text{full}} \approx \frac{z_{(1+p)/2}^{\,2}E[X^2]} {k^2(E[X])^2}.

This formula differs from the count formula by the factor $E[X^2]/(E[X])^2=1+\operatorname{Var}(X)/(E[X])^2$, which is at least 1. When severity is variable, aggregate losses generally require more expected claims than the count-only standard at the same confidence and tolerance. The formula does not apply as stated if severity has no finite second moment, claim amounts are dependent in a way that changes aggregate variance, or the compound Poisson assumptions are inappropriate. A normal approximation may also be unreliable for a highly skewed aggregate with relatively few expected claims.

Worked Example: Allowing for Random Claim Severity

A hypothetical territory has claim frequency $f=0.04$ claims per policy-year. Claim severities follow an exponential distribution with mean $\theta=\$2{,}500$, equivalently rate $1/\theta=1/2{,}500$ per dollar. Claim count is $N\sim\operatorname{Poisson}(\lambda)$, independent of the independent severities, and $\lambda=0.04e$ for $e$ policy-years. Find the approximate full-credibility volume for aggregate losses at 95% confidence within 10%. No interest or mortality assumptions apply.

Step 1: Calculate the severity moments. For an exponential severity with mean $\theta$, $E[X]=\theta$ and $E[X^2]=2\theta^2$. Thus $E[X]=\$2{,}500$, $E[X^2]=2(\$2{,}500)^2=\$12{,}500{,}000$, and $E[X^2]/(E[X])^2=2$.

Step 2: Calculate the expected-claim threshold. For 95% confidence, $z_{0.975}=1.96$, so $z^2=3.8416$. The tolerance is $k=0.10$, so $k^2=0.01$. Therefore,

\lambda_{\text{full}}\approx \frac{3.8416(12{,}500{,}000)} {0.01(2{,}500)^2} = \frac{3.8416(2)}{0.01} =768.32.

The whole-number expected-claim benchmark is 769. As a check, at $\lambda=768.32$, the aggregate-loss standard deviation divided by its mean is $\sqrt{2/768.32}\approx0.05103$. The 10% tolerance is $0.10/0.05103\approx1.96$ standard deviations, as required.

Step 3: Convert the threshold to policy-years.

e_{\text{full}}\approx\frac{768.32}{0.04} =19{,}208\text{ policy-years}.

At 19,208 policy-years, the expected count is $19{,}208(0.04)=768.32$, matching the unrounded threshold.

Method and mistake avoided. This is the compound Poisson aggregate-loss standard. Applying the count-only formula would give $3.8416/0.01=384.16$ expected claims, only half the aggregate-loss threshold. That calculation ignores the additional fluctuation from random severity; the exponential model’s second-moment factor is 2.

Model Checks and Common Exam Traps

Before applying a standard to territorial data, confirm what is being measured. A count standard addresses claim-count fluctuation; an aggregate-loss standard includes both count and severity fluctuation under the stated compound Poisson model. A standard based on a loss ratio or loss cost would also need a clear denominator and an appropriate model for its variability. Do not substitute one measure’s formula for another without establishing the assumptions.

Using a different confidence convention. A central 95% standard uses $z_{0.975}$, not $z_{0.95}$. State the central probability and quantile explicitly.
Misreading the tolerance. A 5% tolerance means $k=0.05$, and the formula uses $k^2$. Entering 5 instead of 0.05 changes the result substantially.
Ignoring severity variation. The count formula is not the aggregate-loss formula. For compound Poisson losses, use $E[X^2]$, not just $E[X]$, and not $\operatorname{Var}(X)$ alone.
Assuming Poisson variation without checking it. If counts are clustered, dependent, or overdispersed, the equality $\operatorname{Var}(N)=E[N]$ may fail. The standard Poisson threshold may then understate the required data volume.
Confusing a benchmark with a credibility adjustment. Meeting or missing the threshold answers whether the stated standard is met under its assumptions. It does not itself specify how to blend an indication or choose a complement.

Key takeaway: State the measure, model, confidence level, and relative error tolerance before calculating a full-credibility threshold. Use the Poisson count formula for counts and the compound Poisson second-moment formula for aggregate losses; translate expected claims into exposure only with a frequency that matches the exposure basis.

These methods are aligned with the CAS Exam 5 (Basic Ratemaking and Estimated Claim Liabilities) syllabus. In practice, a calculated threshold is a model-dependent benchmark, so its assumptions and the quality of the territorial data matter as much as the arithmetic.

Check Your Understanding

For each problem, identify the applicable model and keep the confidence quantile, tolerance, and units consistent.

A hypothetical territory has a Poisson claim count with expected count $\lambda$. Find the whole-number expected-count benchmark for 90% confidence within 10%. (Answer: 271 expected claims)
Under a Poisson count model, find the whole-number expected-count benchmark for 95% confidence within 8%. Use $z_{0.975}=1.96$. (Answer: 601 expected claims)
A territory has expected frequency 0.025 claims per policy-year. Under a Poisson model, how many whole policy-years are needed for 90% confidence within 5%, using $z_{0.95}=1.645$? (Answer: 43,297 policy-years)
Aggregate losses follow a compound Poisson model. Claim severities are independent and take values of \$1,000 or \$5,000 with equal probability. Find the whole-number expected-claim threshold for 90% confidence within 10%, using $z_{0.95}=1.645$. (Answer: 391 expected claims)
A count measure is approximately normal with mean $\lambda$ and variance $1.5\lambda$, rather than Poisson variance $\lambda$. Using that stated variance model, find the whole-number expected-count threshold for 95% confidence within 10%, with $z_{0.975}=1.96$. (Answer: 577 expected claims)

What You'll Learn