Partial Credibility for Claim Severity

Why Severity Credibility Depends on Claim Count

An observed average claim severity is based on a sample of claim amounts. With few claims, that average can vary substantially from the underlying mean severity; with more claims, it is generally more stable. Partial credibility expresses how much of the claim-count volume required for full credibility has been observed.

The calculation is tied to a stated full-credibility standard: a confidence level and an allowable error measured relative to the expected severity. The partial-credibility factor then compares the actual number of claims with the number required by that standard. It is a volume-based factor, not a direct measure of whether the observed average happens to be close to the expected severity.

Definition: Let $X_1,\ldots,X_n$ be independent claim severities with common mean $\mu=E[X]$ and finite variance $\sigma^2=\operatorname{Var}(X)$. The observed average severity is $\bar X=\sum_{j=1}^n X_j/n$, and the severity coefficient of variation is $\operatorname{CV}=\sigma/\mu$. If $n_s$ claims are required for the specified full-credibility standard, the partial-credibility factor for $n<n_s$ observed claims is $Z=\sqrt{n/n_s}$.

From a Full Standard to a Partial Factor

Suppose the full-credibility requirement is that the observed average be within a relative error $k$ of the expected severity, with probability at least $p$. In symbols, the target is approximately $P(|\bar X-\mu|\leq k\mu)\geq p$. The value $k$ is a proportion: a 10% tolerance means $k=0.10$, not $10$.

Under the normal approximation for the sample average, the standard deviation of $\bar X$ is $\sigma/\sqrt n$. Let $z_{(1+p)/2}$ be the standard normal quantile that gives a central probability $p$. Equating the allowable error to this many standard errors gives the severity full-credibility standard.

n_s=\left(\frac{z_{(1+p)/2}\operatorname{CV}}{k}\right)^2

For $n<n_s$, the square-root partial factor is

Z=\sqrt{\frac{n}{n_s}} =\frac{k\sqrt n}{z_{(1+p)/2}\operatorname{CV}}.

The $n$ in these formulas is the number of claims whose severities are averaged. It is not automatically the number of policies, vehicles, insured locations, or exposure units. Those quantities matter only insofar as they determine how many claims are available for the severity calculation.

Why the Square-Root Rule Holds

The rule follows from how the sampling variability of an average changes with sample size. For independent claim severities with common variance $\sigma^2$, the variance of the average is

\operatorname{Var}(\bar X) =\operatorname{Var}\left(\frac{1}{n}\sum_{j=1}^n X_j\right) =\frac{1}{n^2}\sum_{j=1}^n\operatorname{Var}(X_j) =\frac{\sigma^2}{n}.

Consequently, the standard deviation of $\bar X$, expressed as a proportion of the mean $\mu$, is $\operatorname{CV}/\sqrt n$. A central normal-approximation requirement for a relative error $k$ is therefore

\frac{k\mu}{\sigma/\sqrt n} =\frac{k\sqrt n}{\operatorname{CV}} \geq z_{(1+p)/2}.

Solving this inequality for $n$ gives $n\geq n_s$. If a smaller sample of $n$ claims is available, its standard-error scale relative to the full standard is proportional to $\sqrt{n/n_s}$, which gives the partial factor. This reasoning depends on independent, identically distributed claim severities with finite mean and variance, and on the normal approximation being adequate for the average. It may be unreliable when the sample is very small or the severity distribution is extremely skewed.

Worked Calculations

Worked Example: A Gamma Severity Model and a 90% Standard

Method: calculate the full standard from the relative error requirement, then take the square root of the observed-to-required claim-count ratio. This example avoids treating the individual severity distribution as normal merely because a normal approximation is used for the average.

A hypothetical equipment insurer models individual claim severity $X$ with a Gamma distribution having shape $\alpha=16/9$ and rate $\beta=1/675$, in dollars. For a 90% central confidence level, use $z_{0.95}=1.645$, and require the average severity to be within 10% of its mean. The insurer has 64 claims. Find the partial-credibility factor.

For this Gamma parameterization, $E[X]=\alpha/\beta$ and $\operatorname{Var}(X)=\alpha/\beta^2$. Thus $E[X]=(16/9)/(1/675)=1{,}200$ dollars. The variance is $810{,}000$ dollars squared, so the standard deviation is $900$ dollars and $\operatorname{CV}=900/1{,}200=0.75$.

The full-credibility standard is

n_s=\left(\frac{1.645(0.75)}{0.10}\right)^2 =(12.3375)^2 =152.21390625.

The observed count, 64, is below this standard. Using the unrounded standard,

Z=\sqrt{\frac{64}{152.21390625}} =\frac{8}{12.3375} =0.648429\ldots

Rounded to four decimal places, the partial-credibility factor is 0.6484. The calculation uses 64 claims, not the number of policies that generated them.

Worked Example: A Claim Count Near the Full Standard

Method: compute the severity standard from the given mean and standard deviation, then use the observed claim count. The common mistake avoided here is rounding the standard to an integer before calculating the factor.

A hypothetical repair program has Gamma-distributed claim severities with shape $\alpha=100/9$ and rate $\beta=1/90$, in dollars. The target is 95% central confidence with a relative error of 6%. There are 25 claims. Use $z_{0.975}=1.96$.

The mean is $\alpha/\beta=(100/9)/(1/90)=1{,}000$ dollars. The variance is $\alpha/\beta^2=90{,}000$ dollars squared, so the standard deviation is $300$ dollars and the coefficient of variation is $300/1{,}000=0.30$.

The full standard is

n_s=\left(\frac{1.96(0.30)}{0.06}\right)^2 =(9.8)^2 =96.04.

The factor uses $n_s=96.04$, not the rounded-up integer threshold of 97:

Z=\sqrt{\frac{25}{96.04}} =\frac{5}{9.8} =0.510204\ldots

Rounded to four decimal places, $Z=\mathbf{0.5102}$. At least 97 whole claims would be needed to meet or exceed the full standard, but rounding the standard upward is not a reason to replace 96.04 in this partial-factor calculation.

Worked Example: Higher Severity Variability Requires More Claims

Method: determine the coefficient of variation from the stated Gamma parameters, calculate the full standard, and compare it with the actual claim count. This example avoids assuming that the same claim count gives the same severity credibility for every class.

A hypothetical specialty contractor class has Gamma claim severities with shape $\alpha=25/16$ and rate $\beta=1/1{,}280$, in dollars. The standard uses 90% central confidence, $z_{0.95}=1.645$, and a 8% relative error tolerance. The class has 144 claims. Find its partial factor.

The mean is $\alpha/\beta=(25/16)(1{,}280)=2{,}000$ dollars. The variance is $\alpha/\beta^2=2{,}560{,}000$ dollars squared, giving standard deviation $1{,}600$ dollars and $\operatorname{CV}=1{,}600/2{,}000=0.80$.

The full-credibility standard is

n_s=\left(\frac{1.645(0.80)}{0.08}\right)^2 =(16.45)^2 =270.6025.

Since 144 claims are fewer than 270.6025, the partial factor is

Z=\sqrt{\frac{144}{270.6025}} =\frac{12}{16.45} =0.729483\ldots

Rounded to four decimal places, $Z=\mathbf{0.7295}$. The large coefficient of variation raises the claim count required for full credibility, which lowers the factor compared with a less variable severity class having the same claim count and standard settings.

Interpretation and Common Traps

For a fixed full standard, the factor grows with the square root of the claim count, not in direct proportion to the claim count. Doubling the number of claims therefore multiplies the factor by $\sqrt{2}$, provided the count remains below the full standard. Likewise, a more demanding confidence level or a smaller tolerated relative error increases the required standard and reduces the factor for a fixed observed count.

A frequent error is to omit the square root. In the second worked example, the incorrect calculation $25/96.04=0.2603$ is not the partial-credibility factor. The required calculation is $\sqrt{25/96.04}=0.5102$. Another common error is to use a policy count in place of the number of claims. Severity is calculated over claim amounts; exposure units alone do not determine the sampling variation of the observed average severity.

Keep the error tolerance relative. Use $k=0.06$ for 6%, rather than $6$ or a dollar amount.
Use the severity coefficient of variation. The standard depends on $\sigma/\mu$, not on the standard deviation alone.
Do not round intermediate values unnecessarily. Carry the full standard through the square-root calculation and round the final factor as requested.
Check the model basis. The mean and variance must describe the same claim-severity definition as the observed average.

The factor is a measure of volume relative to a specified statistical standard. It does not, by itself, establish that the severity observations are homogeneous, independent, or representative. Those assumptions and the suitability of the normal approximation should be considered before relying on the calculation. Partial-credibility calculations of this form are aligned with the CAS Exam 5 (Basic Ratemaking and Estimated Claim Liabilities) syllabus.

Key takeaway: For an observed average severity based on $n$ claims, first obtain $n_s=(z_{(1+p)/2}\operatorname{CV}/k)^2$ from the specified full standard. When $n<n_s$, calculate the partial factor as $Z=\sqrt{n/n_s}$, using claim count and retaining unrounded intermediate values.

Check Your Understanding

For each problem, calculate the factor using the stated full-credibility standard or its inputs. Round factors to four decimal places.

A severity standard requires 400 claims. A hypothetical class has 100 claims. What is its partial-credibility factor? (Answer: $Z=0.5000$)
For a hypothetical severity portfolio, $\operatorname{CV}=0.50$, $p=0.90$, $z_{0.95}=1.645$, and $k=0.10$. There are 16 claims. Calculate the full standard and partial factor. (Answer: $n_s=67.650625$, so at least 68 claims meet the standard; $Z=0.4863$)
A hypothetical product has mean severity \$1,500 and standard deviation \$600. For 95% central confidence use $z_{0.975}=1.96$; the relative error tolerance is 8%, and 49 claims are observed. Find the factor. (Answer: $\operatorname{CV}=0.40$, $n_s=96.04$, and $Z=0.7143$)
A claim class has $\operatorname{CV}=0.60$, a 90% central confidence target with $z_{0.95}=1.645$, and a 15% error tolerance. If 20 claims are observed, find the factor and the minimum whole-claim count that meets the standard. (Answer: $n_s=43.2964$, $Z=0.6797$, and at least 44 claims are needed)
Two hypothetical severity classes each have 25 claims and use a 90% central confidence target with $z_{0.95}=1.645$ and $k=0.10$. Class A has $\operatorname{CV}=0.40$; Class B has $\operatorname{CV}=0.80$. Find both factors and state which is lower. (Answer: Class A $n_s=43.2964$, $Z=0.7599$; Class B $n_s=173.1856$, $Z=0.3799$; Class B has the lower factor)

What You'll Learn