Coefficient of Variation in Severity Standards

Why Severity Variability Sets the Required Claim Count

A full-credibility standard for average claim severity asks how many observed claims are needed for the sample average to estimate the expected claim size with a specified level of precision. The answer depends on more than the expected dollar amount: it also depends on how variable individual claim sizes are relative to that mean. The coefficient of variation expresses that relative variability in a scale-free way.

This tutorial develops the severity standard using the claim size coefficient of variation, then applies it to several hypothetical severity distributions. The calculations are aligned with the CAS Exam 5 (Basic Ratemaking and Estimated Claim Liabilities) syllabus.

Definition: For a claim size $X$ with finite positive mean and finite variance, its coefficient of variation is $$ \operatorname{CV}(X)=\frac{\sqrt{\operatorname{Var}(X)}}{E[X]}. $$ It is the standard deviation divided by the mean, so it measures dispersion relative to the average claim size. The coefficient of variation is dimensionless: expressing all claim sizes in dollars or in thousands of dollars does not change it.

Severity Full-Credibility Standard

Suppose $X_1,\ldots,X_n$ are independent claim severities drawn from the same distribution as $X$, with $E[X]=\mu>0$ and $\operatorname{Var}(X)=\sigma^2<\infty$. The estimate of expected severity is the sample average $\bar X=(X_1+\cdots+X_n)/n$. A common limited-fluctuation precision requirement specifies a confidence level $p$ and a relative error tolerance $r$: the sample average should be within $r$ of the true mean with probability at least $p$.

Severity full-credibility standard: Under a normal model for $\bar X$, or a normal approximation to its distribution, the minimum claim count satisfying $$ P\left(\left|\bar X-\mu\right|\leq r\mu\right)\geq p $$ is calculated as $$ n_0=\left(\frac{z_{(1+p)/2}\operatorname{CV}(X)}{r}\right)^2. $$ The required whole-number claim count is $N=\lceil n_0\rceil$, the smallest integer greater than or equal to $n_0$. Here $z_{(1+p)/2}$ is the standard normal quantile with cumulative probability $(1+p)/2$.

The model assumptions matter. The claims must be comparable observations from a stable severity distribution, and the observations must be independent. A finite variance is needed for this normal approximation. For normally distributed claim sizes, the sample average is exactly normal. For other distributions, the formula is an approximation based on the central limit theorem; strong skewness or a heavy tail can make the approximation less reliable at modest claim counts. In particular, a finite mean alone is not enough for this variance-based standard if the severity distribution has infinite variance.

Why the Formula Uses the Coefficient of Variation

The derivation shows why the mean and standard deviation enter through their ratio. By independence, the average of $n$ observations has mean $\mu$ and variance $\sigma^2/n$. Under the normal model or approximation, standardizing that average gives a standard normal quantity:

\frac{\bar X-\mu}{\sigma/\sqrt n}\ \approx\ N(0,1).

The required probability is therefore approximately the probability that a standard normal variable lies between two symmetric bounds:

P\left(\left|\bar X-\mu\right|\leq r\mu\right) \approx P\left( \left| \frac{\bar X-\mu}{\sigma/\sqrt n} \right| \leq \frac{r\mu\sqrt n}{\sigma} \right).

For a central probability of $p$, the corresponding standard normal bounds are $-z_{(1+p)/2}$ and $z_{(1+p)/2}$. Thus the precision requirement is met when

\frac{r\mu\sqrt n}{\sigma}\geq z_{(1+p)/2}.

Solving this inequality for $n$, and using $\operatorname{CV}(X)=\sigma/\mu$, gives

n\geq \left(\frac{z_{(1+p)/2}\sigma}{r\mu}\right)^2 = \left(\frac{z_{(1+p)/2}\operatorname{CV}(X)}{r}\right)^2.

Since the number of claims must be a whole number, the calculated threshold is rounded up. The squared relationship has a useful interpretation: doubling the coefficient of variation multiplies the required count by four, while doubling the tolerance divides it by four. The standard depends on the relative, not absolute, size of the severity variation.

Worked Example: Exponential Claim Severities

A hypothetical equipment-breakdown portfolio has claim severities modeled as independent exponential random variables with mean $\$1{,}200$. Determine the full-credibility claim count for a 90% central confidence level and a 10% relative error tolerance.

Method and distribution parameters. For an exponential distribution with mean parameter $\theta$, $E[X]=\theta$ and $\operatorname{Var}(X)=\theta^2$. Here $\theta=\$1{,}200$, so the standard deviation is also $\$1{,}200$, and $\operatorname{CV}(X)=1$. A 90% central confidence level uses $z_{0.95}\approx1.645$, and $r=0.10$.

Substitute these values into the severity standard:

n_0= \left(\frac{1.645(1)}{0.10}\right)^2 = (16.45)^2 = 270.6025.

The required count is $N=\lceil270.6025\rceil=271$ claims. As a check, using the approximate relative error bound at $n=271$ gives $1.645(1)/\sqrt{271}\approx0.0999$, which is below 0.10. At $n=270$, it is approximately 0.1001, which is above the tolerance. The whole-number threshold is therefore 271.

Common mistake avoided: Rounding 270.6025 to the nearest integer would give 271 in this case, but nearest-integer rounding is not the rule. The standard must always be rounded up; if the unrounded result were 270.1, 270 claims would not meet the stated threshold.

Worked Example: Gamma Severity and a Tighter Tolerance

A hypothetical specialty insurer models individual claim severities with a gamma distribution having shape $\alpha=4$ and scale $\theta=\$500$. Calculate the claim count for 95% central confidence and a relative error tolerance of 8%.

Method and distribution parameters. For a gamma distribution parameterized by shape and scale, $E[X]=\alpha\theta$ and $\operatorname{Var}(X)=\alpha\theta^2$. Therefore,

E[X]=4(\$500)=\$2{,}000,\qquad \operatorname{Var}(X)=4(\$500)^2=\$1{,}000{,}000.

The standard deviation is $\sqrt{\$1{,}000{,}000}=\$1{,}000$, so $\operatorname{CV}(X)=\$1{,}000/\$2{,}000=0.5$. For 95% central confidence, $z_{0.975}\approx1.96$; the tolerance is $r=0.08$. Thus,

n_0= \left(\frac{1.96(0.5)}{0.08}\right)^2 = \left(\frac{0.98}{0.08}\right)^2 = (12.25)^2 = 150.0625.

Round up to obtain $N=151$ claims. Checking the approximate relative error bound at 151 gives $1.96(0.5)/\sqrt{151}\approx0.0797$, which is below 0.08. At 150 it is approximately 0.0800, just above 0.08 before rounding. The standard calls for 151 claims.

Common mistake avoided: The gamma scale $\theta$ is not the mean when the shape is greater than one. Using $\$500$ as the mean would produce the wrong coefficient of variation. The gamma mean is $\alpha\theta=\$2{,}000$, and both moments must use the same parameterization.

Comparing Severity Distributions

For a fixed confidence level and tolerance, the required claim count changes with the coefficient of variation. A distribution's name alone does not determine the standard: its parameters determine its mean and variance, and their ratio determines the coefficient of variation. The next example compares two hypothetical models with the same mean but different relative variability.

Worked Example: Comparing Gamma and Lognormal Models

Two hypothetical models are considered for claims on the same type of policy. Model A is gamma with shape $\alpha=4$ and scale $\theta=\$500$. Model B is lognormal with log-scale parameters $\mu_{\log}=\ln(2{,}000)-\tfrac12\ln(1.64)$ and $\sigma_{\log}^2=\ln(1.64)$. For each model, calculate the severity standard at 95% central confidence and a 10% relative error tolerance.

Model A: gamma. Its mean is $4(\$500)=\$2{,}000$, and its variance is $4(\$500)^2=\$1{,}000{,}000$. The standard deviation is $\$1{,}000$, giving $\operatorname{CV}(X)=0.5$. Therefore,

n_{0,A} = \left(\frac{1.96(0.5)}{0.10}\right)^2 = (9.8)^2 = 96.04, \qquad N_A=\lceil96.04\rceil=97.

Model B: lognormal. If $\ln X$ is normal with mean $\mu_{\log}$ and variance $\sigma_{\log}^2$, then $E[X]=e^{\mu_{\log}+\sigma_{\log}^2/2}$ and $\operatorname{CV}(X)=\sqrt{e^{\sigma_{\log}^2}-1}$. The selected parameters give a mean of $e^{\ln(2{,}000)-\ln(1.64)/2+\ln(1.64)/2}=\$2{,}000$. They also give

\operatorname{CV}(X) = \sqrt{e^{\ln(1.64)}-1} = \sqrt{1.64-1} = 0.8.

The lognormal standard is

n_{0,B} = \left(\frac{1.96(0.8)}{0.10}\right)^2 = (15.68)^2 = 245.8624, \qquad N_B=\lceil245.8624\rceil=246.

The lognormal model requires 246 claims, compared with 97 for the gamma model. Before rounding, the ratio of the standards is $245.8624/96.04=2.56$, equal to $(0.8/0.5)^2$. The higher-CV model requires substantially more claims despite having the same mean severity.

Common mistake avoided: Equal expected claim sizes do not imply equal severity standards. The standard is driven by the coefficient of variation, so comparing means alone misses the material difference between these models.

Interpretation, Limits, and Exam Traps

The standard is a precision rule for the average of claim severities under the stated model. It does not, by itself, establish that the underlying severity distribution is appropriate, that the data are representative, or that a resulting rate is adequate. The calculated number counts claims, not policies or exposure units: a policy count cannot be substituted unless the number of claims per policy is known and the problem has appropriately converted the volume.

The coefficient of variation is useful because it makes the standard comparable across currency units and across portfolios with different average claim sizes. For example, multiplying every severity by 100 multiplies both the mean and standard deviation by 100, leaving their ratio unchanged. The required count therefore stays the same when only the units of measurement change.

However, the formula does not capture every feature of a severity distribution. It summarizes variability through the first two moments and relies on the normal model or approximation for the sample mean. When claims are highly skewed, a few extreme observations can have a large effect, and a normal approximation at the calculated count may be questionable. If variance is infinite, the coefficient of variation is not finite and this standard cannot be applied as written. In practice, the model, data quality, claim definition, and stability of the book should be considered alongside the numerical threshold.

Do not use the standard deviation in place of the coefficient of variation. Divide the standard deviation by the mean before substituting.
Do not treat $r$ as a percentage number. A 10% tolerance is $r=0.10$, not $r=10$.
Do not round down or to the nearest whole claim. Apply the ceiling to the unrounded standard.
Do not confuse confidence with tolerance. The confidence level determines the normal quantile; the relative tolerance appears in the denominator.
Do not assume different distributions automatically have different standards. Under this formula, distributions with the same coefficient of variation and the same $p$ and $r$ have the same unrounded standard, though approximation quality may differ.

Key takeaway: Under the normal model or approximation, the severity full-credibility claim count is the square of the normal quantile times the claim size coefficient of variation divided by the relative error tolerance. Calculate the coefficient of variation from the specified distribution, use consistent parameters, and round the resulting count up.

Check Your Understanding

For each hypothetical severity model, find the full-credibility claim count using the stated central confidence level and relative error tolerance. Round only the final count up.

Claim severity is exponential with mean $\$800$. Find the count for 90% central confidence and an 8% tolerance. (Answer: 423 claims)
Claim severity is gamma with shape $\alpha=9$ and scale $\theta=\$200$. Find the count for 95% central confidence and a 10% tolerance. (Answer: 43 claims)
A severity model has mean $\$3{,}000$ and standard deviation $\$1{,}200$. Find the count for 99% central confidence and a 5% tolerance, using $z_{0.995}=2.576$. (Answer: 425 claims)
Two severity models have coefficients of variation 0.6 and 0.9. Find each count for 90% central confidence and a 12% tolerance, using $z_{0.95}=1.645$, and state the ratio of their unrounded standards. (Answer: 68 claims and 153 claims; the unrounded-standard ratio is 2.25)
A lognormal severity model has $\sigma_{\log}^2=\ln(1.49)$ and mean $\$1{,}000$. Find its count for 95% central confidence and a 7% tolerance. Use $z_{0.975}=1.96$. (Answer: 385 claims)

What You'll Learn