How Severity Variation Changes Required Volume

Why Severity Variation Affects Credibility Volume

An aggregate-loss credibility standard asks whether observed total losses are sufficiently stable around their expected value. Even when two portfolios have the same average claim severity, their total losses need not fluctuate equally: a portfolio with more variable claim amounts generally needs more expected claims to meet the same standard. The key measure is the severity coefficient of variation, not just the average claim size.

The result in this tutorial concerns the expected claim-count threshold under a compound Poisson model. It does not convert that threshold into exposure units. The required claim count is determined by the aggregate-loss fluctuation standard and the severity distribution; translating it into exposure also requires a claim frequency.

Definition: Let $N\sim\operatorname{Poisson}(\lambda)$ be the number of claims, and let $X_1,X_2,\ldots$ be independent and identically distributed claim severities, independent of $N$. Assume $E[X]=\mu>0$ and $\operatorname{Var}(X)=\sigma^2<\infty$. The aggregate loss is $S=\sum_{j=1}^{N}X_j$, with $S=0$ when $N=0$. The severity coefficient of variation is $\operatorname{CV}=\sigma/\mu$.

A limited-fluctuation standard specifies a target probability $p$ and a relative tolerance $k$. It seeks enough expected claims that aggregate loss is within $k$ times its mean with probability at least $p$. Under the normal approximation, write $z$ for the standard normal quantile satisfying $P(-z\le Z\le z)=p$, where $Z\sim N(0,1)$. For example, a central 95% standard uses $z=1.96$. The resulting threshold depends on the severity distribution through its second moment.

Deriving the Severity Multiplier

For a fixed claim count, severity variation contributes to total-loss variation. With a Poisson claim count, random variation in the number of claims contributes as well. The two sources combine in the compound Poisson variance.

Condition on the claim count.
Given $N$, the aggregate is a sum of $N$ independent severities, so $E[S\mid N]=N\mu$ and $\operatorname{Var}(S\mid N)=N\sigma^2$.

Apply the laws of total expectation and variance.
$E[S]=E[N\mu]=\lambda\mu$. Also, $\operatorname{Var}(S)=E[\operatorname{Var}(S\mid N)]+\operatorname{Var}(E[S\mid N])$.

Use the Poisson moments.
Because $E[N]=\operatorname{Var}(N)=\lambda$, the variance is $\lambda\sigma^2+\lambda\mu^2=\lambda(\sigma^2+\mu^2)=\lambda E[X^2]$.

The relative variance of aggregate loss is therefore $\operatorname{Var}(S)/(E[S])^2=E[X^2]/(\lambda\mu^2)$. Since $E[X^2]=\sigma^2+\mu^2$, the ratio $E[X^2]/\mu^2$ equals $1+\operatorname{CV}^2$. Applying the normal approximation to the limited-fluctuation criterion gives the expected claim-count standard.

\frac{E[X^2]}{\mu^2} = \frac{\sigma^2+\mu^2}{\mu^2} = 1+\operatorname{CV}^2

Theorem: Under the compound Poisson assumptions above and a normal approximation to aggregate loss, the expected claim count $\lambda$ satisfies the central limited-fluctuation standard when $\lambda \ge (z/k)^2 E[X^2]/\mu^2=(z/k)^2(1+\operatorname{CV}^2)$. If a whole-number claim-count threshold is requested, round this lower bound up to the next integer. The unrounded value is the mathematical threshold for the Poisson mean.

Why the Formula Holds

The standard requires $P(|S-E[S]|\le kE[S])$ to be at least $p$. Approximate $S$ by a normal random variable with mean $\lambda\mu$ and variance $\lambda E[X^2]$. The distance from the mean to either tolerance boundary, measured in standard deviations, is $k\lambda\mu/\sqrt{\lambda E[X^2]}$. To achieve central probability $p$, this standardized distance must be at least $z$. Squaring both nonnegative sides and rearranging yields the threshold.

\frac{k\lambda\mu}{\sqrt{\lambda E[X^2]}}\ge z \quad\Longleftrightarrow\quad \lambda\ge \left(\frac{z}{k}\right)^2\frac{E[X^2]}{\mu^2} =\left(\frac{z}{k}\right)^2(1+\operatorname{CV}^2)

This derivation also explains why the mean severity cancels. If every severity is multiplied by the same positive constant, both aggregate loss and its mean are multiplied by that constant; a relative tolerance is unchanged. What matters is severity variability relative to its mean. A deterministic severity has $\operatorname{CV}=0$ and multiplier 1. Increasing the coefficient of variation increases the required expected count quadratically through $1+\operatorname{CV}^2$.

Worked Applications

Worked Example: Comparing Fixed and Variable Severity

Two hypothetical portfolios each have expected severity \$500. For Portfolio A, every claim is exactly \$500. For Portfolio B, severity is \$250 with probability 0.5 and \$750 with probability 0.5. Find the whole-number expected claim-count threshold for each portfolio under a central 90% standard with 10% tolerance. Use $z=1.645$.

Method. Find each severity second moment and apply the compound Poisson threshold. This comparison isolates the effect of severity variation because both portfolios have the same mean.

For Portfolio A, $\mu=500$, $\sigma^2=0$, and $E[X^2]=500^2=250{,}000$. Its severity multiplier is $250{,}000/500^2=1$. Thus the unrounded threshold is $(1.645/0.10)^2(1)=16.45^2=270.6025$. Rounding up gives a threshold of 271 expected claims.

For Portfolio B, the mean is $0.5(250)+0.5(750)=125+375=500$. The second moment is $0.5(250^2)+0.5(750^2)=0.5(62{,}500)+0.5(562{,}500)=31{,}250+281{,}250=312{,}500$. Its multiplier is $312{,}500/500^2=312{,}500/250{,}000=1.25$. The unrounded threshold is $270.6025(1.25)=338.253125$, so the whole-number threshold is 339 expected claims.

The check is that Portfolio B has positive severity variation while Portfolio A has none, and its multiplier, 1.25, exceeds 1. The resulting claim-count threshold is also higher. A common mistake is to give both portfolios the same threshold because their mean severities match; the second moment shows why that is incorrect.

Worked Example: A Severity Distribution With a Rare Large Claim

A hypothetical coverage has severity \$300 with probability 0.90 and \$2,100 with probability 0.10. Find the whole-number expected claim-count threshold for a central 95% aggregate-loss standard with 12% tolerance. Use $z=1.96$.

Method. Calculate the first and second severity moments, then use their ratio in the compound Poisson standard. The common mistake to avoid is rounding the moment ratio or the final threshold too early.

The mean severity is $\mu=0.90(300)+0.10(2{,}100)=270+210=480$. The second moment is $E[X^2]=0.90(300^2)+0.10(2{,}100^2)=0.90(90{,}000)+0.10(4{,}410{,}000)=81{,}000+441{,}000=522{,}000$. Therefore, the severity multiplier is $522{,}000/480^2=522{,}000/230{,}400=2.265625$.

The fluctuation factor is $(1.96/0.12)^2=(49/3)^2=2401/9\approx266.7777778$. The unrounded threshold is $(2401/9)(2.265625)=(2401/9)(145/64)=348{,}145/576\approx604.4184028$. Rounding up gives 605 expected claims.

As a check, the severity multiplier exceeds 1 because severities are not constant, so the threshold must exceed the corresponding deterministic-severity threshold of about 267 claims. The unrounded result is just over 604, so rounding to the nearest integer would be insufficient; the required whole-number threshold is 605.

Worked Example: Converting a Severity Standard Deviation to a Threshold

A hypothetical equipment policy has mean claim severity \$700 and severity standard deviation \$350. Determine the whole-number expected claim-count threshold for a central 90% standard with a 7.5% tolerance. Use $z=1.645$.

Method. Convert the standard deviation to a coefficient of variation, then apply the severity multiplier. This avoids confusing the standard deviation with the variance.

The coefficient of variation is $\operatorname{CV}=350/700=0.5$, so $1+\operatorname{CV}^2=1+0.25=1.25$. The fluctuation factor is $(1.645/0.075)^2=(21.9333333\ldots)^2\approx481.0711111$. Multiplying by the severity factor gives an unrounded threshold of $481.0711111(1.25)\approx601.3388889$. Rounding up gives 602 expected claims.

For a check, the same confidence and tolerance with deterministic severity would require an unrounded 481.0711 expected claims, so adding variability raises the threshold to about 601.34. A common error is to use $1+\operatorname{CV}$ rather than $1+\operatorname{CV}^2$; that would incorrectly produce a multiplier of 1.5 instead of 1.25.

Interpreting the Result and Its Limits

The severity multiplier is $1+\operatorname{CV}^2$, so a high mean alone does not imply a high count threshold. For instance, scaling every severity by a constant changes neither the coefficient of variation nor the required count under a relative tolerance. A heavier concentration of large outcomes relative to the mean, however, raises the second moment and the required volume.

The result relies on a compound Poisson model: claim severities are identically distributed and independent of one another and of the claim count, and the severity second moment is finite. The normal approximation is most useful when the aggregate distribution is adequately approximated by a normal distribution. It can be inaccurate for small expected counts or highly skewed, heavy-tailed severities. In those settings, the formula is a variance-based approximation, not an exact probability calculation.

The criterion is also a model-based fluctuation standard, not a guarantee that future losses will stay within the stated band. Dependence among claims, changing severity distributions, policy limits that alter the severity basis, or a non-Poisson claim count can change the relevant aggregate variance. The severity calculation must match the actual payment amount being aggregated.

Key takeaway: Under the compound Poisson normal approximation, the severity adjustment to the expected claim-count threshold is $1+\operatorname{CV}^2=E[X^2]/E[X]^2$. Greater relative severity variation means a larger required count, even when mean severity is unchanged.

Check Your Understanding

For each problem, use the stated central normal quantile and round a required whole-number expected claim count upward.

A hypothetical policy has constant severity \$800. What expected claim-count threshold meets a central 90% standard with 8% tolerance, using $z=1.645$? (Answer: 423 expected claims)
Claim severity is \$600 with probability 0.75 and \$1,800 with probability 0.25. Find the whole-number threshold for a central 95% standard with 20% tolerance, using $z=1.96$. (Answer: 129 expected claims)
A severity distribution has mean \$700 and standard deviation \$350. Find the whole-number threshold for a central 90% standard with 7.5% tolerance, using $z=1.645$. (Answer: 602 expected claims)
For a central 99% standard with 10% tolerance, compare these severities: Distribution A is \$400 or \$800 with equal probabilities; Distribution B is \$500 with probability 0.90 and \$1,500 with probability 0.10. Both means are \$600. Using $z=2.576$, find the whole-number thresholds and identify which distribution requires more expected claims. (Answer: A requires 738 claims; B requires 830 claims, so B requires more.)
A severity distribution has coefficient of variation 0.40. With 500 expected claims and a central 95% standard using $z=1.96$, what is the smallest relative tolerance that meets the approximate standard? Give the answer as a percentage to two decimal places. (Answer: 9.44%)

What You'll Learn