Comparing Frequency, Severity, and Aggregate Loss Standards

Why the Three Standards Differ

A credibility standard describes how much experience is needed for an estimate to meet a specified precision target. Frequency, average severity, and aggregate loss are different estimates, so their variability—and therefore their required data volume—can differ even when the confidence level and relative error tolerance are the same. A useful comparison holds those targets fixed and identifies how each estimate responds to claim-count and claim-size variation.

The comparison below uses classical full-credibility standards based on a normal approximation. These are precision rules under stated assumptions, not guarantees that every observed portfolio will meet the target. The credibility material on full standards for claim counts, average severity, aggregate loss, and loss ratios provides the individual methods; the goal here is to place frequency, severity, and aggregate loss on a common basis.

Definition: Let $E$ be exposure, $\lambda$ the expected claim frequency per exposure unit, and $N$ the claim count. Assume $N\sim\operatorname{Poisson}(\lambda E)$. Let $X$ be the payment on one claim, with mean $m=E[X]>0$, variance $\sigma^2$, and coefficient of variation $\operatorname{CV}=\sigma/m$. Assume claim amounts are independent and identically distributed, independent of $N$, and have finite variance. The aggregate loss is $S=\sum_{j=1}^{N}X_j$. For confidence level $p$, let $z_p$ be the positive standard normal quantile satisfying $P(-z_p\le Z\le z_p)=p$, where $Z\sim N(0,1)$. Let $r$ be the allowed relative error.

The full-credibility condition used throughout is that the estimate lie within $r$ of its expected value with approximate probability at least $p$. In the normal approximation, an estimator with mean $\theta$ and variance $V$ meets this condition when $z_p\sqrt{V}\le r\theta$. The estimates considered here are the frequency $N/E$, the average severity $\bar X$, and the aggregate loss per exposure $S/E$.

Standards on a Common Claim-Count Scale

For frequency, the estimated frequency is $N/E$, its expected value is $\lambda$, and its variance is $\lambda/E$. Substituting into the relative-error condition gives an exposure requirement. Multiplying that exposure requirement by $\lambda$ expresses it as required expected claims, which makes comparison with severity and aggregate loss direct.

\begin{aligned} \operatorname{Var}\left(\frac{N}{E}\right)&=\frac{\lambda}{E},\\ \frac{z_p\sqrt{\lambda/E}}{\lambda}&\le r \quad\Longrightarrow\quad E\ge \frac{z_p^2}{r^2\lambda},\\ \lambda E&\ge \frac{z_p^2}{r^2}. \end{aligned}

For severity, with $n$ independent claim payments, the estimate is the sample mean $\bar X$. Its mean is $m$ and variance is $\sigma^2/n$. Applying the same condition gives the required number of claims.

\frac{z_p\sqrt{\sigma^2/n}}{m}\le r \quad\Longrightarrow\quad n\ge \frac{z_p^2\sigma^2}{r^2m^2} =\frac{z_p^2\operatorname{CV}^2}{r^2}.

Aggregate loss incorporates both random claim counts and random claim sizes. Under the compound Poisson assumptions above, the expected aggregate loss per exposure is $\lambda m$, and its variance per exposure is $\lambda(\sigma^2+m^2)/E$. The extra $m^2$ term is the contribution of claim-count variation.

\frac{z_p\sqrt{\lambda(\sigma^2+m^2)/E}}{\lambda m}\le r \quad\Longrightarrow\quad \lambda E\ge \frac{z_p^2}{r^2} \left(1+\operatorname{CV}^2\right).

Why Aggregate Loss Includes Both Sources of Variation

The aggregate-loss variance follows by conditioning on the claim count. Given $N$, the sum $S$ has conditional mean $Nm$ and conditional variance $N\sigma^2$. The law of total variance states that $\operatorname{Var}(S)=E[\operatorname{Var}(S\mid N)]+\operatorname{Var}(E[S\mid N])$. Since a Poisson variable with mean $\lambda E$ has both mean and variance $\lambda E$, the two terms can be evaluated directly.

\begin{aligned} E[S]&=E[Nm]=\lambda E m,\\ \operatorname{Var}(S) &=E[N\sigma^2]+\operatorname{Var}(Nm)\\ &=\lambda E\sigma^2+m^2\lambda E\\ &=\lambda E(\sigma^2+m^2). \end{aligned}

Dividing the variance of $S$ by $E^2$ gives the variance of $S/E$, and dividing by its mean $\lambda m$ in the relative-error condition yields the aggregate standard above. Equivalently, its expected claim-count requirement is the frequency requirement plus the severity requirement:

\frac{z_p^2}{r^2}(1+\operatorname{CV}^2) = \underbrace{\frac{z_p^2}{r^2}}_{\text{frequency}} + \underbrace{\frac{z_p^2\operatorname{CV}^2}{r^2}}_{\text{severity}}.

This addition is a consequence of the compound Poisson model and the specified aggregate-loss target. It does not mean that every other measure of total loss has this variance or standard. For example, dependence between claim count and claim size, changing risk mix, or a frequency model that is not Poisson can change the result.

Estimate	Required expected claims	What drives its variability?
Frequency	$z_p^2/r^2$	Random claim count
Average severity	$z_p^2\operatorname{CV}^2/r^2$	Variation in claim payments
Aggregate loss per exposure	$z_p^2(1+\operatorname{CV}^2)/r^2$	Random count and random payment

Worked Comparisons

Worked Example: Comparing All Three Standards

A hypothetical book has expected frequency $\lambda=0.40$ claims per exposure unit. Claim payments have mean $m=\$4{,}000$ and standard deviation $\sigma=\$6{,}000$. Find the expected-claim and exposure standards for frequency, average severity, and aggregate loss at 95% confidence with a 10% relative error tolerance. Use $z_p=1.96$.

Method and assumptions. Use the Poisson-frequency and independent, identically distributed severity models defined above. First calculate $\operatorname{CV}=6{,}000/4{,}000=1.5$, so $\operatorname{CV}^2=2.25$. The common precision factor is $z_p^2/r^2=1.96^2/0.10^2=3.8416/0.01=384.16$.

Frequency. The expected-claim standard is $384.16$ claims. Since the expected frequency is $0.40$ per exposure unit, the exposure standard is $384.16/0.40=960.40$ exposure units.

Average severity. The required claim count is $384.16(2.25)=864.36$ claims. The corresponding exposure volume at the stated frequency is $864.36/0.40=2{,}160.90$ exposure units.

Aggregate loss. The required expected claim count is $384.16(1+2.25)=384.16(3.25)=1{,}248.52$ claims. The corresponding exposure volume is $1{,}248.52/0.40=3{,}121.30$ exposure units.

Check and common mistake avoided. The aggregate standard, $1{,}248.52$, equals the frequency standard plus the severity standard: $384.16+864.36=1{,}248.52$. Omitting the frequency contribution would incorrectly give only $864.36$ aggregate-standard claims. The exposure figures are obtained by dividing each expected-claim standard by $0.40$, not by multiplying.

Worked Example: How Severity Variability Changes the Comparison

Two hypothetical portfolios each have frequency $\lambda=0.50$ claims per exposure unit. Portfolio A has severity coefficient of variation $0.50$; Portfolio B has coefficient of variation $2.00$. For each portfolio, calculate the claim-count and exposure standards for frequency, average severity, and aggregate loss at 90% confidence with a 20% relative error tolerance. Use $z_p=1.644853626951$.

Method and precision factor. Keep the full stated quantile through the calculations. Its square is approximately $2.705543454$, so the common factor is $2.705543454/0.20^2=2.705543454/0.04=67.63858635$.

Frequency, both portfolios. The required expected claim count is $67.63858635$. At frequency $0.50$, the exposure standard is $67.63858635/0.50=135.2771727$.

Portfolio A severity. The severity claim-count standard is $67.63858635(0.50^2)=67.63858635(0.25)=16.90964659$. Its exposure standard is $16.90964659/0.50=33.81929318$.

Portfolio A aggregate loss. The aggregate claim-count standard is $67.63858635(1+0.25)=84.54823294$. Its exposure standard is $84.54823294/0.50=169.0964659$.

Portfolio B severity. The severity claim-count standard is $67.63858635(2.00^2)=67.63858635(4)=270.5543454$. Its exposure standard is $270.5543454/0.50=541.1086908$.

Portfolio B aggregate loss. The aggregate claim-count standard is $67.63858635(1+4)=338.1929318$. Its exposure standard is $338.1929318/0.50=676.3858635$.

Check and common mistake avoided. For each portfolio, aggregate claims equal frequency claims plus severity claims: $67.63858635+16.90964659=84.54823294$, and $67.63858635+270.5543454=338.1929318$. Replacing the high-variability portfolio’s $\operatorname{CV}^2=4$ with $\operatorname{CV}=2$ in the standard would understate its severity requirement.

Worked Example: A High-Precision Aggregate Standard

A hypothetical specialty program has frequency $\lambda=0.10$ claims per exposure unit. Severity has mean $\$2{,}500$ and standard deviation $\$5{,}000$. Find the expected-claim and exposure standards for frequency, severity, and aggregate loss at 99% confidence and 5% relative error tolerance. Use $z_p=2.576$.

Method and assumptions. The coefficient of variation is $5{,}000/2{,}500=2$, so its square is $4$. The common factor is $2.576^2/0.05^2=6.635776/0.0025=2{,}654.3104$.

Frequency. The standard is $2{,}654.3104$ expected claims, requiring $2{,}654.3104/0.10=26{,}543.104$ exposure units.

Severity. The standard is $2{,}654.3104(4)=10{,}617.2416$ claims, requiring $10{,}617.2416/0.10=106{,}172.416$ exposure units.

Aggregate loss. The standard is $2{,}654.3104(1+4)=13{,}271.552$ claims, requiring $13{,}271.552/0.10=132{,}715.52$ exposure units.

Check and common mistake avoided. The aggregate claim standard equals $2{,}654.3104+10{,}617.2416=13{,}271.552$. Rounding the common factor before multiplying would needlessly change the reported results; the displayed values are rounded only at the final step, to two decimal places.

Conditions, Interpretation, and Exam Traps

The formulas are most useful when their assumptions fit the experience being measured. They assume a homogeneous frequency across exposure, Poisson claim counts, independent claim payments with a stable distribution, and an independent relationship between count and severity. The normal approximation is a further condition for translating variance into a confidence-and-tolerance standard. It may be poor when expected counts are small or when severity is extremely skewed, even if the variance exists. In such cases, these standards are not a substitute for evaluating the model or using an appropriate distributional calculation.

A frequent error is to use the severity formula for aggregate loss. That omits the frequency component: aggregate loss varies when the count changes even if every payment is exactly the same. The converse error is to assume that aggregate loss always needs more claims than severity. The formulas show that this is not necessarily true: aggregate requires $z_p^2(1+\operatorname{CV}^2)/r^2$ claims, while severity requires $z_p^2\operatorname{CV}^2/r^2$, so aggregate exceeds severity by the frequency standard. But severity can require fewer claims than frequency when $\operatorname{CV}<1$, and more when $\operatorname{CV}>1$.

Finally, keep the units straight. The frequency formula first produces expected claims, while exposure is expected claims divided by $\lambda$. Severity and aggregate formulas produce claim-count standards; converting them to exposure also requires the expected frequency. A fractional standard is a mathematical threshold in expected volume, not a promise that a particular observed claim count will be achieved.

Key takeaway: At a common confidence level and relative error tolerance, frequency requires $z_p^2/r^2$ expected claims, severity requires $z_p^2\operatorname{CV}^2/r^2$, and compound Poisson aggregate loss requires their sum. The aggregate standard includes both count variability and payment variability.

Check Your Understanding

Use the stated normal quantiles and keep intermediate precision until the final answer. Assume the models and conditions in this tutorial unless a question says otherwise.

At 95% confidence, with $z_p=1.96$ and $r=20\%$, what is the full-credibility expected-claim standard for frequency? (Answer: 96.04 expected claims.)
At 90% confidence, use $z_p=1.645$. A severity distribution has $\operatorname{CV}=1.25$, and $r=20\%$. Find its full-credibility claim-count standard. (Answer: 105.70 claims.)
At 95% confidence, use $z_p=1.96$. For $\operatorname{CV}=0.75$ and $r=10\%$, find the aggregate-loss expected-claim standard. (Answer: 600.25 expected claims.)
At 90% confidence, use $z_p=1.645$. A hypothetical portfolio has $\lambda=0.20$ claims per exposure unit, $\operatorname{CV}=1.50$, and $r=10\%$. Find the aggregate expected-claim standard and its exposure equivalent. (Answer: 879.46 expected claims and 4,397.29 exposure units.)
For a compound Poisson model, explain how the aggregate standard compares with the frequency and severity standards, and identify the assumption that allows their expected-claim standards to be added. (Answer: Aggregate equals frequency plus severity, $z_p^2/r^2+z_p^2\operatorname{CV}^2/r^2$; this follows from Poisson count variation and independent, identically distributed severities independent of the count.)

Estimate	Required expected claims	What drives its variability?
Frequency	\(z_p^2/r^2\)	Random claim count
Average severity	\(z_p^2\operatorname{CV}^2/r^2\)	Variation in claim payments
Aggregate loss per exposure	\(z_p^2(1+\operatorname{CV}^2)/r^2\)	Random count and random payment

What You'll Learn