Full Credibility for Aggregate Loss

Why Aggregate-Loss Credibility Uses Both Frequency and Severity

A portfolio’s aggregate loss varies for two reasons: the number of claims can change, and the amounts of those claims can differ. A full-credibility standard for aggregate loss must therefore account for both sources of variation. A claim-count standard alone omits claim-size uncertainty; a severity standard for an average claim amount omits fluctuation in the number of claims.

The target is a claim volume large enough that aggregate loss is likely to fall within a specified percentage of its expected value. The resulting standard is a threshold for expected claim count, not a statement that every observed total within the threshold is adequate for every other actuarial purpose.

Definition: Let $N$ be the number of claims in a fixed portfolio period and let $X_j$ be the amount of the $j$th claim. Aggregate loss is $S=\sum_{j=1}^{N}X_j$, with $S=0$ when $N=0$. Under the compound Poisson model, $N\sim\operatorname{Poisson}(\lambda)$, the claim amounts are independent and identically distributed, and $N$ is independent of those amounts. The severity distribution has finite mean $\mu=E[X]$ and finite variance $\sigma^2=\operatorname{Var}(X)$.

Here, $\lambda=E[N]$ is the expected claim count. The full-credibility question is usually expressed using a relative error tolerance $k>0$ and a target probability $p$: how large must $\lambda$ be so that aggregate loss is within $k$ of its mean with probability at least $p$? The standard answer uses a normal approximation to the compound Poisson aggregate loss.

Deriving the Aggregate-Loss Standard

First calculate the mean and variance of aggregate loss. Conditional on $N=n$, the sum has mean $n\mu$ and variance $n\sigma^2$. Applying the laws of total expectation and total variance gives the compound Poisson moments:

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

The variance contains $\mu^2$ as well as $\sigma^2$. This is the contribution from the random claim count: even if individual claim amounts were identical, variation in $N$ would still make aggregate loss fluctuate.

Theorem: Under the compound Poisson assumptions, define the severity coefficient of variation by $CV=\sigma/\mu$, assuming $\mu>0$. Using a normal approximation for $S$, a full-credibility standard for relative error tolerance $k$ and central probability $p$ is $$ \lambda_{\text{full}} = \frac{z_{(1+p)/2}^{\,2}}{k^2} \left(1+CV^2\right) = \frac{z_{(1+p)/2}^{\,2}}{k^2} \frac{E[X^2]}{(E[X])^2}, $$ where $z_{(1+p)/2}$ is the corresponding standard normal quantile. The required expected claim count is at least this value; if a whole-claim threshold is required, round upward.

To see why the formula holds, approximate $S$ by a normal random variable with mean $\lambda\mu$ and variance $\lambda(\sigma^2+\mu^2)$. The condition that aggregate loss be within $k$ of its mean is $P(|S-\lambda\mu|\leq k\lambda\mu)\geq p$. Standardizing the endpoints gives the following condition:

\frac{k\lambda\mu} {\sqrt{\lambda(\sigma^2+\mu^2)}} \geq z_{(1+p)/2} \quad\Longleftrightarrow\quad \lambda \geq \frac{z_{(1+p)/2}^{\,2}}{k^2} \frac{\sigma^2+\mu^2}{\mu^2}.

Since $(\sigma^2+\mu^2)/\mu^2=1+CV^2$, this is the stated standard. The derivation also shows why aggregate-loss variability cannot be obtained by using the severity variance alone.

Applying the Standard

For each application, identify the aggregate quantity, determine the claim-size second moment or coefficient of variation, and then apply the selected confidence and error tolerance. The normal approximation is most useful when the aggregate is sufficiently regular for a normal approximation to be reasonable. It is not an exact probability calculation for every compound Poisson distribution, particularly when expected counts are small or severities are highly skewed or heavy-tailed.

Worked Example: Exponential Claim Amounts and a 95% Target

A hypothetical equipment-repair portfolio has a Poisson claim count with expected value $\lambda$. Individual repair amounts are independent exponential random variables with mean $1,000. Find the expected claim count needed for aggregate loss to be within 10% of its mean with approximate probability 95%. If the portfolio produces an expected 0.04 claims per policy-year, find the corresponding policy-year volume.

Model and inputs. The count is Poisson, and severities are independent exponential amounts with mean $\mu=$1,000$. An exponential distribution with mean $\mu$ has variance $\sigma^2=\mu^2$, so $CV=1$ and $1+CV^2=2$. Losses are undiscounted; no interest rate or mortality assumption applies. Use $k=0.10$ and the 95% two-sided normal quantile $z_{0.975}=1.96$.

Apply the aggregate standard. The expected claim count threshold is

\lambda_{\text{full}} = \frac{1.96^2}{0.10^2}(2) = \frac{3.8416}{0.01}(2) = 768.32.

Thus the threshold is 768.32 expected claims. If a whole-claim threshold is required, round upward to 769. At $\lambda=769$, the normal-approximation standardized tolerance is $0.10\sqrt{769/2}\approx1.961$, just above 1.96, which checks the rounding direction.

Convert to exposure. At 0.04 expected claims per policy-year, the exposure needed for 769 expected claims is $769/0.04=19{,}225$ policy-years. The full-credibility threshold is therefore 769 expected claims, corresponding to 19,225 policy-years at the stated frequency.

Method and mistake avoided. This uses the compound Poisson aggregate-loss standard. Using the severity variability factor alone would miss the additional aggregate fluctuation from the random count.

Worked Example: A Discrete Severity Distribution

A hypothetical glass-cover portfolio has Poisson claim count with expected value $\lambda$. Each claim is $500 with probability 0.8 and $2,500 with probability 0.2, independently of all other claims and of the count. Find the full-credibility expected claim count for aggregate loss within 5% of its mean with approximate probability 90%.

Model and inputs. Claim amounts have a two-point distribution with the probabilities specified above. Losses are undiscounted; no interest rate or mortality assumption applies. Use $k=0.05$ and $z_{0.95}=1.645$.

Find the severity moments. The mean claim amount is $E[X]=0.8($500)+0.2($2{,}500)=$900$. The second moment is $E[X^2]=0.8($500^2)+0.2($2{,}500^2)=$1{,}450{,}000$. Hence $\mu^2=$810{,}000$, $\sigma^2=E[X^2]-\mu^2=$640{,}000$, and

1+CV^2 = \frac{E[X^2]}{\mu^2} = \frac{1{,}450{,}000}{810{,}000} \approx1.790123.

Calculate the standard.

\lambda_{\text{full}} = \frac{1.645^2}{0.05^2}(1.790123) = \frac{2.706025}{0.0025}(1.790123) \approx1{,}937.65.

The whole-claim threshold is 1,938 expected claims. As a check, at $\lambda=1{,}938$, the standardized tolerance is $0.05\sqrt{1{,}938/1.790123}\approx1.645$, matching the selected normal quantile to the displayed precision.

Method and mistake avoided. The second-moment form of the aggregate standard was used. It would be incorrect to substitute $CV^2$ alone: the correct multiplier is $1+CV^2$, which includes count fluctuation.

Worked Example: Gamma Severities and Exposure Volume

A hypothetical specialty-liability portfolio has a Poisson claim count. Individual losses follow a gamma distribution with shape parameter $\alpha=16/9$ and scale parameter $\theta=$1{,}125$. Determine the claim-count threshold for aggregate loss to be within 8% of its mean with approximate probability 99%. If the expected frequency is 0.04 claims per policy-year, convert the threshold to policy-years.

Model and inputs. Under the shape-scale parameterization, a gamma severity has mean $\alpha\theta$ and variance $\alpha\theta^2$. Thus $\mu=(16/9)($1{,}125)=$2{,}000$ and $\sigma^2=(16/9)($1{,}125)^2=$2{,}250{,}000$. Losses are undiscounted; no interest rate or mortality assumption applies. Use $k=0.08$ and $z_{0.995}=2.576$.

Calculate the severity factor. Since $\mu^2=$4{,}000{,}000$,

1+CV^2 = 1+\frac{2{,}250{,}000}{4{,}000{,}000} = 1.5625.

Calculate the claim-count threshold.

\lambda_{\text{full}} = \frac{2.576^2}{0.08^2}(1.5625) = \frac{6.635776}{0.0064}(1.5625) = 1{,}620.0625.

Round upward to 1,621 expected claims. At that threshold, $0.08\sqrt{1{,}621/1.5625}\approx2.577$, which exceeds the selected quantile. At the unrounded threshold, the standardized value is exactly 2.576.

Convert to exposure. The exposure required is $1{,}621/0.04=40{,}525$ policy-years. The result is 1,621 expected claims, or 40,525 policy-years at the specified frequency.

Method and mistake avoided. The gamma moments are converted to the second-moment multiplier before applying the aggregate standard. Treating the severity variance as the aggregate variance would omit the contribution from random claim count.

Assumptions, Interpretation, and a Common Trap

The calculation depends on a compound Poisson model: the count is Poisson, claim amounts are identically distributed and independent of one another, and claim amounts are independent of the count. The severity must have a finite, positive mean and finite variance. If claim counts are fixed rather than random, or if the count model is not Poisson, the aggregate variance and resulting standard may differ. Dependence among claims, changing claim-size distributions, and material trends within the period also require appropriate modeling rather than uncritical use of this formula.

The normal approximation is a modeling approximation, not an exact guarantee. The stated probability is approximate, and the required volume is a threshold under the specified assumptions. For highly skewed severities or limited expected counts, the approximation may be poor; a model-specific probability calculation or simulation may be more appropriate.

Common exam trap: Applying the severity-only variability factor $CV^2$ rather than the aggregate factor $1+CV^2$. For exponential severity, $CV=1$. With 95% confidence and 10% tolerance, the incorrect calculation gives $1.96^2(1)/0.10^2=384.16$ expected claims. The correct standard is $1.96^2(2)/0.10^2=768.32$. At 384.16 claims, the correct standardized tolerance is only $0.10\sqrt{384.16/2}\approx1.386$, below 1.96. The error nearly halves the required volume.

The aggregate standard matters when the quantity being evaluated is the total loss rather than a claim count or average claim amount. Its multiplier makes the role of severity explicit: as claim-size variation increases, the required expected count increases, even when the mean severity stays fixed. The standard is aligned with the CAS Exam 5 (Basic Ratemaking and Estimated Claim Liabilities) syllabus.

Check Your Understanding

For each problem, use the compound Poisson normal-approximation standard and round a whole-claim threshold upward.

A hypothetical portfolio has Poisson claim count and exponential claim amounts with mean $800. Find the full-credibility expected claim count for a 10% tolerance and 95% probability, using $z_{0.975}=1.96$. (Answer: 769 expected claims)
Claim amounts are $1,000 with probability 0.75 and $3,000 with probability 0.25. Find the threshold for aggregate loss within 10% of its mean with 90% probability, using $z_{0.95}=1.645$. (Answer: 361 expected claims)
Claim severity has mean $2,000 and coefficient of variation 1.5. Find the threshold for 5% tolerance and 95% probability, using $z_{0.975}=1.96$. (Answer: 4,995 expected claims)
Severity has a gamma distribution with shape 4 and scale $500. Find the threshold for 5% tolerance and 99% probability, using $z_{0.995}=2.576$. (Answer: 3,318 expected claims)
A full-credibility aggregate-loss standard is 800 expected claims. If the expected frequency is 0.08 claims per exposure-year, how many exposure-years are required to reach that standard? (Answer: 10,000 exposure-years)

What You'll Learn