Normal Approximation To Aggregate Loss

Why Approximate an Aggregate Loss?

An aggregate loss can have a complicated distribution even when the claim count and individual claim amounts are simple. Exact convolution, recursion, or numerical methods may be available, but a normal approximation offers a quick way to estimate a probability once the aggregate mean and variance are known. Its usefulness depends on more than those two moments: the shape of the aggregate distribution matters, especially when a small number of large claims can dominate the total.

Let $N$ be the number of claims in a period, let $X_1,X_2,\ldots$ be independent identically distributed claim amounts, and assume the claim amounts are independent of $N$. The aggregate loss is $S_N=\sum_{j=1}^{N}X_j$, with $S_0=0$. Here $X_j$ denotes a severity, while $S_N$ denotes the total loss.

Definition: A moment-matched normal approximation replaces the distribution of $S_N$ by $N(E[S_N],\operatorname{Var}(S_N))$. For an approximate probability, standardize the relevant threshold using the aggregate mean and standard deviation, then evaluate the standard normal distribution function $\Phi$. The approximation is not an assertion that the aggregate is actually normal.

Aggregate Mean and Variance

For a random claim count independent of identically distributed severities with finite second moments, conditioning on $N$ gives the aggregate moments:

E[S_N]=E[N]E[X], \qquad \operatorname{Var}(S_N)=E[N]\operatorname{Var}(X)+\operatorname{Var}(N)(E[X])^2.

The variance has two components. The first reflects variation in claim amounts given the count; the second reflects variation in the count itself. Omitting either component generally understates or misstates aggregate variability.

For a compound Poisson model, $N\sim\operatorname{Poisson}(\lambda)$, so $E[N]=\operatorname{Var}(N)=\lambda$. Consequently,

E[S_N]=\lambda E[X], \qquad \operatorname{Var}(S_N)=\lambda E[X^2].

The second formula follows because $\operatorname{Var}(X)+(E[X])^2=E[X^2]$. In particular, it is the second raw moment of severity, not its variance alone, that appears in the compound Poisson aggregate variance.

Worked Example: A Compound Poisson Portfolio With Exponential Claims

A hypothetical portfolio has a Poisson claim count with $\lambda=100$ claims per year. Each claim amount is exponentially distributed with mean $\theta=\$1{,}000$, independently of the count and of the other claims. Use a normal approximation to estimate the probability that annual aggregate loss is at most $\$120{,}000$. State the aggregate moments and assess the approximation using skewness.

Method. First calculate the severity moments, then use the compound Poisson aggregate formulas. For an exponential severity with mean $\theta$, $E[X]=\theta$ and $E[X^2]=2\theta^2$.

The aggregate mean is

E[S_N]=100(1{,}000)=\$100{,}000.

The aggregate variance and standard deviation are

\operatorname{Var}(S_N)=100\bigl(2(1{,}000)^2\bigr) =200{,}000{,}000, \qquad \operatorname{SD}(S_N)=\sqrt{200{,}000{,}000} =\$14{,}142.14.

Standardizing the threshold gives

z=\frac{120{,}000-100{,}000}{14{,}142.14}=1.4142.

Thus the normal estimate is $P(S_N\leq \$120{,}000)\approx\Phi(1.4142)=0.9213$, to four decimal places.

For exponential severity, $E[X^3]=6\theta^3$. The compound Poisson aggregate skewness, derived below, is $3/(\sqrt{2}\sqrt{\lambda})$, which here is $3/(\sqrt{2}\sqrt{100})=0.2121$. This moderate positive skewness is a useful diagnostic, not a guarantee that every tail probability is accurate. The calculation is internally consistent: the threshold is above the mean, so the estimated probability is above one-half.

Continuity Correction for Discrete Aggregates

A normal distribution is continuous, but an aggregate formed from discrete claim amounts often takes values on a lattice, such as multiples of $\$1{,}000$. When approximating a discrete cumulative probability, a continuity correction moves the boundary by half a lattice unit. If the lattice spacing is $h$, then for an aggregate threshold $x$ on the lattice, approximate $P(S_N\leq x)$ using the normal probability up to $x+h/2$. The correction can improve the approximation; it does not make a poor-fitting normal model reliable.

Worked Example: A Discrete Compound Binomial Aggregate

A hypothetical insurer covers 80 independent exposures. Each exposure produces a claim with probability $0.4$, so $N\sim\operatorname{Binomial}(80,0.4)$. Every claim has a fixed amount of $\$1{,}000$, and claims are independent across exposures. Approximate the probability that aggregate loss is at most $\$37{,}000$, using a continuity correction.

Here $E[N]=80(0.4)=32$ and $\operatorname{Var}(N)=80(0.4)(0.6)=19.2$. Since every claim is $\$1{,}000$, $S_N=1{,}000N$. Therefore,

E[S_N]=\$32{,}000,\qquad \operatorname{Var}(S_N)=1{,}000^2(19.2)=19{,}200{,}000, \qquad \operatorname{SD}(S_N)=\$4{,}381.78.

The aggregate takes values in increments of $\$1{,}000$, so $h=\$1{,}000$. For the event $S_N\leq \$37{,}000$, use the corrected boundary $\$37{,}500$. Then

z=\frac{37{,}500-32{,}000}{4{,}381.78}=1.2552, \qquad P(S_N\leq \$37{,}000)\approx\Phi(1.2552)=0.8953.

The estimate is $0.8953$, to four decimal places. Using $\$37{,}000$ without the correction would instead use $z=1.1411$, which ignores the discrete boundary. The corrected value is an approximation; the exact event is $P(N\leq37)$.

Why Compound Poisson Skewness Matters

A normal distribution is symmetric. A compound Poisson aggregate is generally right-skewed, because a small number of unusually large claims can push the total far above its mean. For a compound Poisson model with finite third severity moment, its third cumulant is $\lambda E[X^3]$. Its variance is $\lambda E[X^2]$, so its standardized skewness is

\gamma_{S_N} =\frac{\lambda E[X^3]}{(\lambda E[X^2])^{3/2}} =\frac{E[X^3]}{\sqrt{\lambda}\,(E[X^2])^{3/2}}.

One justification uses the moment-generating function. When it exists near zero, the compound Poisson aggregate has cumulant-generating function $K_{S_N}(t)=\lambda(M_X(t)-1)$, where $M_X(t)=E[e^{tX}]$. Differentiating $r$ times at zero gives the $r$th cumulant $\kappa_r=\lambda E[X^r]$. In particular, $\kappa_2=\lambda E[X^2]$ and $\kappa_3=\lambda E[X^3]$. The standardized third cumulant $\kappa_3/\kappa_2^{3/2}$ is the skewness formula above.

For exponential severity with mean $\theta$, $E[X^2]=2\theta^2$ and $E[X^3]=6\theta^3$. Hence,

\gamma_{S_N} =\frac{\lambda 6\theta^3}{(\lambda 2\theta^2)^{3/2}} =\frac{3}{\sqrt{2}\sqrt{\lambda}}.

The skewness declines as the expected claim count increases, but even a relatively small skewness does not by itself establish accuracy at an extreme threshold. Severity tails, a small expected count, and whether the question concerns a central probability or a remote tail all matter.

Worked Example: A Low-Count Model With Noticeable Skewness

A hypothetical pool has $N\sim\operatorname{Poisson}(9)$ claims in a year. Severities are independent exponential amounts with mean $\theta=\$1{,}000$, and are independent of $N$. Use the matched normal model to estimate $P(S_N>\$15{,}000)$, then judge whether the estimate should be treated cautiously.

The aggregate mean is $9(\$1{,}000)=\$9{,}000$. Since $E[X^2]=2(\$1{,}000)^2$, its variance and standard deviation are

\operatorname{Var}(S_N)=9\bigl(2(1{,}000)^2\bigr)=18{,}000{,}000, \qquad \operatorname{SD}(S_N)=\$4{,}242.64.

The standardized threshold is

z=\frac{15{,}000-9{,}000}{4{,}242.64}=1.4142.

The normal estimate of the upper-tail probability is $1-\Phi(1.4142)=0.0787$, to four decimal places.

For these exponential severities, $E[X^3]=6\theta^3$ and $E[X^2]=2\theta^2$, giving

\gamma_{S_N} =\frac{9(6\theta^3)}{(9(2\theta^2))^{3/2}} =\frac{3}{\sqrt{2}\sqrt{9}} =\frac{3}{\sqrt{18}} =0.7071.

This is substantial positive skewness for a normal approximation, and the expected claim count is only 9. Treat $0.0787$ as a rough estimate, not a precise tail result. Also note that the compound Poisson aggregate has a point mass at zero, whereas a normal distribution does not. That distinction is especially important when estimating probabilities near zero.

Practical Checks and Common Traps

Check the model and independence assumptions.
Confirm that the frequency and severity moments are available and that the assumed independence between count and severities is appropriate for the formula being used.

Calculate both aggregate moments.
For a general independent frequency-severity model, use the count-variance and severity-variance components. For a compound Poisson model, use $\lambda E[X]$ and $\lambda E[X^2]$.

Match the probability event to its boundary.
Standardize with the aggregate standard deviation. If the aggregate is lattice-valued, consider a half-lattice-unit continuity correction in the appropriate direction.

Assess shape before relying on the result.
Check skewness and consider the claim count, severity tail, and location of the threshold. A normal approximation can be weak in a skewed upper tail even when its moments are correctly matched.

A frequent exam trap is to use the variance of an individual claim in place of its second raw moment for a compound Poisson aggregate. Another is to compute exponential-severity skewness incorrectly. The incorrect calculation $ \sqrt{2}/\sqrt{\lambda}$ does not follow from the third raw moment. The correct calculation is

\frac{\lambda E[X^3]}{(\lambda E[X^2])^{3/2}} =\frac{\lambda 6\theta^3}{(\lambda 2\theta^2)^{3/2}} =\frac{3}{\sqrt{2}\sqrt{\lambda}}.

At $\lambda=9$, this equals $3/\sqrt{18}=0.7071$, not $0.4714$. The error can materially understate the warning that the aggregate remains right-skewed. For aligned work on the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus, show the moment calculation before rounding and distinguish a diagnostic about approximation quality from a proof of accuracy.

Key takeaway: A normal approximation uses the aggregate mean and variance, but its credibility depends on distributional shape. For compound Poisson exponential claims, skewness is $3/(\sqrt{2}\sqrt{\lambda})$; a small expected count should prompt particular caution.

Check Your Understanding

For each problem, state the relevant moments or correction before giving the requested approximation or assessment.

A hypothetical compound Poisson portfolio has $\lambda=36$ and a fixed claim amount of $\$2{,}500$. Approximate $P(S_N\leq \$120{,}000)$, applying a continuity correction. (Answer: $E[S_N]=\$90{,}000$, $\operatorname{SD}(S_N)=\$15{,}000$, corrected $z=2.0833$, probability $0.9814$.)
For a compound Poisson model with $\lambda=25$ and independent exponential severities with mean $\$400$, approximate $P(S_N\leq \$14{,}000)$. (Answer: $E[S_N]=\$10{,}000$, $\operatorname{SD}(S_N)=\$2{,}828.43$, $z=1.4142$, probability $0.9213$.)
For compound Poisson frequency with $\lambda=16$ and exponential severity, calculate aggregate skewness and state what it suggests about a normal approximation. (Answer: $\gamma_{S_N}=3/(\sqrt{2}\sqrt{16})=0.5303$; the positive skewness warrants caution, particularly for upper-tail estimates.)
A hypothetical portfolio has $N\sim\operatorname{Binomial}(50,0.2)$ and fixed claim amounts of $\$1{,}000$. Approximate $P(S_N\geq \$12{,}000)$ using a continuity correction. (Answer: mean $\$10{,}000$, standard deviation $\$2{,}828.43$, corrected $z=(11{,}500-10{,}000)/2{,}828.43=0.5303$, upper-tail probability $0.2981$.)
A compound Poisson model has only 4 expected claims, and severity is highly right-skewed. Is matching the aggregate mean and variance enough to rely on a normal estimate of a very high loss threshold? (Answer: No. The small expected count and skewed severity indicate that the upper-tail normal approximation may be unreliable; matching two moments does not ensure tail accuracy.)

What You'll Learn