Normal Approximation for Aggregate Loss

Why Approximate Aggregate Loss?

An aggregate loss is the sum of the payments arising from a random number of claims. Exact probabilities can require a detailed convolution or recursion, and an aggregate distribution may have many possible values. When the expected claim count is sufficiently large and no one claim dominates the total, a normal distribution can provide a quick approximation using only the aggregate mean and variance.

The approximation is useful for estimating a tail probability or a high aggregate-loss threshold. It is not a replacement for checking whether a normal shape is plausible. A low claim count, highly skewed severity, or a coarse discrete payment structure can make the approximation misleading, especially in the tail.

Definition: Let $N$ be the number of claims in a fixed period, and let $Y_1,Y_2,\ldots$ be the claim payments. The aggregate loss is $S=\sum_{j=1}^{N}Y_j$, with $S=0$ when $N=0$. A normal approximation replaces the distribution of $S$ by a normal distribution with the same mean and variance: $$ S \ \dot{\sim}\ N\big(E[S],\operatorname{Var}(S)\big). $$ Here the second argument is the variance, not the standard deviation. The symbol $\dot{\sim}$ indicates an approximation.

Aggregate Moments to Use

Assume the claim count $N$ is independent of the severities, and that, conditional on $N$, claim payments are independent and identically distributed with finite mean $\mu_Y=E[Y]$ and variance $\sigma_Y^2=\operatorname{Var}(Y)$. These assumptions give the general aggregate moments below. Severity may include policy modifications, provided $Y$ represents the payment being aggregated.

E[S]=E[N]\mu_Y,\qquad \operatorname{Var}(S)=E[N]\sigma_Y^2+\operatorname{Var}(N)\mu_Y^2.

For a Poisson claim count with parameter $\lambda$, $E[N]=\operatorname{Var}(N)=\lambda$. Therefore the variance simplifies to $\lambda E[Y^2]$, not merely $\lambda\operatorname{Var}(Y)$. For other frequency models, retain both terms in the general variance formula.

Model conditions: The formulas above require independence between frequency and severity, and independent, identically distributed severities that do not depend on the claim count. If these assumptions fail, the displayed moments need not be correct. A normal approximation also requires a distributional justification beyond matching moments; the moment formulas alone do not establish that the aggregate distribution is approximately normal.

Why the Moment Formula Holds

Conditional on $N=n$, the aggregate is a sum of $n$ independent severities. Thus $E[S\mid N=n]=n\mu_Y$ and $\operatorname{Var}(S\mid N=n)=n\sigma_Y^2$. Applying the laws of total expectation and total variance gives a short derivation:

Condition on the claim count.
$E[S\mid N]=N\mu_Y$ and $\operatorname{Var}(S\mid N)=N\sigma_Y^2$, by independence of the severities.

Take the expectation.
$E[S]=E[E[S\mid N]]=E[N]\mu_Y$.

Separate the sources of variance.
$\operatorname{Var}(S)=E[\operatorname{Var}(S\mid N)]+\operatorname{Var}(E[S\mid N])=E[N]\sigma_Y^2+\operatorname{Var}(N)\mu_Y^2$.

The two variance terms have distinct roles. The first reflects variation in claim sizes at a given count; the second reflects variation in the count itself. For Poisson frequency, their sum is $\lambda(\sigma_Y^2+\mu_Y^2)=\lambda E[Y^2]$, since $E[Y^2]=\sigma_Y^2+\mu_Y^2$.

Using the Approximation

Write $\mu_S=E[S]$ and $\sigma_S=\sqrt{\operatorname{Var}(S)}$. The continuous normal approximation estimates a tail probability by standardizing the threshold:

P(S>x)\ \approx\ 1-\Phi\!\left(\frac{x-\mu_S}{\sigma_S}\right),

where $\Phi$ is the standard normal cumulative distribution function. A corresponding approximate $p$-quantile is $\mu_S+z_p\sigma_S$, where $z_p$ satisfies $\Phi(z_p)=p$. If payments lie on a regular grid, a continuity correction may improve a probability calculation. For example, for a lattice spacing $h$, approximate $P(S\geq x)$ using the normal probability above $x-h/2$, when $x$ is on that lattice. The correction is an approximation aid, not a change to the underlying event.

Worked Example: A Poisson Portfolio Quantile

A hypothetical portfolio has $N\sim\operatorname{Poisson}(100)$ claims in one year. Independently, each claim payment is $1,000 or $3,000, each with probability $0.5$. Estimate the 95th percentile of annual aggregate payments. Report the result in thousands of dollars to four decimal places. Use $z_{0.95}=1.6449$.

Method. Find the compound Poisson mean and variance, then use the normal quantile formula. Payments are stated in thousands of dollars, so let $Y$ take values 1 and 3 in those units.

First calculate the severity moments:

E[Y]=0.5(1)+0.5(3)=2,\qquad E[Y^2]=0.5(1^2)+0.5(3^2)=5.

For Poisson frequency, the aggregate mean is $100(2)=200$ thousand dollars, and its variance is $100(5)=500$ thousand-dollars squared. Thus $\sigma_S=\sqrt{500}=22.3606798$ thousand dollars. The requested approximate quantile is

200+1.6449\sqrt{500} =200+36.7811 =236.7811\text{ thousand dollars}.

The approximation is therefore about $236,781.10. This is a continuous normal estimate, not a claim that the aggregate can take every value near that amount. The common mistake avoided here is using $\operatorname{Var}(Y)$ instead of $E[Y^2]$ in the compound Poisson aggregate variance. That error would give $100(1)=100$, rather than 500, thousand-dollars squared, and substantially understate the spread.

Check. The mean is $200,000, and the estimated 95th percentile is above the mean, as it should be. The calculation used the displayed $z$-value consistently; all intermediate amounts are in thousands.

Tail Probabilities and Continuity Corrections

For a discrete aggregate, the event at a threshold must be translated carefully before applying the normal approximation. A strict inequality and a weak inequality can differ by a full payment increment. If all possible aggregate payments are multiples of a fixed amount, place the normal boundary halfway between adjacent attainable values. For non-lattice or irregular payments, a single continuity correction may not be well-defined.

Worked Example: A Discrete Aggregate Tail

A hypothetical insurer covers 400 independent exposures. Each produces one claim with probability $0.10$, and no claim otherwise, so $N\sim\operatorname{Binomial}(400,0.10)$. Every claim pays exactly $2,500. Estimate the probability that aggregate payments exceed $130,000, using a continuity correction. Give the probability to four decimals.

Method. Since the payment is fixed, $S=2{,}500N$. The frequency moments are $E[N]=400(0.10)=40$ and $\operatorname{Var}(N)=400(0.10)(0.90)=36$. Therefore $E[S]=2{,}500(40)=100{,}000$, and $\operatorname{Var}(S)=2{,}500^2(36)=225{,}000{,}000$, giving $\sigma_S=15{,}000$.

The event $S>130{,}000$ means $2{,}500N>130{,}000$, or $N>52$, equivalently $N\geq53$. The count is integer-valued, so the continuity-corrected boundary is 52.5 claims, corresponding to $131,250. Hence

z=\frac{131{,}250-100{,}000}{15{,}000} =\frac{31{,}250}{15{,}000} =2.0833, \qquad P(S>130{,}000)\approx 1-\Phi(2.0833)=0.0186.

Check. The threshold is more than two standard deviations above the mean, so a tail probability below 0.05 is reasonable. A common error is to standardize $130,000 directly. That ignores the fact that the event is $N\geq53$, whose continuity-corrected boundary is $131,250, not $130,000.

When the Normal Approximation Is Weak

A normal distribution is symmetric and continuous. Aggregate loss, by contrast, is nonnegative and may have a large probability at zero, substantial right skew, or many widely spaced attainable values. A high expected count often helps the approximation, but it is not a guarantee: very variable severities or a small number of dominant claims can preserve a strongly skewed aggregate distribution. Tail estimates are usually more sensitive to these features than estimates near the center.

The next example compares a normal approximation with an exact probability for a deliberately small, highly discrete portfolio. It illustrates a diagnostic, not a rule that the corrected approximation must always be close.

Worked Example: Recognizing a Poor Approximation

A hypothetical portfolio has $N\sim\operatorname{Poisson}(2)$ claims. Every claim pays exactly $1,000. Approximate the probability that aggregate payments are at most $1,000 using a continuity-corrected normal distribution, then compare it with the exact probability. Give probabilities to four decimals.

Method. The aggregate is $S=1{,}000N$. The Poisson aggregate mean is $2(1{,}000)=2{,}000$, and its variance is $2(1{,}000^2)=2{,}000{,}000$. Thus the standard deviation is $\sqrt{2{,}000{,}000}=1{,}414.2136$.

The event $S\leq1{,}000$ is $N\leq1$. With lattice spacing $1,000, the continuity-corrected boundary is $1,500:

z=\frac{1{,}500-2{,}000}{1{,}414.2136}=-0.3536, \qquad P(S\leq1{,}000)\approx\Phi(-0.3536)=0.3618.

The exact probability is available from the Poisson count:

P(N\leq1)=P(N=0)+P(N=1) =e^{-2}+2e^{-2}=3e^{-2}=0.4060.

The approximation is low by about $0.4060-0.3618=0.0442$, or 4.42 percentage points. With only two expected claims, the aggregate has substantial mass at zero and at $1,000; it cannot resemble a smooth symmetric distribution particularly well. The common mistake avoided is treating a continuity correction as a cure for poor shape. It adjusts a boundary but cannot remove skewness or concentrated probability masses.

Check. The exact probability is greater than the normal estimate, consistent with the large probability of zero or one claim in this low-frequency model. The discrepancy is itself a warning against relying on the normal estimate without qualification.

Practical Checks Before Reporting an Estimate

A normal approximation should be presented as an estimate, with its assumptions and limitations visible. In a fitted model, the moments may be plug-in estimates derived from estimated frequency and severity parameters. Using those fitted moments treats the inputs as fixed; it does not automatically account for parameter uncertainty.

Check the model chain. Confirm that the frequency and severity assumptions match the aggregate being approximated, including any payment limits or deductibles already reflected in severity.
Check units. If severity is measured in thousands of dollars, its variance is measured in squared thousands; convert consistently before combining moments.
Check event boundaries. Translate $S>x$, $S\geq x$, or $S\leq x$ into the corresponding attainable claim counts or payment amounts before applying a continuity correction.
Check plausibility. A negative normal value has no counterpart for a nonnegative aggregate. A substantial chance of negative values, strong right skew, low frequency, or a few dominant severities signals caution.
Check what the result means. A moment-matched normal approximation is not an exact aggregate distribution and may be especially inaccurate in extreme tails.

Key takeaway: Compute the aggregate mean and variance from the stated frequency-severity model, standardize the desired threshold, and use a continuity correction when a regular discrete lattice makes one appropriate. Treat the result cautiously when low frequency, skewness, or concentrated discrete mass conflicts with the smooth symmetric normal shape.

Check Your Understanding

For each problem, state the aggregate moments needed and apply a normal approximation only as requested. Answers use the assumptions given in each hypothetical model.

A hypothetical portfolio has Poisson claim count with mean 64, and each claim pays exactly $500. Estimate the 90th percentile of aggregate loss, using $z_{0.90}=1.2816$. (Answer: $37,126.40.)
Claim count is Poisson with mean 25. Each payment is independently $500 or $1,500, each with probability 0.5. Estimate $P(S>$35,000)$ using a continuity correction and give four decimals. (Answer: approximately 0.0333.)
A hypothetical book has $N\sim\operatorname{Binomial}(200,0.20)$, with a fixed $1,000 payment per claim. Estimate $P(S\geq$45,000)$ using a continuity correction and give four decimals. (Answer: approximately 0.2132.)
A compound Poisson model has mean claim count 9 and fixed payment $2,000 per claim. Estimate $P(S\leq$12,000)$ with a continuity correction and give four decimals. (Answer: approximately 0.2023.)
A hypothetical frequency model has $E[N]=50$ and $\operatorname{Var}(N)=60$. Independent claim payments have mean $3,000 and variance $4$ million dollars squared. Estimate $P(S>$200,000)$ with a continuous normal approximation and give four decimals. (Answer: approximately 0.0330.)

What You'll Learn