Expected Aggregate Loss By Conditioning

Why Conditioning Simplifies Expected Aggregate Loss

An aggregate loss combines two sources of uncertainty: how many claims occur and how large each claim is. Its full probability distribution can be complicated, but its expected value has a particularly direct form. Conditioning on the claim count separates the frequency part from the severity part and leads to a useful formula for a wide range of models.

This result applies whether the claim count is binomial, Poisson, negative binomial, or another count model. The frequency distribution matters through its mean; the severity distribution matters through its mean. The key modeling condition is that the individual claim severities have the same distribution and are independent of the claim count.

Definition: Let $N$ be a nonnegative integer-valued claim count. Let $X_1,X_2,\ldots$ be the individual claim severities, and define the aggregate loss by $S=\sum_{j=1}^{N}X_j$. When $N=0$, the sum is empty and $S=0$. Suppose the $X_j$ are identically distributed with finite mean $E[X]$, are independent of $N$, and $E[N]$ is finite. Then $E[S]=E[N]E[X]$.

Here $E[N]$ is the expected number of claims over the period being modeled, and $E[X]$ is the expected severity per claim. Their product has units of currency: claims multiplied by dollars per claim gives dollars. The result concerns the mean aggregate loss; it does not, by itself, describe the probability distribution or the variability of that loss.

Deriving the Formula by Conditioning

If the claim count is known to be $n$, the aggregate is the sum of exactly $n$ severities. By linearity of expectation, the expected value of that sum is $nE[X]$. The independence condition ensures that conditioning on the value of $N$ does not change the distribution or mean of an individual severity.

E[S\mid N=n] =E\left[\sum_{j=1}^{n}X_j\middle|N=n\right] =\sum_{j=1}^{n}E[X_j\mid N=n] =nE[X].

The law of total expectation now averages these conditional means across the possible values of the claim count:

\begin{aligned} E[S] &=E\bigl[E[S\mid N]\bigr]\\ &=E[NE[X]]\\ &=E[N]E[X]. \end{aligned}

In the second line, $E[X]$ is a constant, not a random quantity, because the severity distribution is the same for each claim and independent of $N$. This derivation explains why the frequency distribution’s detailed shape is not needed to calculate the mean once $E[N]$ is known. It also shows where the assumptions enter: if the expected severity depends on the claim count, then $E[S\mid N=n]$ need not equal $nE[X]$, and the product formula may not apply.

Theorem: Under the assumptions in the definition, the expected aggregate loss is the expected claim count multiplied by the expected severity per claim. The claim count may have any distribution on the nonnegative integers for which $E[N]$ is finite; the formula does not require a Poisson model.

Applying the Result to Frequency Models

For common frequency models, calculate $E[N]$ using the parameters and conventions of that model, then multiply by the mean severity. For example, if $N\sim\operatorname{Binomial}(m,q)$, then $E[N]=mq$. If $N\sim\operatorname{Poisson}(\lambda)$, then $E[N]=\lambda$. For a negative binomial model, state the parameterization before using its mean, because multiple conventions are in use.

For the negative binomial convention used below, $N$ is the number of failures before the $r$th success in independent trials, each with success probability $p$. Under that convention, $E[N]=r(1-p)/p$. A model that instead counts total trials or uses a different success-probability convention requires its corresponding mean.

Worked Example: Binomial Frequency and Discrete Severity

A hypothetical portfolio covers a one-year period. The number of claims is $N\sim\operatorname{Binomial}(12,0.10)$. Claim severities are independent of $N$ and of one another. Each claim is \$1,000 with probability 0.50, \$3,000 with probability 0.30, or \$7,000 with probability 0.20. Find the expected aggregate loss.

Method and assumptions. Use conditioning on the claim count. The binomial count has $m=12$ trials and claim probability $q=0.10$. The stated severity probabilities sum to $0.50+0.30+0.20=1.00$, and the independence assumptions required for the product formula are given.

First calculate the expected claim count:

E[N]=mq=12(0.10)=1.2\text{ claims}.

Next calculate the expected severity, weighting each possible amount by its probability:

\begin{aligned} E[X] &=1000(0.50)+3000(0.30)+7000(0.20)\\ &=500+900+1400\\ &=\$2{,}800. \end{aligned}

Therefore,

E[S]=E[N]E[X]=1.2(\$2{,}800)=\$3{,}360.

Check. The result is in dollars, as expected: 1.2 expected claims multiplied by \$2,800 per claim gives \$3,360. A common mistake is to use the mean severity alone as the aggregate mean, implicitly treating the portfolio as if it had exactly one claim. The count is random and its expected value is 1.2, so the correct aggregate mean is \$3,360, not \$2,800.

Worked Example: Poisson Frequency Over a Multi-Year Period

A hypothetical insurer models claim arrivals at a constant rate of 2.4 claims per year, with independent arrivals across nonoverlapping periods. Each claim severity has an exponential distribution with mean \$1,800, independent of the count. Find the expected aggregate loss over three years.

Method and assumptions. A Poisson count with rate 2.4 per year over three years has mean $2.4(3)$, using the stated constant-rate model. The exponential severity mean is given directly, so no distributional calculation beyond identifying that mean is needed.

The expected number of claims over the full three-year exposure is

E[N]=2.4\text{ claims per year}\times3\text{ years}=7.2\text{ claims}.

The expected severity is

E[X]=\$1{,}800\text{ per claim}.

Thus,

E[S]=7.2(\$1{,}800)=\$12{,}960.

Check. The calculation uses the three-year expected count, not the one-year rate, and the units reduce to dollars. A common time-unit error is to multiply the annual frequency by the severity but forget the three-year exposure; that would give \$4,320, the expected loss for one year rather than three.

Worked Example: Negative Binomial Frequency and Gamma Severity

For a hypothetical coverage period, let $N$ be the number of failures before the fifth success in independent trials, with success probability $p=0.75$. Each claim severity has a gamma distribution with shape parameter 3 and scale parameter \$600. Assume the severities are independent of the count and of one another. Calculate the expected aggregate loss.

Method and parameterization. This negative binomial variable counts failures before the $r$th success, so its mean is $r(1-p)/p$. The gamma distribution is specified using shape and scale; for shape $\alpha$ and scale $\theta$, its mean is $\alpha\theta$.

For the count, $r=5$ and $1-p=0.25$, giving

E[N]=\frac{r(1-p)}{p} =\frac{5(0.25)}{0.75} =\frac{1.25}{0.75} =\frac{5}{3}\text{ claims}.

For severity, $\alpha=3$ and $\theta=\$600$, so

E[X]=\alpha\theta=3(\$600)=\$1{,}800.

The expected aggregate loss is therefore

E[S]=\frac{5}{3}(\$1{,}800)=\$3{,}000.

Check. Multiplying \$1,800 by $5/3$ gives \$3,000 exactly. A common mistake is to use the wrong negative binomial convention, for example counting the successful trials as part of $N$. Here $N$ counts failures only, so the mean is $5(0.25)/0.75=5/3$, not a mean based on total trials.

Assumptions, Scope, and Exam Traps

The formula is useful because it avoids constructing the aggregate distribution when only its mean is requested. For the expected value, you do not need to calculate the probability of each possible aggregate amount or enumerate every possible claim count. You do need to identify the modeled period, determine the expected number of claims for that period, and calculate the expected severity per claim under the specified severity model.

The assumptions matter. If severities are independent of the count and have a common finite mean, then each conditional severity mean remains $E[X]$, giving the conditional sum $nE[X]$. If claim severities vary systematically with $N$, or if only a count-dependent severity mean is supplied, that reasoning cannot be used without modification. Similarly, $E[N]$ and $E[X]$ must be finite for the stated finite-mean calculation.

The aggregate expectation does not require the individual severities to be independent of one another if, conditional on the count, each has the same expected value $E[X]$. Independence among severities is commonly assumed in compound models and is stated in the examples, but the expectation calculation itself uses linearity of expectation. Independence of severity from the count is what ensures the conditional mean per claim remains unchanged.

Match the exposure period. A rate per month must be multiplied by the number of months in the modeled period before it is combined with severity.
Use a probability-weighted severity mean. For discrete outcomes, multiply each amount by its probability and add the products.
State the count convention. In particular, verify what the negative binomial variable counts before using its mean.
Keep the target clear. $E[S]$ is the expected total loss, not the expected loss per claim and not a percentile or tail measure.

These distinctions are useful in compound aggregate calculations and in integrated problems that combine frequency and severity. Recognizing the conditional-expectation structure is aligned with the SOA Exam FAM (Fundamentals of Actuarial Mathematics) syllabus.

Key takeaway: When the expected severity per claim does not change with the claim count, condition on $N$: the conditional expected aggregate is $NE[X]$, and averaging over $N$ gives $E[S]=E[N]E[X]$. Calculate the count mean for the correct exposure period and the severity mean under the stated distribution.

Check Your Understanding

For each hypothetical model, identify the frequency mean and severity mean before finding the expected aggregate loss. Assume the frequency and severities satisfy the independence conditions stated in the definition.

A one-year count is $N\sim\operatorname{Binomial}(10,0.20)$. Each claim is exactly \$750. Find $E[S]$. (Answer: \$1,500)
Claims arrive according to a Poisson model at 1.5 claims per month for an eight-month period. Each severity is \$400 with probability 0.25, \$1,000 with probability 0.50, and \$2,600 with probability 0.25. Find $E[S]$. (Answer: \$15,000)
$N$ counts failures before the third success, with success probability 0.60. Severities have a gamma distribution with shape 2 and scale \$900. Find $E[S]$. (Answer: \$3,600)
A count is binomial with 20 trials and claim probability 0.05. Each severity is \$1,000 with probability 0.70 or \$5,000 with probability 0.30. Find $E[S]$. (Answer: \$2,200)
A hypothetical portfolio has a Poisson claim rate of 0.30 claims per week for 10 weeks. Severities are exponential with mean \$4,000 per claim. Find the expected aggregate loss over the full period. (Answer: \$12,000)

What You'll Learn