The Random-Sum Model for Aggregate Losses

Why Model Aggregate Loss as a Random Sum?

An insurer’s total loss over a period varies for two distinct reasons: the number of claims can vary, and the amount of each claim can vary. A model that represents both sources of variation begins by treating aggregate loss as a sum with a random number of terms. This representation is useful for describing a portfolio’s losses and is a foundation for later credibility calculations.

The two parts of the model should not be confused. The claim count tells us how many losses occur during the period. The claim severities give the amounts of those losses. For example, knowing that a period has three claims does not determine its aggregate loss unless the three claim amounts are also specified.

Definition: Let $N$ be the number of claims in a specified exposure period, and let $X_1,X_2,\ldots$ be the claim amounts. The aggregate loss is $$ S=\sum_{j=1}^{N}X_j, $$ with the convention $S=0$ when $N=0$. In the standard random-sum model, $N$ is independent of the claim amounts, and the $X_j$ are independent and identically distributed with a common severity distribution. Write $E[X_j]=\mu_X$, when the mean exists.

The period and the meaning of “claim amount” must be specified consistently. For example, the count might refer to claims incurred during one policy year, while each $X_j$ might be the corresponding ultimate amount for one such claim. If the amount is instead a payment made during that year, the severity definition has changed. The model does not itself decide whether amounts are gross or net of a deductible, limit, or other adjustment; that choice must be made when defining $X_j$.

Model Ingredients and Assumptions

A random-sum model is a representation, not a claim that every portfolio has a particular count distribution or severity distribution. The analyst must select distributions that suit the risk and the exposure period. The following assumptions describe the usual basic model:

A well-defined count: $N$ is a nonnegative integer-valued random variable for the specified portfolio and period.
Consistent severity units: each $X_j$ measures the same type of loss on the same basis, such as dollars per claim.
Identically distributed severities: each claim amount has the same distribution in the modeled population.
Independent severities: knowing one claim amount does not change the distribution of another claim amount.
Count-severity independence: the number of claims does not change the distribution of the individual claim amounts.

These assumptions can be inappropriate. For example, if a storm produces many claims and also tends to produce unusually large claims, the count and severity may be dependent. If the portfolio combines distinct risk groups with different severity distributions, the severities may not be identically distributed. A more detailed model may then be needed; one should not apply the standard random-sum assumptions automatically.

A Poisson model is one possible choice for $N$, not part of the definition of a random sum. If $N\sim\operatorname{Poisson}(\lambda)$, then $\lambda$ is the expected claim count over the stated period. Other count distributions can also be used. The severity distribution is selected separately.

Worked Example: A Fixed Amount per Claim

A hypothetical equipment policy portfolio has claim count $N\sim\operatorname{Poisson}(2)$ during one year. Assume each claim has a fixed amount of $\$750$, and the claim amount is independent of the count. Find the aggregate-loss model, the probability that aggregate loss is $\$1{,}500$, and the expected aggregate loss. There is no interest or mortality assumption in this one-year loss model.

Method and model. Let $X_j=\$750$ for every claim. The model is $$ S=\sum_{j=1}^{N}750=750N. $$ The count is Poisson with parameter $\lambda=2$; severity is deterministic at $\$750$.

Aggregate loss equals $\$1{,}500$ exactly when $750N=1{,}500$, or $N=2$. Therefore, $$ P(S=\$1{,}500)=P(N=2) =\frac{2^2e^{-2}}{2!} =2e^{-2} \approx 0.2707. $$ The count probabilities sum to one over all possible nonnegative counts, and the event $S=\$1{,}500$ corresponds to precisely one count value, $N=2$.

For the mean, $E[N]=2$, so the expected aggregate amount is $$ E[S]=750E[N]=750(2)=\$1{,}500. $$ This is an expected value, not a statement that the aggregate loss will equal $\$1{,}500$ in every year. The probability calculation confirms that this exact amount occurs with probability about $0.2707$, not certainty. The method avoids treating the mean aggregate as a guaranteed outcome.

Calculating Aggregate-Loss Probabilities

When severities are variable, aggregate-loss probabilities are found by considering which claim counts and severity combinations can produce the target amount. For a discrete severity distribution, each possible combination contributes its probability, conditional on the count. Count-severity independence then allows the probability of a count and a severity combination to be multiplied.

The number of combinations can grow quickly. A useful first step is to identify feasible counts. If every claim is at least $a>0$, then $N$ claims produce aggregate loss of at least $aN$. This can rule out counts that are too large. Similarly, if a count is fixed at $n$, the aggregate is the sum of exactly $n$ severities.

Worked Example: A Discrete Severity Distribution

A hypothetical small-business portfolio has $N\sim\operatorname{Poisson}(0.6)$ claims during a specified period. Independently, each claim amount is $\$500$ with probability $0.5$, or $\$1{,}500$ with probability $0.5$. Find $P(S=\$1{,}500)$. Assume the stated amounts are the claim amounts used in the aggregate, with no interest or mortality adjustment.

Method and model. We enumerate the claim counts and severity combinations that can sum to $\$1{,}500$. With one claim, the amount must be $\$1{,}500$. With two claims, the smallest possible total is $\$1{,}000$, and the possible totals are $\$1{,}000,\$2{,}000,\$3{,}000$, so two claims cannot produce $\$1{,}500$. With three claims, the total is $\$1{,}500$ only if all three claims are $\$500$. Four or more claims have minimum aggregate loss $\$2{,}000$, so they cannot contribute.

For one claim, the Poisson probability is $$ P(N=1)=\frac{0.6^1e^{-0.6}}{1!}=0.6e^{-0.6}. $$ The conditional probability that its amount is $\$1{,}500$ is $0.5$. Thus this case contributes $$ P(N=1)P(X_1=\$1{,}500) =(0.6e^{-0.6})(0.5) =0.3e^{-0.6}. $$

For three claims, $$ P(N=3)=\frac{0.6^3e^{-0.6}}{3!} =\frac{0.216}{6}e^{-0.6} =0.036e^{-0.6}. $$ Given three claims, the probability all three are $\$500$ is $$ (0.5)^3=0.125. $$ So this case contributes $$ P(N=3)P(X_1=X_2=X_3=\$500) =(0.036e^{-0.6})(0.125) =0.0045e^{-0.6}. $$

Adding the disjoint cases gives $$ P(S=\$1{,}500)=(0.3+0.0045)e^{-0.6} =0.3045e^{-0.6} \approx 0.1671. $$ The answer is a probability, between zero and one. The method avoids omitting the three-claim case or counting impossible two-claim and four-or-more-claim cases.

A Short Derivation for the Expected Aggregate

The expected aggregate loss follows directly from conditioning on the claim count. This result requires a finite expected count and a finite severity mean, as well as the standard independent, identically distributed severity model.

Condition on the claim count.
If $N=n$, aggregate loss is the sum of $n$ severities. By linearity of expectation, $E[S\mid N=n]=n\mu_X$, including $n=0$, for which the sum is zero.

Average over possible counts.
The law of total expectation gives $E[S]=E[E[S\mid N]]=E[N\mu_X]$.

Use the constant severity mean.
Since $\mu_X$ does not depend on $N$ under count-severity independence, $E[S]=\mu_XE[N]$.

$$ E[S]=E[N]E[X_1]. $$

This identity gives the expected aggregate amount; it does not describe the full distribution of $S$. In particular, the expected aggregate alone does not determine the probability of a particular total or the amount of variation around the mean. Those questions require further analysis.

Worked Example: A Binomial Claim Count and Variable Severity

A hypothetical inspection program covers four independent units during one period. Each unit has probability $0.10$ of producing one claim, so $N\sim\operatorname{Binomial}(4,0.10)$. Conditional on a claim, its amount is $\$1{,}000$ with probability $0.60$ and $\$2{,}000$ with probability $0.40$. Assume severity outcomes are independent across claims and independent of the count. Find $P(S=\$2{,}000)$. No interest or mortality assumption is involved.

Method and model. For one claim, the total is $\$2{,}000$ only if the claim amount is $\$2{,}000$. For two claims, the total is $\$2{,}000$ only if both amounts are $\$1{,}000$. Three or four claims have minimum totals $\$3{,}000$ and $\$4{,}000$, respectively, so they do not contribute. Zero claims also cannot produce the target.

The probability of exactly one claim is $$ P(N=1)=\binom{4}{1}(0.10)(0.90)^3 =4(0.10)(0.729) =0.2916. $$ The probability that this one claim is $\$2{,}000$ is $0.40$, so the one-claim contribution is $$ 0.2916(0.40)=0.11664. $$

The probability of exactly two claims is $$ P(N=2)=\binom{4}{2}(0.10)^2(0.90)^2 =6(0.01)(0.81) =0.0486. $$ Given two claims, both must be $\$1{,}000$, with probability $$ (0.60)(0.60)=0.36. $$ The two-claim contribution is therefore $$ 0.0486(0.36)=0.017496. $$

The cases are mutually exclusive, so $$ P(S=\$2{,}000)=0.11664+0.017496=0.134136. $$ The total is between zero and one, and it includes every feasible claim count. The method avoids assuming that an aggregate of $\$2{,}000$ can arise only from one claim.

Common Mistakes and Model Checks

A common arithmetic trap is multiplying the probability of a count by the probability of a particular severity pattern incorrectly. In the discrete-severity example, the three-claim contribution is $$ \left(\frac{0.6^3e^{-0.6}}{3!}\right)(0.5^3) =(0.036e^{-0.6})(0.125) =0.0045e^{-0.6}, $$ not $0.009e^{-0.6}$. Using the incorrect value would give $(0.3+0.009)e^{-0.6}=0.309e^{-0.6}$, rather than the correct $(0.3+0.0045)e^{-0.6}\approx0.1671$. Keep the Poisson probability and the conditional severity-pattern probability separate until both have been calculated.

Confusing claim count with aggregate loss. $N$ is a number of claims; $S$ is a dollar amount. They are not interchangeable.
Using the average severity as each actual claim amount. $E[X_1]$ is a mean. Unless severity is deterministic, replacing every $X_j$ by its mean changes the aggregate distribution.
Forgetting the zero-claim case. The random sum is $S=0$ when $N=0$. This convention is important when calculating the distribution of aggregate loss.
Assuming Poisson counts without justification. The random-sum definition permits other count distributions. Choose the count model based on the modeled risk and period.
Multiplying count and severity probabilities without independence. The multiplication is justified by the standard model’s independence assumption. If larger counts are associated with different severities, that assumption may fail.

Key takeaway: Define the period, the claim count $N$, and the per-claim amounts $X_j$ before calculating aggregate loss. Under the standard model, $S=\sum_{j=1}^{N}X_j$, with $S=0$ when there are no claims; the count and severities are independent, and the severities are independent and identically distributed.

Check Your Understanding

For each problem, identify the relevant random-sum structure before calculating. Amounts are hypothetical, and no interest or mortality adjustment applies unless stated.

A portfolio has $N\sim\operatorname{Poisson}(1.5)$ claims in one period, and every claim is exactly $\$1{,}200$. Write $S$ in terms of $N$, then find $P(S=\$2{,}400)$. (Answer: $S=1{,}200N$; $P(S=\$2{,}400)=1.125e^{-1.5}\approx0.2510$.)
Five independent devices each produce at most one claim, with probability $0.10$ per device. Every claim is $\$500$. Find the probability that aggregate loss is $\$1{,}500$. (Answer: $P(N=3)=\binom{5}{3}(0.10)^3(0.90)^2=0.0081$.)
Let $N\sim\operatorname{Poisson}(0.4)$. Each independent claim is $\$500$ with probability $0.75$ and $\$1{,}500$ with probability $0.25$, independently of $N$. Find $P(S=\$1{,}500)$, showing which counts can contribute. (Answer: $0.1045e^{-0.4}\approx0.0700$; the contributing cases are one $\$1{,}500$ claim and three $\$500$ claims.)
A hypothetical portfolio has $E[N]=2.4$. Claim amounts are $\$800$ with probability $0.70$ and $\$1{,}800$ with probability $0.30$, under the standard independent severity model. Find the expected aggregate loss. (Answer: $E[X_1]=0.70(800)+0.30(1{,}800)=\$1{,}100$, so $E[S]=2.4(1{,}100)=\$2{,}640$.)
Suppose a shared weather event can both increase the number of claims and make each claim more severe. Which standard random-sum assumption may fail, and what should an actuary avoid doing without further justification? (Answer: Count-severity independence may fail; do not multiply count and severity probabilities as if they were independent without justification or a model that accounts for the dependence.)

What You'll Learn