Independence And Identical Severity Assumptions

Why the Independence Assumptions Matter

In a compound aggregate model, total loss is a random sum: the number of claims is random, and each claim contributes a random severity. These are two separate sources of randomness. A model must specify both how the severities relate to one another and how the claim count relates to them. Treating either relationship incorrectly can produce the wrong distribution for aggregate loss, even when the count model and the individual severity distribution appear reasonable.

Let $N$ be the number of claims in a fixed period, and let $X_j$ be the severity of claim $j$, for $j=1,2,\ldots$. Define aggregate loss by

S_N=\sum_{j=1}^{N}X_j,

with $S_N=0$ when $N=0$. The standard compound model makes two distinct assumptions: the severities $X_1,X_2,\ldots$ are independent and identically distributed, and the count $N$ is independent of the severity sequence. Neither assumption means that severities are zero, nor that the number of claims is fixed.

Definition: The severities are independent and identically distributed (iid) if every $X_j$ has the same probability distribution and any finite collection of different severities is mutually independent. The count $N$ is independent of the severity sequence if knowing $N$ provides no information about any finite collection of severities, and knowing those severities provides no information about $N$.

Two Different Kinds of Independence

It is useful to state the two assumptions separately. Independence among severities concerns relationships such as whether a large first claim makes a large second claim more likely. Independence between $N$ and the severity sequence concerns whether the claim count contains information about the severities. A model can satisfy one assumption and fail the other.

Assumption	What it says	What it does not say
Severities are iid	Each claim has the same severity distribution, and different claim severities are mutually independent.	It does not make $N$ independent of the severities.
$N$ is independent of the severity sequence	The count gives no information about the severities, and vice versa.	It does not make the severities independent of one another.

“Iid” does not mean that every claim has the same realized dollar amount. It means the claims are generated by the same probability law, independently. Conversely, saying that claim severities have the same distribution does not establish independence: two claims can have identical marginal distributions but move together because of a shared cause.

The Conditional Aggregate Law

Suppose the severity sequence is iid with distribution $F_X$, and $N$ is independent of the entire sequence. If the count is fixed at $n$, the aggregate consists of $n$ independent draws from $F_X$. Thus, for $n\ge 1$, the conditional distribution of $S_N$ given $N=n$ is the distribution of a sum of $n$ iid severities. In distribution notation, it is the $n$-fold convolution of $F_X$. When $n=0$, the conditional aggregate is zero with probability one.

S_N\mid(N=n)\ \overset{d}{=}\ X_1+\cdots+X_n, \qquad F_{S_N\mid N=n}=F_X^{*n}\quad(n\ge 1).

Here, $F_X^{*n}$ denotes the distribution obtained by adding $n$ independent variables with distribution $F_X$; it does not mean that the severities are multiplied. For a discrete severity distribution, the conditional aggregate probabilities are found by adding the possible severity values in every way that produces the specified total. For continuous severities, the corresponding distribution is obtained by convolution of the severity distributions.

The role of count-severity independence can be seen in a direct factorization. For events $A_1,\ldots,A_n$ describing the first $n$ severities, independence gives

P(X_1\in A_1,\ldots,X_n\in A_n\mid N=n) = \frac{P(N=n)P(X_1\in A_1,\ldots,X_n\in A_n)}{P(N=n)} = \prod_{j=1}^{n}P(X_j\in A_j),

provided $P(N=n)>0$. The final equality uses mutual independence of the severities. Therefore, after conditioning on $N=n$, the severities still have their original joint iid distribution. If $N$ is associated with severity, that cancellation does not hold: conditioning on the count can change which severities are likely.

Worked Example: The Conditional Aggregate Under the Standard Assumptions

A hypothetical portfolio has a claim count $N$ that is independent of an iid severity sequence. Each claim is either a loss of $1,000 or $3,000, each with probability $0.5$. Given that $N=2$, find the distribution of aggregate loss and its conditional expected value.

Method. Apply the iid severity assumption to the two claims. The count-severity independence assumption ensures that conditioning on $N=2$ does not change the severity probabilities.

There are four equally likely ordered pairs of severities. Each pair has probability $0.5\times0.5=0.25$:

$($1,000,$1,000)$ gives a total of $2,000, with probability $0.25$.
$($1,000,$3,000)$ gives a total of $4,000, with probability $0.25$.
$($3,000,$1,000)$ gives a total of $4,000, with probability $0.25$.
$($3,000,$3,000)$ gives a total of $6,000, with probability $0.25$.

Combining the two different orders that produce $4,000 gives $P(S_N=$4,000\mid N=2)=0.25+0.25=0.50$. The full conditional distribution is therefore: $2,000 with probability $0.25$, $4,000 with probability $0.50$, and $6,000 with probability $0.25$.

The mean severity is $0.5($1,000)+0.5($3,000)=$2,000$. Therefore the conditional expected aggregate is $2($2,000)=$4,000$. Checking the distribution directly gives $0.25($2,000)+0.50($4,000)+0.25($6,000)=$4,000$, as required.

Common mistake avoided: The two mixed orders are different outcomes, but both contribute to the probability of a $4,000 total. Counting only one mixed order would incorrectly assign that total probability $0.25$, rather than $0.50$.

Why Identical Marginals Are Not Enough

Suppose every claim severity has the same marginal distribution, but two or more severities are dependent. Knowing one severity may then change the probability distribution of another. The conditional aggregate distribution need not be the convolution of the marginal severity distribution with itself, because convolution in this setting relies on independent summands.

For example, a shared event could affect every claim in a period: on a quiet day, all claims might be small, while on a severe-weather day, all might be large. Each claim could have the same marginal distribution across days, yet severities within a day would be positively dependent. A model that uses only the individual severity distribution would miss that clustering.

Worked Example: Same Marginal Severity, Dependent Claims

A hypothetical portfolio has a claim count $N$ independent of the severity pair $(X_1,X_2)$, with $P(N=2)>0$. Each severity equals $1,000 with probability $0.5$ and $3,000 with probability $0.5$, but the claims are perfectly dependent: $X_1=X_2$ with probability one. Find the conditional aggregate distribution given $N=2$, and compare it with the iid result.

Method. Use the stated joint dependence rather than assuming that equal marginal distributions imply independence. Since $N$ is independent of the severity pair, conditioning on $N=2$ leaves the pair’s joint distribution unchanged.

With probability $0.5$, both severities are $1,000, so the total is $2($1,000)=$2,000$. With probability $0.5$, both are $3,000, so the total is $2($3,000)=$6,000$. Thus $P(S_N=$2,000\mid N=2)=0.5$, $P(S_N=$6,000\mid N=2)=0.5$, and the conditional probability of $4,000 is zero.

The marginal mean severity is still $0.5($1,000)+0.5($3,000)=$2,000$, so the conditional mean aggregate is $4,000. The result can be checked from the two aggregate outcomes: $0.5($2,000)+0.5($6,000)=$4,000$. The mean matches the iid example, but the distributions differ: under iid severities, the aggregate had probability $0.50$ at $4,000; here it has probability zero there.

Common mistake avoided: Using the iid convolution because both claims have the same marginal distribution. Identical distributions alone do not justify treating claims as independent.

Why Count-Severity Independence Is Also Separate

Even if the severity sequence is iid, the count can be associated with it. In that case, knowing $N=n$ may change the distribution of the severities included in the aggregate. The conditional aggregate law then cannot generally be obtained by taking $n$ independent draws from the unconditional severity distribution.

Worked Example: An Iid Severity Sequence with an Informative Count

A hypothetical model has an iid severity sequence, where each $X_j$ is $1,000 or $3,000 with probability $0.5$ each. Define the count by $N=1$ when $X_1=$1,000$, and $N=0$ when $X_1=$3,000$. Find the aggregate distribution conditional on $N=1$, and explain why the standard conditional law does not apply.

Method. Condition directly on the rule defining $N$. The severity sequence is iid, but the count is not independent of that sequence.

The event $N=1$ occurs exactly when $X_1=$1,000$. Therefore $P(N=1)=0.5$, and, conditional on $N=1$, the aggregate is $S_N=X_1=$1,000$ with probability one. In particular, $P(S_N=$3,000\mid N=1)=0$.

If one incorrectly applied the standard compound-model conditional law, one would assign probability $0.5$ to each of $1,000 and $3,000 when $N=1$. That contradicts the rule defining the count. The conditional mean under the actual model is $1,000, not the unconditional mean severity of $2,000.

Check. The count takes value 1 with probability $0.5$, and on that event aggregate loss is $1,000. It takes value 0 with probability $0.5$, and aggregate loss is zero. These outcomes agree with the construction and demonstrate that count-severity independence fails.

A Practical Assumption Check

Before using a standard compound aggregate model, translate the wording of the problem into two questions. First, are claim severities mutually independent and governed by the same distribution? Second, is the count independent of the severity sequence? A statement that each claim has a common severity distribution answers only part of the first question; it does not establish either mutual independence or count-severity independence.

Identify the count and severity sequence.
Write $N$ for the number of claims and $X_j$ for the severity of claim $j$. This makes clear which random quantities are being compared.

Check the severity assumptions.
Look for support that claim severities have the same distribution and are mutually independent. Equal marginal distributions alone are insufficient.

Check count-severity independence.
Ask whether the number of claims could reveal anything about claim sizes, or whether a shared event affects both count and severity.

Use the conditional sum law only when justified.
Under both standard assumptions, given $N=n$, aggregate loss has the distribution of a sum of $n$ iid severities. If either assumption fails, the conditional distribution must reflect the dependence that was specified.

A common exam trap is to see “the severity distribution is $F_X$” and immediately treat all claim amounts as iid and independent of the count. That statement alone supplies a marginal distribution, not the full dependence structure. Use only the assumptions stated or clearly implied by the model.

Key takeaway: The standard conditional compound law requires both iid severities and independence of the claim count from the severity sequence. Independence among severities and independence between count and severities are distinct assumptions; neither implies the other.

Check Your Understanding

For each problem, identify the relevant independence assumption before finding the requested conditional aggregate result.

A hypothetical portfolio has $N$ independent of an iid severity sequence. Each severity is $2,000 or $5,000 with probability $0.5$ each. Given $N=2$, find the probability that aggregate loss is $7,000. (Answer: $0.5000$)
Two hypothetical claim severities always equal one another. Each is $4,000 or $8,000 with probability $0.5$ each, and $N$ is independent of the pair. Given $N=2$, give the aggregate support and its probabilities. (Answer: $8,000 with probability $0.5000$; $16,000 with probability $0.5000$)
Claim severities are iid, each $1,000 or $3,000 with equal probability. A count rule sets $N=1$ exactly when $X_1=$3,000$, and otherwise sets $N=0$. Find $P(S_N=$3,000\mid N=1)$. (Answer: $1.0000$)
Suppose severities are independent but alternate distributions by claim order: $X_1$ is always $1,000 and $X_2$ is always $4,000. Assume $N=2$ with certainty. Is the iid severity assumption satisfied, and what is aggregate loss? (Answer: No; aggregate loss is $5,000 with probability $1.0000$)
A model states that $N$ is independent of the severity sequence and that all severities have the same marginal distribution, but it does not state whether different severities are independent. Is the standard $n$-fold convolution conditional law justified from these facts alone? (Answer: No; mutual independence among severities is also required.)

Assumption	What it says	What it does not say
Severities are iid	Each claim has the same severity distribution, and different claim severities are mutually independent.	It does not make \(N\) independent of the severities.
\(N\) is independent of the severity sequence	The count gives no information about the severities, and vice versa.	It does not make the severities independent of one another.

What You'll Learn