Conditioning Aggregate Loss On Claim Count

Why Condition on the Claim Count?

An aggregate loss is a sum of claim payments, but both the number of terms in the sum and the size of each term can vary. Fixing the claim count isolates the severity part of the model: with the count held at $n$, the remaining randomness comes from the claim amounts. This separation is useful when a problem gives an observed claim count or asks what aggregate losses look like for a specified number of claims.

Let $N$ be the claim count, let $X_1,X_2,\ldots$ denote claim severities, and define the aggregate loss by $S_N=\sum_{i=1}^{N}X_i$. When $N=n$, this sum has exactly $n$ terms. The familiar conditional formulas require a model assumption: conditional on the count, the claim severities are independent and identically distributed, with the same distribution as $X$.

Definition: Under the independent, identically distributed severity assumption, write $\mu=E[X]$ and $\sigma^2=\operatorname{Var}(X)$. For a fixed count $n$, the conditional aggregate loss is $S_N\mid N=n=\sum_{i=1}^{n}X_i$. Its conditional mean and variance are $E[S_N\mid N=n]=n\mu$ and $\operatorname{Var}(S_N\mid N=n)=n\sigma^2$. For $n=0$, the sum is empty and is defined to be zero, so both moments are zero.

The count distribution—whether binomial, Poisson, negative binomial, or another model—determines how likely different values of $N$ are. But once the count is fixed at $n$, those frequency probabilities do not appear in these two formulas. The conditional moments depend on $n$ and on the severity moments. This does not mean frequency is irrelevant to the overall aggregate model; it means that conditioning has temporarily removed count variation from the calculation.

Why the Conditional Formulas Hold

For a fixed value $N=n$, aggregate loss is the sum of $n$ severities. Linearity of expectation gives its mean without requiring independence. Independence is needed for the usual variance formula: it makes every covariance between distinct severities zero.

E[S_N\mid N=n] =E\left[\sum_{i=1}^{n}X_i\mid N=n\right] =\sum_{i=1}^{n}E[X_i\mid N=n] =n\mu

\begin{aligned} \operatorname{Var}(S_N\mid N=n) &=\sum_{i=1}^{n}\operatorname{Var}(X_i\mid N=n) +2\sum_{1\le i<j\le n}\operatorname{Cov}(X_i,X_j\mid N=n)\\ &=n\sigma^2. \end{aligned}

The last equality uses both parts of the assumption: each severity has conditional variance $\sigma^2$, and distinct severities have conditional covariance zero. If the claim amounts have a shared risk factor, their covariance may not be zero. In that case, simply multiplying one severity variance by $n$ can give the wrong aggregate variance.

Important condition: The standard conditional formulas apply when the severities retain the specified mean and variance given $N=n$, and are conditionally independent. If severity depends on the count, use the conditional severity moments instead. If severities are dependent, include their conditional covariances in the variance.

Worked Conditional Calculations

Worked Example: Four Claims Under a Poisson Frequency Model

A hypothetical portfolio has $N\sim\operatorname{Poisson}(3)$. Claim severities are independent of $N$ and independent of one another. Each severity is $\$80$ with probability $0.25$ and $\$200$ with probability $0.75$. Given that four claims occurred, calculate the conditional mean and variance of aggregate loss.

Step 1: Find the severity mean.

E[X]=0.25(80)+0.75(200)=20+150=\$170.

Step 2: Find the severity variance. First calculate the second moment, then subtract the squared mean:

E[X^2]=0.25(80^2)+0.75(200^2) =0.25(6{,}400)+0.75(40{,}000) =1{,}600+30{,}000=31{,}600.

\operatorname{Var}(X)=31{,}600-170^2 =31{,}600-28{,}900=\$2{,}700^2.

Step 3: Apply the fixed-count formulas. Here $n=4$, so the conditional mean is $4(170)=\$680$, and the conditional variance is $4(2{,}700)=\$10{,}800$ squared.

The Poisson mean of 3 describes the frequency model before conditioning. It does not replace the given count of four in the conditional calculation. As a check, the conditional mean is four times the average severity, and the conditional variance is four times the variance of one severity.

Method and trap avoided: Calculate severity moments first, then multiply by the observed count. Using $3$, the Poisson mean, in place of the conditioned count $4$ would answer a different question.

Worked Example: A Binomial Count and a Three-Point Severity Model

A hypothetical equipment plan has $N\sim\operatorname{Binomial}(5,0.4)$. Conditional on a claim occurring, its severity is $\$100$, $\$300$, or $\$500$, with probabilities $0.2$, $0.5$, and $0.3$, respectively. Severity is independent across claims and of $N$. Given $N=2$, find the aggregate-loss mean and variance.

Step 1: Verify the conditioning event is compatible with the frequency model. The binomial model allows counts from 0 through 5, so $N=2$ is possible. Its probability is

P(N=2)=\binom{5}{2}(0.4)^2(0.6)^3 =10(0.16)(0.216)=0.3456.

Step 2: Calculate the severity mean.

E[X]=0.2(100)+0.5(300)+0.3(500)=20+150+150=\$320.

Step 3: Calculate the severity variance.

E[X^2]=0.2(10{,}000)+0.5(90{,}000)+0.3(250{,}000) =2{,}000+45{,}000+75{,}000=122{,}000,

\operatorname{Var}(X)=122{,}000-320^2 =122{,}000-102{,}400=19{,}600.

Step 4: Condition on two claims. The aggregate mean is $2(320)=\$640$, and its variance is $2(19{,}600)=\$39{,}200$ squared.

The value $0.3456$ confirms that the given count is possible, but it is not a multiplier in either conditional moment. The conditional question fixes the count at two.

Method and trap avoided: Separate the frequency calculation from the severity calculation. Do not multiply the conditional mean or variance by $P(N=2)$; that probability is relevant to unconditional event probabilities, not to moments conditional on $N=2$.

When Claim Severities Are Dependent

Independence is not a cosmetic assumption. If all claims in a period are affected by the same uncertain condition, claim severities can move together. For $n$ exchangeable severities with common conditional variance $\sigma^2$ and common pairwise conditional covariance $c$, the variance expansion is

\operatorname{Var}(S_N\mid N=n)=n\sigma^2+n(n-1)c.

There are $n(n-1)/2$ distinct pairs in the variance expansion, and each covariance appears twice. Thus the total covariance contribution is $2[n(n-1)/2]c=n(n-1)c$. When $c=0$, this reduces to $n\sigma^2$; when $c$ is positive, the independent-severity formula understates the variance.

Worked Example: A Shared Risk Factor Across Two Claims

A hypothetical two-claim event has claim payments $X_i=Z+U_i$ for $i=1,2$. The common component $Z$ is $\$0$ or $\$200$, each with probability $0.5$. The individual components $U_1,U_2$ are independent, each $\$0$ or $\$400$ with equal probability. All three variables are mutually independent, and $N=2$ is fixed. Find the mean and variance of aggregate loss.

Step 1: Find the component means and variances.

E[Z]=0.5(0)+0.5(200)=\$100, \qquad \operatorname{Var}(Z)=0.5(0.5)(200-0)^2=\$10{,}000^2.

E[U_i]=0.5(0)+0.5(400)=\$200, \qquad \operatorname{Var}(U_i)=0.5(0.5)(400-0)^2=\$40{,}000^2.

Step 2: Calculate the aggregate mean. Since $S_2=X_1+X_2=2Z+U_1+U_2$,

E[S_2]=2E[Z]+E[U_1]+E[U_2] =2(100)+200+200=\$600.

Step 3: Calculate the aggregate variance. The three components $Z,U_1,U_2$ are mutually independent, so their variances add, with the coefficient of $Z$ squared:

\begin{aligned} \operatorname{Var}(S_2) &=4\operatorname{Var}(Z)+\operatorname{Var}(U_1)+\operatorname{Var}(U_2)\\ &=4(10{,}000)+40{,}000+40{,}000\\ &=\$120{,}000^2. \end{aligned}

Check using claim-level moments. Each claim has mean $100+200=\$300$ and variance $10{,}000+40{,}000=50{,}000$ squared. The two claims share $Z$, so their covariance is $\operatorname{Var}(Z)=10{,}000$ squared. Therefore the aggregate variance is $2(50{,}000)+2(10{,}000)=\$120{,}000$ squared, matching the component calculation.

Method and trap avoided: Keep the common component visible, or include the covariance between claim payments. Treating the two claims as independent would give $2(50{,}000)=\$100{,}000$ squared, which is too small because it omits the shared risk factor.

Common Exam Traps and Practical Interpretation

Using the expected count after conditioning. If the question states $N=n$, use $n$ in the conditional formulas. Do not substitute $E[N]$.
Confusing second moment with variance. Compute $\operatorname{Var}(X)=E[X^2]-(E[X])^2$ before multiplying by the count. Multiplying $E[X^2]$ by $n$ does not give the aggregate variance.
Assuming that a fixed count guarantees independent claims. The count is fixed, but a common environment, policy feature, or other shared risk factor can still create dependence between payments.
Forgetting the units. A mean is measured in dollars; a variance is measured in squared dollars. Multiplication by $n$ changes neither type of unit.
Applying the standard formulas to changing severity distributions. If the conditional severity mean or variance depends on the count, use the moments conditional on $N=n$, rather than unconditional severity moments.

Conditioning is valuable because it clarifies which source of uncertainty is being measured. At a fixed count, severity variation remains while frequency variation is held fixed. The result is a clean calculation, provided the severity assumptions are checked before using it. This distinction is aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus.

Key takeaway: For $n$ conditionally independent, identically distributed claim severities, aggregate loss has conditional mean $nE[X]$ and conditional variance $n\operatorname{Var}(X)$. If claims share a risk factor, add the covariance contributions instead of assuming them away.

Check Your Understanding

For each problem, condition on the stated claim count before applying aggregate-loss formulas. Variances are in squared dollars.

A hypothetical portfolio has $N\sim\operatorname{Poisson}(4)$. Each independent severity is $\$100$ with probability $0.5$, $\$200$ with probability $0.25$, and $\$400$ with probability $0.25$. Given $N=2$, find the conditional mean and variance. (Answer: mean $\$400$; variance $\$30{,}000$ squared.)
A hypothetical warranty has $N\sim\operatorname{Binomial}(6,0.3)$. Each independent severity is $\$100$ with probability $0.2$ and $\$600$ with probability $0.8$. Given $N=1$, find $P(N=1)$, then give the conditional aggregate mean and variance. (Answer: $P(N=1)=0.3025$; mean $\$500$; variance $\$40{,}000$ squared.)
Two independent severities have the following distributions: $X_1$ is $\$0$ or $\$200$ with equal probability, and $X_2$ is $\$100$ or $\$500$ with equal probability. If the count is fixed at two, find the aggregate mean and variance. (Answer: mean $\$400$; variance $\$50{,}000$ squared.)
For two claims, let $X_i=Z+U_i$, where $Z$ is $\$0$ or $\$100$ with equal probability and each independent $U_i$ is $\$0$ or $\$200$ with equal probability. All components are mutually independent. Find the aggregate mean and variance, and identify the variance that would result from incorrectly treating the claim payments as independent. (Answer: mean $\$300$; correct variance $\$30{,}000$ squared; incorrect independent-severity variance $\$25{,}000$ squared.)
Suppose a fixed count of three has conditionally exchangeable severities with conditional variance $\$8{,}000$ squared per claim and pairwise conditional covariance $\$1{,}000$ squared. Find the aggregate variance. (Answer: $\$30{,}000$ squared.)

What You'll Learn