Predicting Aggregate Losses With Frequency and Severity Uncertainty

Why Aggregate Loss Forecasting Requires Both Frequency and Severity

A forecast of total loss depends on two uncertain components: how many claims occur and how much each claim costs. Even if the expected number of claims and the expected claim amount are known, aggregate loss can remain highly variable. In a Bayesian forecast, uncertainty about the underlying frequency and severity parameters adds another source of variation beyond the randomness of future claims.

This tutorial combines posterior predictive frequency and severity distributions in a compound Poisson model. The goal is to calculate the predictive mean and variance of aggregate loss while keeping distinct the variability in future claim counts, future claim amounts, and the unknown parameters governing those outcomes. All insurance scenarios and data below are hypothetical.

Definition: Let $N$ be the number of claims in a future period with exposure $t$, and let $X_1,X_2,\ldots$ be the claim amounts. Aggregate loss is $S=\sum_{j=1}^{N}X_j$, with $S=0$ when $N=0$. Let $\Lambda$ be the claim frequency per unit exposure and let $\Theta$ index the severity distribution. Conditional on $(\Lambda,\Theta)$, assume $N\sim\operatorname{Poisson}(t\Lambda)$, and assume the claim amounts are independent and identically distributed with conditional mean $m(\Theta)=E[X_j\mid\Theta]$ and conditional second moment $q(\Theta)=E[X_j^2\mid\Theta]$. The claim amounts are conditionally independent of $N$. Expectations and variances without conditioning are taken over the posterior distribution of the parameters.

The posterior distribution of $(\Lambda,\Theta)$ may be dependent. Conditional independence of the future count and claim amounts given the parameters does not require the posterior parameters themselves to be independent. The moment $q(\Theta)$ is the severity second moment, not its variance: $q(\Theta)=\operatorname{Var}(X_j\mid\Theta)+m(\Theta)^2$.

Predictive Mean and Variance

Given the parameters, the compound Poisson mean and variance are $E[S\mid\Lambda,\Theta]=t\Lambda m(\Theta)$ and $\operatorname{Var}(S\mid\Lambda,\Theta)=t\Lambda q(\Theta)$. Applying the laws of total expectation and total variance gives the unconditional predictive moments.

E[S]=tE[\Lambda m(\Theta)]

\operatorname{Var}(S) =tE[\Lambda q(\Theta)] +t^2\operatorname{Var}(\Lambda m(\Theta)).

The first variance term averages the future claim-process variability over parameter uncertainty. It includes both Poisson count variation and the variation in individual claim amounts. The second term measures the variation in the conditional aggregate mean as the unknown parameters vary over their posterior distribution.

Why the Variance Formula Holds

For a fixed pair $(\Lambda,\Theta)$, a Poisson count with mean $t\Lambda$ yields the compound Poisson moments. To see the variance, condition first on the number of claims. The conditional mean given $N$ and $\Theta$ is $Nm(\Theta)$, and the conditional variance is $N\operatorname{Var}(X_j\mid\Theta)$. Therefore,

\begin{aligned} \operatorname{Var}(S\mid\Lambda,\Theta) &=E[N\mid\Lambda]\operatorname{Var}(X_j\mid\Theta) +\operatorname{Var}(N\mid\Lambda)m(\Theta)^2\\ &=t\Lambda\operatorname{Var}(X_j\mid\Theta) +t\Lambda m(\Theta)^2\\ &=t\Lambda q(\Theta), \end{aligned}

because a Poisson random variable has equal conditional mean and variance, both $t\Lambda$, and $q(\Theta)=\operatorname{Var}(X_j\mid\Theta)+m(\Theta)^2$. Now apply the law of total variance:

\begin{aligned} \operatorname{Var}(S) &=E[\operatorname{Var}(S\mid\Lambda,\Theta)] +\operatorname{Var}(E[S\mid\Lambda,\Theta])\\ &=tE[\Lambda q(\Theta)] +\operatorname{Var}(t\Lambda m(\Theta)). \end{aligned}

Since $t$ is fixed, $\operatorname{Var}(t\Lambda m(\Theta))=t^2\operatorname{Var}(\Lambda m(\Theta))$, establishing the stated result. If $\Lambda$ and $\Theta$ are independent in the posterior, products of their moments may sometimes be factored; if they are dependent, the joint expectation and variance must be used as written.

Key result: For a compound Poisson aggregate with uncertain posterior parameters, calculate $E[S]$ from the posterior expectation of the conditional aggregate mean. Calculate $\operatorname{Var}(S)$ as the expected conditional compound Poisson variance plus the posterior variance of the conditional aggregate mean.

Worked Applications

Worked Example: Uncertain Frequency and a Fixed Claim Amount

A hypothetical book has future exposure $t=2$. Its posterior annual frequency is $\Lambda\sim\operatorname{Gamma}(4,2)$ in shape-rate form. Conditional on $\Lambda$, the future count is Poisson. Every claim amount is exactly $\$400.00$. Find the predictive mean and variance of aggregate loss. No interest or mortality component is involved.

Method and posterior moments. For a shape-rate Gamma distribution, the mean is shape divided by rate and the variance is shape divided by the squared rate. Thus $E[\Lambda]=4/2=2$, $\operatorname{Var}(\Lambda)=4/2^2=1$, correction: for shape $4$, the variance is $4/2^2=1$. Therefore $E[\Lambda^2]=\operatorname{Var}(\Lambda)+E[\Lambda]^2=1+4=5$. The conditional severity mean is $m=400$, and its second moment is $q=400^2=160{,}000$.

Predictive mean. Substitute into the aggregate mean formula:

E[S]=2(2)(400)=\$1{,}600.00.

Predictive variance. The expected conditional variance is $tE[\Lambda q]=2(2)(160{,}000)=640{,}000$. The parameter-uncertainty term is $t^2\operatorname{Var}(\Lambda m)=4(400^2)\operatorname{Var}(\Lambda)=4(160{,}000)(1)=640{,}000$. Hence

\operatorname{Var}(S)=640{,}000+640{,}000 =1{,}280{,}000\ \text{dollars}^2.

Check and trap avoided. The mean agrees with expected count $tE[\Lambda]=4$ times $\$400.00$. A common mistake is to use the variance $2$ for $\operatorname{Gamma}(4,2)$, confusing the rate with a scale parameter; the shape-rate variance is $4/2^2=1$. This example's stated posterior parameters therefore produce variance $\$1{,}280{,}000$, not the result obtained under a different Gamma convention. No rounding was needed.

Worked Example: A Joint Posterior for Frequency and Severity

A hypothetical risk has one unit of future exposure. Its joint posterior parameter distribution has two equally likely states: $(\Lambda,\Theta)=(1,600)$ or $(3,1{,}000)$, with probability $0.5$ each. Conditional on the state, $N$ is Poisson with mean $\Lambda$, and claim amounts are exponential with mean $\Theta$. Find the predictive aggregate mean and variance. No interest or mortality component is involved.

Method and severity moments. For an exponential severity with mean $\Theta$, $m(\Theta)=\Theta$ and $q(\Theta)=E[X_j^2\mid\Theta]=2\Theta^2$. In the first state, $m=600$ and $q=2(600^2)=720{,}000$. In the second state, $m=1{,}000$ and $q=2(1{,}000^2)=2{,}000{,}000$.

Predictive mean. The conditional aggregate means in the two states are $1(600)=600$ and $3(1{,}000)=3{,}000$. Their posterior average is

E[S]=\frac{600+3{,}000}{2}=\$1{,}800.00.

Expected conditional variance. The values of $\Lambda q(\Theta)$ are $1(720{,}000)=720{,}000$ and $3(2{,}000{,}000)=6{,}000{,}000$. Their average is $3{,}360{,}000$. Since $t=1$, this is also $tE[\Lambda q(\Theta)]$.

Parameter-uncertainty term. The conditional means, 600 and 3,000, differ from their mean of 1,800 by $-1,200$ and $1,200$. Thus $\operatorname{Var}(\Lambda m(\Theta))=[(-1{,}200)^2+(1{,}200)^2]/2=1{,}440{,}000$.

Total and check. The predictive variance is $3{,}360{,}000+1{,}440{,}000=4{,}800{,}000$ dollars squared. The two contributions are nonnegative, and the mean is the average of the two state-specific means. The joint states matter: frequency and severity parameters are associated here, so replacing their joint moments with products of separate averages would be inappropriate. No rounding was needed.

Worked Example: Independent Gamma Frequency and Severity Parameters

For one unit of future exposure, suppose the posterior parameters are independent, with $\Lambda\sim\operatorname{Gamma}(2,1)$ and $\Theta\sim\operatorname{Gamma}(2,0.002)$, both in shape-rate form. Conditional on $\Theta$, claim amounts are exponential with mean $\Theta$; conditional on $\Lambda$, the count is Poisson with mean $\Lambda$. Find the predictive aggregate mean and variance. No interest or mortality component is involved.

Method and parameter moments. For $\Lambda$, $E[\Lambda]=2/1=2$, $\operatorname{Var}(\Lambda)=2/1^2=2$, and $E[\Lambda^2]=2+2^2=6$. For $\Theta$, $E[\Theta]=2/0.002=1{,}000$, $\operatorname{Var}(\Theta)=2/0.002^2=500{,}000$, and $E[\Theta^2]=500{,}000+1{,}000^2=1{,}500{,}000$.

Predictive mean. The exponential severity mean is $m(\Theta)=\Theta$. Independence gives

E[S]=E[\Lambda]E[\Theta]=2(1{,}000)=\$2{,}000.00.

Expected conditional variance. For exponential severity, $q(\Theta)=2\Theta^2$, so $E[q(\Theta)]=2(1{,}500{,}000)=3{,}000{,}000$. Independence then gives $E[\Lambda q(\Theta)]=2(3{,}000{,}000)=6{,}000{,}000$.

Parameter-uncertainty term. Independence also gives $E[(\Lambda\Theta)^2]=E[\Lambda^2]E[\Theta^2]=6(1{,}500{,}000)=9{,}000{,}000$. Therefore,

\operatorname{Var}(\Lambda\Theta) =9{,}000{,}000-(2{,}000)^2 =5{,}000{,}000.

With $t=1$, total predictive variance is $6{,}000{,}000+5{,}000{,}000=11{,}000{,}000$ dollars squared. The mean and variance use the posterior severity moments, including $E[\Theta^2]$, rather than only $E[\Theta]$. A common mistake is to omit the second-moment contribution from the conditional aggregate variance or to use $\operatorname{Var}(\Theta)$ in place of $E[\Theta^2]$. No rounding was needed.

Separating the Sources of Uncertainty

The identity $q(\Theta)=\operatorname{Var}(X_j\mid\Theta)+m(\Theta)^2$ makes the expected conditional variance term more transparent:

tE[\Lambda q(\Theta)] =tE[\Lambda\operatorname{Var}(X_j\mid\Theta)] +tE[\Lambda m(\Theta)^2].

The first component represents variation in claim amounts around their conditional mean. The second reflects Poisson count variation: even with a fixed severity mean, the number of claims fluctuates. The additional term $t^2\operatorname{Var}(\Lambda m(\Theta))$ accounts for posterior uncertainty in the conditional expected aggregate. These sources should not be conflated; in particular, a fixed claim amount has zero severity variance but aggregate loss can still vary because the count and frequency parameter vary.

A reliable calculation follows the same sequence each time: determine the conditional severity mean and second moment, calculate the posterior expectation of the conditional aggregate variance, then calculate the posterior variance of the conditional aggregate mean. Keep frequency exposure $t$ in both places: it appears to the first power in the expected conditional variance and to the second power in the parameter-uncertainty term.

Common exam trap: Using only $tE[\Lambda q(\Theta)]$ reports the expected conditional variance, not the full predictive variance. For example, in the first worked example that shortcut gives $640{,}000$, while the posterior variance of the conditional mean contributes another $640{,}000$. The correct predictive variance is $1{,}280{,}000$ dollars squared.

Aggregate predictive variance is useful when assessing the range of outcomes implied by a ratemaking model, especially when the available experience is limited and posterior parameter uncertainty is material. The formula is specific to the stated conditional Poisson frequency and conditionally independent severity setup. A different frequency distribution, dependence among claim amounts, or dependence between frequency and severity beyond the modeled parameters may require a different conditional variance calculation.

Check Your Understanding

For each hypothetical scenario, calculate the requested predictive aggregate moments. Gamma distributions use shape-rate form, and exponential distributions are parameterized by their mean. No interest or mortality component is involved.

For exposure $t=1.5$, let $\Lambda\sim\operatorname{Gamma}(3,1.5)$ and let every claim amount be $\$250.00$. Find the predictive mean and variance. (Answer: mean $\$750.00$; variance $375{,}000$ dollars squared.)
For $t=1$, let the frequency be fixed at $\Lambda=2$. The posterior severity mean $\Theta$ is 500 or 900, each with probability $0.5$, and claim amounts are exponential conditional on $\Theta$. Find the predictive mean and variance. (Answer: mean $\$1{,}400.00$; variance $2{,}280{,}000$ dollars squared.)
For $t=2$, let $\Lambda\sim\operatorname{Gamma}(3,1.5)$, and let claim amounts be exponential with fixed mean $\$400.00$. Find the predictive mean and variance. (Answer: mean $\$1{,}600.00$; variance $2{,}133{,}333.33$ dollars squared.)
For $t=1$, suppose the joint posterior assigns probability $0.5$ to each state $(\Lambda,\Theta)=(0.5,400)$ and $(2,1{,}000)$. Claim amounts are exponential conditional on $\Theta$, and counts are Poisson conditional on $\Lambda$. Find the predictive mean and variance. (Answer: mean $\$1{,}100.00$; variance $2{,}890{,}000$ dollars squared.)

What You'll Learn