Why Risk-Specific Means Matter
Two policyholders can have different expected claim costs even when each policyholder’s claims are random. The collective risk model represents this with a latent risk parameter: an unobserved characteristic that helps determine the distribution of a risk’s claims. Once that parameter is fixed, the model describes the claim observations for that risk. The resulting conditional expected value is the risk-specific mean.
This distinction is central to credibility theory. An observed claim total can differ from its risk-specific mean just by chance. Separately, the risk-specific means themselves can differ across risks. The first is variation within a risk; the second is variation between risks. This tutorial defines these elements and shows how to calculate conditional means, including for compound aggregate claims. It is aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus.
The parameter is called latent because it is not directly observed as a label in the claim data. It may represent unmeasured differences in risk characteristics. The model does not require that \(\Theta\) be a single numerical measurement in practice; it can encode risk classes or other unobserved heterogeneity.
Conditional independence and identical distributions are assumptions of the basic model, not automatic facts about real claims. They mean that, once \(\Theta\) is given, each period’s observation follows the same distribution and knowing one period’s observation does not alter the distribution of another. Without these assumptions, formulas relying on them must be reconsidered.
Conditional Observations and Two Sources of Variation
For a fixed value \(\theta\), the conditional distribution of \(X_j\) describes the possible claim outcomes for a risk with that parameter value. The conditional mean \(m(\theta)\) summarizes the expected observation for that risk; it does not say that every observation equals that mean. A risk can have a low expected claim amount and still experience a large claim.
Across risks, \(\Theta\) can take different values, and therefore the risk-specific mean \(m(\Theta)\) can differ. If \(\Theta\) is held fixed, random observations can still vary around \(m(\Theta)\). If the model considers different risks, their conditional means can vary with their latent parameters. Keeping these two comparisons separate is essential: an individual risk’s noisy experience is not itself the same thing as a difference in expected claims between risks.
| Question | What is held fixed? | What varies? |
|---|---|---|
| Within-risk variation | The risk parameter \(\Theta=\theta\) | Claim observations \(X_j\) around \(m(\theta)\) |
| Between-risk variation | The model for generating risks | The risk parameter and its associated mean \(m(\Theta)\) |
Worked Example: Risk-Specific Means for Claim Counts
A hypothetical book has two latent risk types. A randomly selected risk has type A with probability \(0.6\) and type B with probability \(0.4\). In each year, its claim count \(N\), conditional on type, is Poisson: the mean is \(1.5\) for type A and \(3\) for type B. Find each risk-specific mean count and the unconditional mean count for a randomly selected risk.
Method: condition on the latent type first. For a Poisson random variable, the mean equals its parameter. Thus, for type A, \(E[N\mid\Theta=A]=1.5\), and for type B, \(E[N\mid\Theta=B]=3\). These are the two risk-specific means.
For a randomly selected risk, average the conditional means using the type probabilities: \(E[N]=0.6(1.5)+0.4(3)=0.9+1.2=2.1\) claims per year.
The result is checked by the weights: \(0.6+0.4=1\), and \(2.1\) lies between the two conditional means \(1.5\) and \(3\). The calculation distinguishes the expected count for a type from the expected count across the mixed population. It does not imply that an individual risk will have exactly its conditional mean count in a year.
Risk-Specific Means for Aggregate Claims
An aggregate claim amount combines a random number of claims with random claim severities. Let \(N\) denote the claim count and \(Y_1,Y_2,\ldots\) the severities, so that \(S=\sum_{k=1}^{N}Y_k\), with \(S=0\) when \(N=0\). The conditional expected aggregate depends on the conditional expected count and the conditional severity mean.
The assumptions matter. In particular, the severity mean in this result is conditional on the latent parameter. An unconditional mean \(E[Y_k]\) alone does not establish that the severity mean is the same for every \(\theta\). If severity distributions differ by risk, their conditional means must remain attached to the corresponding risk parameter.
Why the Aggregate-Mean Formula Holds
Condition first on both the risk parameter and the claim count. Given \(\Theta=\theta\) and \(N=n\), the aggregate is a sum of \(n\) severities. Conditional independence of severity and count given \(\Theta\) ensures that conditioning on \(N=n\) does not change the conditional severity mean.
Now average over the claim count while keeping \(\theta\) fixed. The conditional expectation of \(N\) is \(\lambda(\theta)\), which gives the formula:
If the severity mean is constant across risks, so that \(\mu_Y(\theta)=\mu_Y\) for every \(\theta\), this simplifies to \(\lambda(\theta)\mu_Y\). If the conditional severity mean changes with \(\theta\), the product must use that risk-specific mean.
Worked Example: A Compound Poisson Model with Risk-Dependent Severity
Consider a hypothetical insurance portfolio with two equally likely latent risk types. Conditional on type A, \(N\) is Poisson with mean \(2\), and every claim has severity \(400\) dollars. Conditional on type B, \(N\) is Poisson with mean \(5\), and every claim has severity \(1{,}000\) dollars. Given the type, severity is independent of the count. Find the conditional expected aggregate for each type and the expected aggregate for a randomly selected risk.
Method: use the conditional frequency and severity means for each type. For type A, \(\lambda(A)=2\) and \(\mu_Y(A)=400\), so \(E[S\mid\Theta=A]=2(400)=800\) dollars.
For type B, \(\lambda(B)=5\) and \(\mu_Y(B)=1{,}000\), so \(E[S\mid\Theta=B]=5(1{,}000)=5{,}000\) dollars. Since each type has probability \(0.5\), \(E[S]=0.5(800)+0.5(5{,}000)=400+2{,}500=2{,}900\) dollars.
The answer is checked by averaging the two conditional aggregate means, with probabilities summing to one. The calculation uses the severity mean appropriate to each type; it does not assume that both types have the same expected payment per claim.
Common Exam Trap: Substituting an Unconditional Severity Mean
Suppose in the preceding example a severity \(Y\) is generated by first selecting a risk type with equal probabilities and then taking a claim from that type. Its unconditional mean is \(E[Y]=0.5(400)+0.5(1{,}000)=700\) dollars. That unconditional mean does not equal either conditional severity mean.
The incorrect calculation would substitute \(700\) dollars into each type’s conditional aggregate formula. It would produce \(2(700)=1{,}400\) dollars for type A and \(5(700)=3{,}500\) dollars for type B, rather than the correct \(800\) and \(5{,}000\). Averaging those incorrect values gives \(0.5(1{,}400)+0.5(3{,}500)=2{,}450\) dollars, not the correct \(2{,}900\).
The error is using \(E[Y]\) where the derivation requires \(E[Y\mid\Theta=\theta]\). In general, conditional on a risk type, the correct result is \(E[S\mid\Theta=\theta]=\lambda(\theta)\mu_Y(\theta)\). Only if \(E[Y\mid\Theta=\theta]=\mu_Y\) for every \(\theta\) may the common unconditional severity mean be used as \(\mu_Y\) in every risk-specific calculation.
Worked Example: Separate Conditional Claim Fluctuation from Risk Differences
A hypothetical one-year policy has a fixed payment of \(500\) dollars if a claim occurs and zero otherwise. A risk has latent type L with probability \(0.75\) or type H with probability \(0.25\). Conditional on L, the claim probability is \(0.1\); conditional on H, it is \(0.3\). Find each risk-specific mean payment, the mean payment across randomly selected risks, and interpret a year with no claim for a type H risk.
Method: calculate the conditional expectation for each type. Let \(X\) be the one-year payment. For type L, \(E[X\mid\Theta=L]=0.1(500)+0.9(0)=50\) dollars. For type H, \(E[X\mid\Theta=H]=0.3(500)+0.7(0)=150\) dollars.
The mean across randomly selected risks is \(E[X]=0.75(50)+0.25(150)=37.50+37.50=75\) dollars.
If a type H risk has no claim in a particular year, its observed payment is \(0\), but its risk-specific mean remains \(150\) dollars per year. The zero is a possible outcome under the type H conditional distribution; it does not change the type’s mean by itself. Conversely, the difference between \(50\) and \(150\) dollars reflects different risk-specific means, not merely different outcomes for the same fixed type.
The check is direct: the type probabilities sum to one, and the overall mean \(75\) dollars lies between the two conditional means. The example avoids treating one observed year as if it revealed a risk’s expected payment exactly.
How to Set Up a Collective-Model Calculation
A reliable solution separates the conditioning step from the averaging step. First identify what is random about a single risk’s claims after its latent parameter is fixed. Then calculate that risk’s conditional expected value. Only when the question asks about a randomly selected risk should the conditional means be averaged over the distribution of \(\Theta\).
Identify the risk type or parameter value and the probability model for it, if provided.
Record the distribution and parameters of the claim observation, frequency, or severity given \(\Theta=\theta\).
Use the conditional distribution; do not replace a conditional mean with an unconditional one without justification.
Use the probabilities of the latent risk types to obtain an overall expected value for a randomly selected risk.
These steps also clarify what the calculation does not establish. A mean is not a prediction of the exact outcome in a period, and averaging risk-specific means does not identify the latent type of any particular policyholder. The detailed variance formulas for within-risk and between-risk components are separate results; the essential setup here is to keep the conditional distribution and the distribution of the latent parameter distinct.
Check Your Understanding
For each problem, identify the relevant conditional mean before calculating any overall expected value.
- A randomly selected risk is type A with probability \(0.7\) and type B with probability \(0.3\). Its annual claim count is conditionally Poisson with means \(1\) and \(4\), respectively. Find both risk-specific mean counts and the overall mean count. (Answer: \(1\) and \(4\) claims; overall mean \(1.9\) claims.)
- A one-period payment is \(800\) dollars with conditional probability \(0.05\) for type L and \(0.20\) for type H, and otherwise is zero. Find the two risk-specific means. (Answer: \(40\) dollars for L and \(160\) dollars for H.)
- Two equally likely types have conditional Poisson claim counts with means \(1\) and \(3\). Severity is \(600\) dollars for each type. Find the conditional expected aggregate for each type and the overall expected aggregate. (Answer: \(600\) dollars, \(1{,}800\) dollars, and \(1{,}200\) dollars.)
- Two equally likely risk types have conditional Poisson means \(2\) and \(4\). Their conditional severity means are \(300\) and \(900\) dollars, respectively, and severities are independent of counts given type. Find the conditional aggregate means and the overall aggregate mean. (Answer: \(600\) dollars, \(3{,}600\) dollars, and \(2{,}100\) dollars.)
- In the preceding problem, a solver uses the unconditional severity mean \(600\) dollars for both risk types. What overall aggregate mean does that incorrect method produce, and what is the correct value? (Answer: Incorrect \(1{,}800\) dollars; correct \(2{,}100\) dollars.)