Conditional Means and the Individual Risk Premium

Why Separate a Risk’s Mean From Its Observed Losses?

A risk’s observed losses vary from period to period, even when the underlying risk has not changed. A small number of favorable or unfavorable observations therefore need not describe the risk’s long-run expected loss. Bühlmann credibility begins by separating the latent characteristics of a risk from the random outcomes those characteristics generate.

The key quantity for one risk is its conditional expected loss: the expected loss given that risk’s underlying characteristics. The portfolio-level counterpart is the collective mean, which averages those conditional means over the portfolio’s distribution of risks. Neither quantity is, in general, the same as the observed average for one risk.

Definition: Let $\Theta$ be a random risk parameter representing underlying characteristics that can differ across risks. For a loss $X$ in one observation period, the conditional expected loss is $\mu(\Theta)=E[X\mid\Theta]$. For a particular risk whose parameter is $\theta$, its individual risk premium (or hypothetical mean) is $\mu(\theta)=E[X\mid\Theta=\theta]$. The collective mean is $m=E[\mu(\Theta)]$, the average conditional expected loss over the portfolio’s distribution of risk parameters.

This setup treats the portfolio as a population of risks: $\Theta$ varies from risk to risk, while observations for a given risk are random. In the basic repeated-observation model, each risk has a fixed parameter over the periods being modeled, and the period losses $X_1,\ldots,X_n$ are conditionally independent and identically distributed given $\Theta$. Their conditional means and variances are finite. The observation period and exposure basis must be consistent across risks before their means are compared.

The individual risk premium is a mean conditional on the risk, not a promise about the loss in any one period. The collective mean is not necessarily the premium of a typical risk, either: it depends on the proportions of different risk types in the portfolio. The model does not require a particular distribution for $\Theta$ or for $X$, but any calculation must specify the distributions or conditional means being used.

How the Collective Mean Follows From the Risk Mix

The collective mean is the expected loss before conditioning on a particular risk’s parameter. The law of total expectation gives $E[X]=E[E[X\mid\Theta]]$. Since $E[X\mid\Theta]=\mu(\Theta)$, it follows that the unconditional expected loss is the collective mean.

m=E[\mu(\Theta)]=E[E[X\mid\Theta]]=E[X].

For a portfolio with a finite set of risk types, this is a weighted average. If type $j$ occurs with probability $p_j$ and has conditional expected loss $\mu_j$, then $m=\sum_j p_j\mu_j$, where the probabilities sum to one. For a continuous risk parameter, the same averaging is expressed as an expectation (or, when a density exists, an integral) over its distribution.

Specify the risk mix.
State the possible values or distribution of $\Theta$, which describes how risks differ in the portfolio.

Find each conditional mean.
Calculate $\mu(\theta)=E[X\mid\Theta=\theta]$ from the loss model for each risk value.

Average over the mix.
Weight the conditional means by the probabilities of their risk values to obtain $m$.

This derivation also clarifies the role of the observed average $\bar X=n^{-1}\sum_{t=1}^nX_t$. It is calculated from realized losses, so it is random even when $\Theta$ is fixed. Under the repeated-observation assumptions, $E[\bar X\mid\Theta]=\mu(\Theta)$; that equality says the observed average is centered on the individual risk premium in repeated samples, not that the two quantities are equal in a particular sample.

Worked Examples

Worked Example: Two Risk Types With Variable Annual Loss

A hypothetical insurer models one-year losses for two risk types. Type A has probability $0.40$ in the portfolio; type B has probability $0.60$. For either type, annual loss $X$ is $0$ with probability $1-p$ and $\$1{,}000$ with probability $p$. Type A has $p=0.10$, and type B has $p=0.30$. Find each individual risk premium and the collective mean. Then, for a type A risk with observed losses of $\$0$ and $\$1{,}000$ over two years, calculate its observed average and compare it with its premium.

Method. Calculate conditional expected loss for each type from its two-point loss distribution, then average those means using the portfolio type probabilities. There is no interest or mortality assumption; the specified Bernoulli loss distributions fully determine the example.

For type A, the conditional mean is $0(0.90)+\$1{,}000(0.10)=\$100$. Thus $\mu(A)=\$100$. For type B, it is $0(0.70)+\$1{,}000(0.30)=\$300$, so $\mu(B)=\$300$.

The collective mean is $m=0.40(\$100)+0.60(\$300)=\$40+\$180=\$220$ per risk-year.

The type A risk’s observed two-year average is $(\$0+\$1{,}000)/2=\$500$. Its individual risk premium remains $\$100$ per year under the stated model. The $\$500$ is the realized average in this short sample, not a replacement for the conditional expected loss.

Check. The collective mean lies between the two type-specific means, and the portfolio weights sum to $0.40+0.60=1.00$. The calculation avoids treating a high observed average from a small sample as the risk’s underlying mean.

Worked Example: A Poisson Frequency Model With Gamma Risk Variation

In a hypothetical portfolio, each risk has a latent annual claim frequency rate $\Theta$. Across risks, $\Theta\sim\operatorname{Gamma}(2,0.01)$, using the shape-rate parameterization, so the density is proportional to $\theta^{2-1}e^{-0.01\theta}$. Conditional on $\Theta=\theta$, the annual claim count $N$ has a Poisson distribution with mean $\theta$. Each claim costs exactly $\$800$, independently of frequency, and annual aggregate loss is $X=\$800N$. Find the individual risk premium as a function of $\theta$, the collective mean, and the observed average for a risk with $\theta=1.5$ and counts $1,3,0$ over three years.

Method. First calculate expected aggregate loss conditional on the frequency parameter. Then average over the Gamma risk distribution. The loss severity is deterministic at $\$800$; there is no interest or mortality assumption.

For $N\mid\Theta=\theta\sim\operatorname{Poisson}(\theta)$, $E[N\mid\Theta=\theta]=\theta$. Therefore $\mu(\theta)=E[\$800N\mid\Theta=\theta]=\$800\theta$.

For a Gamma distribution with shape $\alpha$ and rate $\beta$, the mean is $\alpha/\beta$. Here $E[\Theta]=2/0.01=200$. Hence the collective mean is $m=\$800E[\Theta]=\$800(200)=\$160{,}000$ per risk-year.

For the risk with $\theta=1.5$, the individual risk premium is $\mu(1.5)=\$800(1.5)=\$1{,}200$ per year. Its three-year observed average count is $(1+3+0)/3=4/3$ claims per year. The corresponding average loss is $\$800(4/3)=\$1{,}066.666\ldots$, or $\$1{,}066.67$ to the nearest cent.

The observed average is below the $\$1{,}200$ conditional mean, but that does not change the stated risk parameter or its premium. The three-year average is a random outcome. The Gamma distribution’s mean rate of 200 describes the risk mix, not the specified risk with $\theta=1.5$.

Check. The population mean frequency is 200 claims per risk-year, so the implied mean aggregate loss is $200\times\$800=\$160{,}000$. The calculation avoids confusing a portfolio-level expected loss with the premium for a particular risk.

Worked Example: Exposure-Adjusted Individual Premium

A hypothetical property portfolio has two risk classes. A randomly selected risk is class L with probability $0.25$ and class H with probability $0.75$. Conditional on class, claims follow a Poisson model: the annual claim rate per exposure unit is 2 for class L and 6 for class H. Each claim has a deterministic severity of $\$500$. A particular class H policy has 2.5 exposure units in the period. Find the premium per exposure unit, its expected loss for that policy period, and the observed loss per exposure unit if 13 claims occur.

Method. Use the conditional Poisson mean to find expected claim counts, multiply by fixed severity to obtain expected loss, and keep the per-unit and total-policy quantities distinct. Interest and mortality do not enter this frequency-severity model.

For class L, expected claims per exposure unit are 2, so its individual risk premium is $2(\$500)=\$1{,}000$ per exposure unit. For class H, the corresponding premium is $6(\$500)=\$3{,}000$ per exposure unit.

The collective mean per exposure unit is $0.25(\$1{,}000)+0.75(\$3{,}000)=\$250+\$2{,}250=\$2{,}500$.

For the class H policy with 2.5 exposure units, expected claim count is $6(2.5)=15$. Its expected total loss is $15(\$500)=\$7{,}500$, equivalently $2.5(\$3{,}000)=\$7{,}500$.

If 13 claims occur, the observed total loss is $13(\$500)=\$6{,}500$. The observed loss per exposure unit is $\$6{,}500/2.5=\$2{,}600$. This observed rate is not the class H individual risk premium of $\$3{,}000$ per exposure unit.

Check. The policy’s expected loss is $2.5$ times its per-unit premium, and its observed loss rate is $13/2.5=5.2$ claims per exposure unit times $\$500=\$2{,}600$. The calculation avoids applying an exposure total to a per-unit premium without scaling.

Common Mistakes and Scope

The definition of an individual risk premium depends on the conditioning information in the model. If $\Theta$ is a risk type, then $\mu(\theta)$ is the expected loss for that type. If the risk parameter is a vector of characteristics, the conditional mean is based on that vector. The model must state what $\Theta$ represents and use a consistent observation period and exposure basis.

Substituting an observed average for a conditional mean. The observed average is calculated from realized outcomes. It can differ substantially from $\mu(\theta)$, especially with few observations.
Using the collective mean as every risk’s premium. The collective mean averages over the risk mix. It equals an individual risk premium only when that risk’s conditional mean happens to equal the portfolio average.
Weighting risk types equally without justification. The collective mean uses the probabilities or portfolio proportions of the types, not an unweighted average unless the types are equally represented.
Mixing exposure bases. A premium per exposure unit must be multiplied by the applicable exposure to obtain an expected total loss.

The conditional means and collective mean define the quantities used in the Bühlmann model. They do not, by themselves, specify how much weight to assign to observed experience when estimating a premium. That is a separate credibility question. The distinction here is foundational: first identify the risk-specific expected loss and the portfolio average; do not label a realized average as either one without justification. This material is aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus.

Key takeaway: For a risk with parameter $\theta$, the individual risk premium is $\mu(\theta)=E[X\mid\Theta=\theta]$. The collective mean $m=E[\mu(\Theta)]=E[X]$ averages those premiums across the risk mix. An observed sample average is random experience, not automatically either expected value.

Check Your Understanding

For each problem, distinguish the conditional expected loss from the average across risks or observed periods.

A hypothetical portfolio has type A probability $0.30$ and type B probability $0.70$. Annual losses are $\$0$ or $\$1{,}000$: type A has probability $0.10$ of the positive loss, and type B has probability $0.25$. Find both individual risk premiums and the collective mean. (Answer: $\mu(A)=\$100$, $\mu(B)=\$250$, $m=\$205$ per year.)
For a hypothetical risk, $N\mid\Theta=\theta\sim\operatorname{Poisson}(\theta)$, each claim costs $\$1{,}200$, and the risk has $\theta=2$. It has observed counts 1 and 5 in two years. Find its individual risk premium and observed average loss. (Answer: premium $\$2{,}400$ per year; observed average $\$3{,}600$ per year.)
A portfolio has two classes with probabilities $0.40$ and $0.60$. Conditional annual claim counts are Poisson with means 3 and 7, respectively; each claim costs $\$400$. Find each class premium and the collective mean. (Answer: class premiums $\$1{,}200$ and $\$2{,}800$; collective mean $\$2{,}160$ per year.)
In a hypothetical model, $\Theta\sim\operatorname{Gamma}(3,0.02)$ with shape-rate parameters, $N\mid\Theta=\theta\sim\operatorname{Poisson}(\theta)$, and each claim costs $\$500$. Find $\mu(\theta)$ and the collective mean. (Answer: $\mu(\theta)=\$500\theta$; $m=\$75{,}000$ per risk-year.)
Two risk types have probabilities $0.50$ each and conditional means $\$900$ and $\$1{,}500$ per exposure unit. A type with the $\$1{,}500$ mean has 4 exposure units and records total loss of $\$5{,}200$. Find the collective mean per unit and the risk’s observed loss per unit. (Answer: collective mean $\$1{,}200$ per unit; observed loss $\$1{,}300$ per unit.)

What You'll Learn