Underlying Risk and Observed Claims
Two insureds can face the same coverage and exposure period yet have different underlying propensities to claim. At the same time, one insured can have very different claim outcomes from one period to the next even if that underlying propensity stays the same. A useful model separates these two sources of variation: persistent differences among risks and random outcomes for a given risk.
The Bühlmann risk parameter is the model’s representation of an insured’s underlying risk characteristics. It is not a claim, a claim count, or an observed average. Rather, it identifies which conditional claim distribution applies to that insured. This distinction is foundational to credibility models and aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus.
For example, if a policyholder has a fixed but unobserved annual claim probability \(\theta\), an annual claim indicator \(I\), conditional on \(\Theta=\theta\), can be modeled as Bernoulli with probability \(\theta\). The parameter \(\theta\) describes the risk’s underlying claim propensity; the realized indicator \(I\) is the observed outcome. A policyholder with \(\theta=0.10\) can have a year with no claim, and a policyholder with \(\theta=0.02\) can have a year with a claim.
A risk parameter can be a single number or a vector. It might represent a claim probability, a claim frequency rate, a severity-distribution parameter, or several characteristics together. The choice depends on the outcome being modeled. The model must specify the conditional distribution before any calculations using the parameter are meaningful.
Conditional Distributions and Risk Heterogeneity
A conditional distribution describes outcomes for risks with a particular parameter value. In notation, \(P(X\in A\mid \Theta=\theta)\) is the probability that an outcome \(X\) belongs to a set \(A\), given that the risk parameter equals \(\theta\). Different parameter values can imply different claim behavior even when the form of the conditional distribution is otherwise the same.
The parameter itself may be treated as fixed for a particular insured over the period being modeled, while differing across insureds in a portfolio. To describe a portfolio, a model may assign a probability distribution to \(\Theta\). This distribution represents how risks are distributed across underlying risk types; it does not mean the parameter changes randomly from year to year for a given insured.
| Model component | Meaning | Example |
|---|---|---|
| \(\Theta\) | Underlying risk characteristic | An insured’s claim probability |
| \(X\mid\Theta=\theta\) | Outcome distribution for a specified risk type | A Bernoulli claim indicator with probability \(\theta\) |
| Observed \(X\) | Realized outcome for an insured and period | Zero or one claim indicator |
| Distribution of \(\Theta\) | Representation of risk types in a portfolio | A specified mix of low- and high-frequency risks |
This setup is not the same as assuming that every insured has the portfolio-average claim probability. With a latent parameter, different insureds can have different conditional distributions. The population distribution of \(\Theta\) describes that heterogeneity; the conditional model describes random outcomes within each risk type.
Worked Example: A Two-Type Claim-Probability Model
A hypothetical insurer models one-year claim indicators for a book of policies using two latent risk types. A randomly selected policy has \(\Theta=0.04\) with probability \(0.80\) and \(\Theta=0.16\) with probability \(0.20\). Conditional on its risk parameter, the policy has independent Bernoulli claim indicators in each of two years. Find the probability of no claims, exactly one claim, and claims in both years.
Model and assumptions. For each year, the indicator is Bernoulli with success probability \(\theta\), conditional on \(\Theta=\theta\). The two annual indicators are independent conditional on \(\Theta\). The parameter remains the same for the policy in both years.
For a fixed value \(\theta\), the conditional probabilities are \((1-\theta)^2\) for no claims, \(2\theta(1-\theta)\) for exactly one claim, and \(\theta^2\) for claims in both years.
No claims. For the low-risk type, the probability is \((1-0.04)^2=0.96^2=0.9216\). For the high-risk type, it is \((1-0.16)^2=0.84^2=0.7056\). Weight these conditional probabilities by the type proportions:
Exactly one claim. The low-risk conditional probability is \(2(0.04)(0.96)=0.0768\); the high-risk conditional probability is \(2(0.16)(0.84)=0.2688\). Therefore,
Claims in both years. The corresponding conditional probabilities are \(0.04^2=0.0016\) and \(0.16^2=0.0256\). Thus,
Check. The three mutually exclusive outcomes have probabilities summing to \(0.8784+0.1152+0.0064=1.0000\). The type parameter is not itself an observed claim; it determines the conditional probabilities used for each possible outcome.
Common mistake avoided: using \(0.80(0.04)+0.20(0.16)=0.064\) as if every policy had claim probability \(0.064\), then calculating two-year probabilities from that single value. That collapses the risk types into one average and does not preserve the specified mixture of conditional distributions.
Why Mixture Probabilities Use Conditional Probabilities
When a risk parameter has several possible values, a probability for the portfolio is obtained by combining the probabilities conditional on each value. This follows directly from partitioning the population into its parameter types. For a discrete parameter with possible values \(\theta_1,\ldots,\theta_k\), the events \(\{\Theta=\theta_j\}\) are mutually exclusive and cover the population. Therefore, for any outcome event \(A\),
For a continuous parameter with density \(f_\Theta(\theta)\), the corresponding mixture is an integral:
These formulas are applications of the law of total probability. They explain why one first determines the outcome probability for a specified risk type and then averages those probabilities using the distribution of types. They do not imply that the parameter is observed, nor that the outcomes for a given insured are deterministic.
State what latent characteristic it represents and its possible values or distribution.
Give the distribution of the claim outcome for a fixed parameter value, including any conditional-independence assumptions.
Find the requested probability conditional on each possible parameter value.
Use the discrete weighted sum or continuous integral, as appropriate.
Risk Parameters Can Describe Severity as Well as Frequency
A latent parameter need not govern whether a claim occurs. It can instead govern the distribution of claim sizes, or another outcome of interest. Separating the parameter from the observed claim amount is equally important in a severity model: the parameter identifies a conditional severity distribution, while each reported amount is a realization from that distribution.
Worked Example: A Latent Parameter for Large-Claim Probability
A hypothetical specialty policy has three claims in a year. Each claim is either a small claim of $1,000 or a large claim of $9,000. The insured’s latent type has probability \(0.70\) of being the low-large-claim type, for which each claim is large with probability \(0.10\); with probability \(0.30\), the insured is the high-large-claim type, for which each claim is large with probability \(0.40\). Conditional on type, claim sizes are independent. Find the probability that exactly two of the three claims are large.
Model and assumptions. Let \(\Theta\) be the conditional probability that a claim is large. It equals \(0.10\) with probability \(0.70\), or \(0.40\) with probability \(0.30\). Given \(\Theta=\theta\), the number of large claims among three claims has a binomial distribution with three trials and probability \(\theta\).
For a fixed \(\theta\), the probability of exactly two large claims is \(\binom{3}{2}\theta^2(1-\theta)=3\theta^2(1-\theta)\). For the low-large-claim type, this is \(3(0.10)^2(0.90)=3(0.01)(0.90)=0.027\). For the high-large-claim type, it is \(3(0.40)^2(0.60)=3(0.16)(0.60)=0.288\).
Weighting by the probabilities of the two types gives
Check. The requested probability is \(0.1053\), or \(10.53\%\). The $1,000 and $9,000 amounts define the severity categories, but the event asks only how many claims fall in the large category; no dollar total is needed.
Common mistake avoided: multiplying the total claim count by an averaged large-claim probability and treating the three claims as if the insured’s type were redrawn for each claim. The model states that one latent type applies to the insured and that claim sizes are independent conditional on that type.
Continuous Risk Parameters
A risk parameter need not have just a few categories. For example, a portfolio model may represent claim probabilities by a continuous distribution. Conditional on a particular parameter value, outcomes still follow a specified distribution; portfolio probabilities are then obtained by integrating over possible values. The conditional model and the distribution assigned to the parameter are both necessary parts of the specification.
Worked Example: Mixing Over a Continuous Claim Probability
For a hypothetical group of policies, each insured has a latent annual claim probability \(\Theta\) distributed uniformly from \(0.02\) to \(0.10\). Conditional on \(\Theta=\theta\), the insured’s claim indicators in two years are independent Bernoulli variables with probability \(\theta\). Find the probability that the insured has claims in both years, and compare it with the result from incorrectly assigning every insured the midpoint probability.
Model and assumptions. The density of \(\Theta\) is \(1/(0.10-0.02)=12.5\) on the interval from \(0.02\) to \(0.10\). Given \(\Theta=\theta\), the probability of claims in both years is \(\theta^2\).
The mixture probability is the integral of this conditional probability against the density:
The midpoint of the interval is \(0.06\). If one incorrectly replaces the heterogeneous risk parameters with that single value, the two-year probability is \(0.06^2=0.0036\). The mixture result is \(0.0041333\), not \(0.0036\), because the conditional two-year probability is calculated for each risk type before averaging over the distribution of types. To four decimal places, the correct probability is \(0.0041\), compared with \(0.0036\) under the midpoint substitution.
Check. The integral is over the full support of \(\Theta\), and the density integrates to \(12.5(0.10-0.02)=1\). The midpoint calculation is a comparison, not a valid replacement for the specified mixture model.
Common mistake avoided: using a representative parameter value in a multi-period outcome probability without checking whether that substitution preserves the model. In this example, squaring the midpoint probability does not equal mixing the squared conditional probabilities.
Interpretation and Limits of the Model
The latent risk parameter is a modeling device for representing underlying differences among insureds. It is not necessarily a directly measurable physical trait, and the model does not require that an insurer know the parameter for each policy. Instead, the assumed distribution of \(\Theta\) and the conditional outcome model together describe how observed outcomes can arise from a heterogeneous portfolio.
The distinction is useful whenever repeated observations are available or when risks differ in underlying behavior. Observed claims include random variation even after a risk type is fixed. Conversely, observed differences between insureds may reflect both random variation and underlying heterogeneity. A model that ignores this distinction may misrepresent probabilities across multiple periods or among groups of insureds.
The risk parameter alone does not determine a model. A complete specification says what outcome is being modeled, how that outcome is distributed conditional on the parameter, what values or distribution the parameter can take, and whether repeated outcomes are independent conditional on it. If the parameter is allowed to change over time, or if outcomes are dependent for other reasons, the conditional model must reflect those features. A fixed-parameter model should not be used where that assumption is unsuitable.
Check Your Understanding
For each problem, distinguish the latent risk parameter from the observed outcome and apply the stated conditional model.
- A hypothetical portfolio has claim probability \(\Theta=0.02\) for \(90\%\) of policies and \(\Theta=0.12\) for \(10\%\). Conditional on \(\Theta\), one annual claim indicator is Bernoulli with probability \(\Theta\). What is the probability that a randomly selected policy has a claim? (Answer: \(0.0300\).)
- A policy has \(\Theta=0.05\) with probability \(0.60\) and \(\Theta=0.25\) with probability \(0.40\). Its two annual indicators are independent conditional on \(\Theta\). What is the probability of no claims in either year? (Answer: \(0.7665\).)
- For each of two claims on a hypothetical policy, severity is either small or large. The latent probability of a large claim is \(0.08\) for \(75\%\) of policies and \(0.32\) for \(25\%\), with claim severities independent conditional on type. What is the probability both claims are large? (Answer: \(0.0304\).)
- An insured’s annual claim probability \(\Theta\) is uniform from \(0.01\) to \(0.07\). Conditional on \(\Theta\), annual claim indicators are independent Bernoulli variables. What is the probability of at least one claim over two years? (Answer: \(0.0781\).)
- A hypothetical risk has a latent annual Poisson claim rate \(\Lambda=0.5\) with probability \(0.80\), or \(\Lambda=2\) with probability \(0.20\). Conditional on its rate, annual counts in two years are independent Poisson variables with that rate. What is the probability of zero claims over both years? (Answer: \(0.2980\).)