Risk Classes and Conditional Claim Probabilities
An insurer may divide a hypothetical portfolio into risk classes, such as lower-risk and higher-risk applicants. Each class can have a different claim probability, and the classes may make up different proportions of the portfolio. The overall claim probability depends on both pieces of information: the class-specific probabilities and the proportions in each class.
Once a claim is observed, the question changes. Instead of asking for the probability of a claim given a class, we may ask for the probability that the claimant belonged to a particular class. These are different conditional probabilities. The law of total probability combines class-specific rates into a portfolio rate; Bayes’ theorem uses that portfolio rate to update the probabilities of the classes.
For each model below, the class is treated as fixed for the policyholder over the period being studied, and the stated claim probabilities apply to that period. No interest-rate or mortality assumptions are needed: this is a probability model for claim occurrence, not a present-value or life-contingency calculation. Any portfolio and claim data used are hypothetical.
“Prior” means before using the observed claim outcome, not necessarily before the policy was issued. The class proportions must describe the population to which the calculation applies. If the portfolio mix changes, the overall claim probability and posterior class probabilities can change even when all class-specific claim rates stay the same.
Combining Class Rates: The Law of Total Probability
Because the classes are mutually exclusive and exhaustive, a claim must occur in exactly one of the classes. Thus, the claim event can be split into the disjoint events \(A\cap C_1,\ldots,A\cap C_k\). For each class, the multiplication rule gives \(P(A\cap C_j)=P(A\mid C_j)P(C_j)\). Adding those disjoint probabilities yields the portfolio claim probability.
The formula requires a complete partition: every policyholder must belong to one of the listed classes, and no policyholder may belong to two of them at once. If a class is omitted, the class proportions will not represent the full population, and the calculation will not give the portfolio probability.
Updating the Class Probability: Bayes’ Theorem
After a claim, the posterior probability of class \(C_j\) is the probability that both the policyholder is in \(C_j\) and a claim occurs, divided by the probability that a claim occurs. The numerator is \(P(A\mid C_j)P(C_j)\), and the denominator is the total probability across all classes.
For a nonclaim, use the same reasoning with the event \(A^c\). Within class \(C_j\), its probability is \(P(A^c\mid C_j)=1-P(A\mid C_j)\). It is not generally correct to assume that observing no claim leaves the class probabilities unchanged: a nonclaim may be more likely in one class than another.
Worked Applications
A reliable calculation keeps three quantities separate: the class proportion, the class-specific outcome probability, and the posterior class probability. The worked examples show how to calculate each one and check that the results are coherent.
Worked Example: A Claim in a Two-Class Portfolio
A hypothetical portfolio has 70% of policyholders in class \(L\) and 30% in class \(H\). Over one year, the probability of at least one claim is 0.04 for class \(L\) and 0.12 for class \(H\). Find (a) the overall one-year claim probability, (b) the probability that a policyholder with a claim was in class \(H\), and (c) the probability that a policyholder with no claim was in class \(H\).
Model and assumptions. The two classes are mutually exclusive and exhaustive, with \(P(L)=0.70\) and \(P(H)=0.30\). The class-specific one-year claim probabilities are \(P(A\mid L)=0.04\) and \(P(A\mid H)=0.12\). Interest and mortality assumptions are not applicable to this claim-occurrence model.
(a) Overall claim probability. Apply the law of total probability:
The overall one-year claim probability is 0.064, or 6.4%.
(b) Class \(H\) given a claim. The joint probability of class \(H\) and a claim is \(P(A\mid H)P(H)=(0.12)(0.30)=0.036\). Therefore,
Given a claim, the probability the policyholder was in class \(H\) is 0.5625, or 56.25%.
(c) Class \(H\) given no claim. The probability of no claim is \(1-0.064=0.936\). The joint probability of class \(H\) and no claim is \((0.30)(1-0.12)=(0.30)(0.88)=0.264\). Thus,
Check. The class \(H\) posterior is higher after a claim (0.5625) than before the claim (0.30), and lower after a nonclaim (0.2821). That direction is consistent with class \(H\) having the higher claim probability. The common mistake to avoid is reporting \(P(A\mid H)=0.12\) as the answer to part (b): 0.12 is the claim probability for someone already known to be in \(H\), not the probability of \(H\) after observing a claim.
Worked Example: Three Classes and a Nonclaim
A hypothetical pool has 50% of members in class \(A\), 30% in class \(B\), and 20% in class \(C\). Their respective one-year claim probabilities are 0.02, 0.08, and 0.20. Find the overall claim probability and the probability that a policyholder with no claim was in class \(C\).
Model and assumptions. The three classes form a complete partition. The stated values are class proportions and one-year claim probabilities. No interest or mortality assumptions are involved.
Overall claim probability. Weight each class rate by its class proportion:
So \(P(A^c)=1-0.074=0.926\). The joint probability of being in class \(C\) and having no claim is
Bayes’ theorem now gives
Thus, after observing no claim, the probability of class \(C\) is 0.1728, or 17.28%. The result is less than the prior proportion of 0.20 because class \(C\) has the highest claim probability, so a nonclaim is relatively less likely in that class.
Check. The joint probabilities of a nonclaim in classes \(A\), \(B\), and \(C\) are \((0.50)(0.98)=0.490\), \((0.30)(0.92)=0.276\), and 0.160. They sum to \(0.490+0.276+0.160=0.926\), the nonclaim probability used in the denominator.
Worked Example: Two-Year Experience and Conditional Independence
A hypothetical portfolio has 60% of policyholders in class \(L\) and 40% in class \(H\). The annual probability of a claim is 0.05 in \(L\) and 0.15 in \(H\). Given a policyholder’s class, assume claim occurrence in the two years is independent and the annual claim probability remains constant. Find the probability of at least one claim over two years and the probability of class \(H\) given at least one claim.
Model and assumptions. Classes are exhaustive and mutually exclusive. Within each class, annual outcomes are conditionally independent, with the stated class-specific probability in each year. Independence is assumed conditional on class; it is not assumed across the whole portfolio without conditioning on class.
For class \(L\), the probability of no claim in either year is \((1-0.05)^2=(0.95)^2=0.9025\). Therefore, the probability of at least one claim in two years is \(1-0.9025=0.0975\). For class \(H\), the corresponding probability is \(1-(0.85)^2=1-0.7225=0.2775\).
Overall probability of at least one claim. Let \(B\) denote that event. Apply total probability:
Posterior probability of class \(H\). The joint probability of class \(H\) and event \(B\) is \((0.40)(0.2775)=0.1110\). Hence,
The probability of class \(H\), given at least one claim over two years, is 0.6549. As a check, the two class contributions to \(P(B)\) sum to 0.1695, and the posterior is between 0 and 1. The key trap is to use the one-year rates 0.05 and 0.15 directly as the probabilities of at least one claim over two years. The two-year probabilities must account for both years, here using conditional independence within each class.
A Short Justification of Bayes’ Formula
Bayes’ theorem in this setting follows directly from the definition of conditional probability and the partition of the population. For any class \(C_j\) with \(P(A)>0\), the definition gives \(P(C_j\mid A)=P(C_j\cap A)/P(A)\). The multiplication rule gives \(P(C_j\cap A)=P(A\mid C_j)P(C_j)\). Finally, since the classes partition the population, the law of total probability expresses \(P(A)\) as the sum of these joint probabilities. Substitution gives the posterior formula.
The denominator must be positive: if \(P(A)=0\), conditioning on a claim is undefined in the usual conditional-probability formula. For ordinary claim models with positive class proportions and positive claim probabilities in at least one class, this condition is satisfied.
Interpretation and Common Traps
A posterior class probability is an updated probability, not a definitive classification of an individual. For example, a value of 0.6549 means that, under the stated model and population mix, the probability of class \(H\) among policyholders with the specified experience is 0.6549. It does not mean that every claimant with that experience must be in class \(H\).
- Reversing the condition. \(P(A\mid C_j)\) asks for a claim probability among members of a known class. \(P(C_j\mid A)\) asks for a class probability among those with an observed claim. They are generally unequal.
- Ignoring class proportions. A high class-specific rate does not by itself determine the portfolio claim probability or posterior. The class’s population share matters.
- Using the wrong denominator. For a posterior after a claim, divide the joint probability for the class and a claim by the overall claim probability—not by the class-specific rate.
- Forgetting the complement rate. For a nonclaim, use \(1-P(A\mid C_j)\) in each class, then combine those probabilities with the class proportions.
- Applying independence at the wrong level. In multi-year models, conditional independence within class does not imply unconditional independence in the mixed portfolio. The class mix can create dependence between outcomes when class is not known.
- Leaving out a class. The total-probability calculation requires all classes and their proportions to sum to 1.
Check Your Understanding
For each problem, identify the relevant class proportions and conditional outcome probabilities before calculating. Give probabilities to four decimal places where needed.
- A hypothetical portfolio is 80% class \(L\) and 20% class \(H\). One-year claim probabilities are 0.03 and 0.13, respectively. Find the overall claim probability and \(P(H\mid\text{claim})\). (Answer: overall probability \(0.0500\); posterior \(0.5200\).)
- A hypothetical pool has class proportions 0.45, 0.35, and 0.20 for classes \(A,B,C\), with claim probabilities 0.02, 0.06, and 0.16. Find the overall claim probability and \(P(B\mid\text{claim})\). (Answer: overall probability \(0.0620\); posterior \(0.3387\).)
- A hypothetical portfolio has 65% class \(L\) and 35% class \(H\), with respective one-year claim probabilities 0.04 and 0.10. Find the probability of no claim and \(P(H\mid\text{no claim})\). (Answer: no-claim probability \(0.9390\); posterior \(0.3355\).)
- A hypothetical two-class pool is 70% class \(L\) and 30% class \(H\), with annual claim probabilities 0.03 and 0.12. Assume annual outcomes are independent conditional on class, with probabilities constant over two years. Find \(P(H\mid\text{exactly one claim in two years})\). (Answer: \(0.6086\).)
- A hypothetical portfolio has proportions 0.50, 0.30, and 0.20 in classes \(A,B,C\), with one-year claim probabilities 0.01, 0.05, and 0.25. Find the overall claim probability, \(P(C\mid\text{claim})\), and \(P(C\mid\text{no claim})\). (Answer: overall probability \(0.0700\); posterior given claim \(0.7143\); posterior given no claim \(0.1613\).)