Beta-Binomial Models for Claim Probabilities

Why Model Claim Probabilities as Risk-Specific?

A portfolio may contain risks with different underlying probabilities of a claim in a given exposure period. For a particular insured or exposure unit, let $\theta$ denote its underlying probability of a claim. The observed outcomes are random even when $\theta$ is fixed: a risk with a relatively high $\theta$ can have a claim-free period, and a risk with a relatively low $\theta$ can have a claim. The Beta-Binomial model represents both sources of variation: uncertainty in the risk-specific probability and random claim/no-claim outcomes conditional on that probability.

The model is suited to a fixed number of comparable opportunities for a binary outcome. For example, each opportunity might be one hypothetical policy-year, with the outcome recorded as whether at least one claim occurred. It is not automatically appropriate for claim counts that can exceed one per opportunity, or for exposures whose claim probabilities differ substantially from one trial to the next.

Definition: In the Beta-Binomial model, a risk has a claim probability $\theta$, where $0<\theta<1$. Conditional on $\theta$, the $n$ claim/no-claim outcomes are independent Bernoulli trials with probability $\theta$ of a claim. Thus, if $K$ is the number of claims in the $n$ trials, $K\mid\theta\sim\operatorname{Binomial}(n,\theta)$. Across risks, $\theta$ is modeled as a random variable with a Beta distribution.

The conditional binomial probability of observing exactly $k$ claims is

P(K=k\mid\theta)=\binom{n}{k}\theta^k(1-\theta)^{n-k}, \qquad k=0,1,\ldots,n.

Here, $n$ is the number of trials for the risk, and $k$ is the observed number of claim outcomes. If the actual ordered sequence of outcomes is specified, its likelihood is $\theta^k(1-\theta)^{n-k}$. The binomial coefficient is included when calculating the probability of the count $K=k$, because there are $\binom{n}{k}$ different sequences with that count.

The Beta Prior and Its Parameters

The Beta distribution is a flexible probability distribution on the interval from zero to one. It is parameterized here by $\alpha>0$ and $\beta>0$, both shape parameters. Its density is proportional to $\theta^{\alpha-1}(1-\theta)^{\beta-1}$, with the beta function providing the normalizing constant.

Definition: If $\theta\sim\operatorname{Beta}(\alpha,\beta)$, its density is $$ f_\theta(\theta)=\frac{\theta^{\alpha-1}(1-\theta)^{\beta-1}}{B(\alpha,\beta)},\quad 0<\theta<1, $$ where $B(\alpha,\beta)=\int_0^1 t^{\alpha-1}(1-t)^{\beta-1}\,dt$. The prior mean is $E[\theta]=\alpha/(\alpha+\beta)$, and the prior variance is $\operatorname{Var}(\theta)=\alpha\beta/[(\alpha+\beta)^2(\alpha+\beta+1)]$.

The ratio of the parameters determines the prior mean, while their sum affects how concentrated the distribution is around that mean. Multiplying both parameters by the same positive constant preserves the mean but changes the variance. For example, Beta(2, 3) and Beta(4, 6) both have mean $0.4$, but Beta(4, 6) has the smaller variance. The parameters are often described as prior claim and no-claim strength, respectively, but they are not generally literal counts of observations. In particular, the powers in the density are $\alpha-1$ and $\beta-1$; calling $\alpha$ and $\beta$ exact observed counts can lead to an incorrect interpretation.

Why the Posterior Is Also Beta

The Beta prior is conjugate to the binomial likelihood: after observing a count, the updated distribution for $\theta$ remains a Beta distribution. This result follows directly by multiplying the likelihood and prior densities. The derivation also clarifies why the claim count adds to one shape parameter and the no-claim count to the other.

Write the likelihood.
For $k$ claims in $n$ trials, the likelihood as a function of $\theta$ is proportional to $\theta^k(1-\theta)^{n-k}$. The factor $\binom{n}{k}$ does not depend on $\theta$.

Multiply by the prior density.
The prior contributes a factor proportional to $\theta^{\alpha-1}(1-\theta)^{\beta-1}$. The product is proportional to $\theta^{\alpha+k-1}(1-\theta)^{\beta+n-k-1}$.

Recognize the Beta density.
The resulting powers match a Beta distribution with parameters $\alpha+k$ and $\beta+n-k$. The normalizing constant is $B(\alpha+k,\beta+n-k)$.

Theorem: With a Beta($\alpha,\beta$) prior and $K=k$ claims in $n$ conditionally independent Bernoulli trials, the posterior distribution is $$ \theta\mid K=k\sim\operatorname{Beta}(\alpha+k,\ \beta+n-k). $$ This update describes the distribution of the risk-specific probability after the observed experience; it does not change the definition of a claim or the number of trials.

Integrating out $\theta$ gives the marginal probability of a count across risks:

P(K=k)=\binom{n}{k}\frac{B(\alpha+k,\beta+n-k)}{B(\alpha,\beta)}.

This is the Beta-Binomial probability. It accounts for heterogeneity in $\theta$ across risks: unlike a binomial model with one fixed probability shared by the whole population, it averages the conditional binomial probability over the prior distribution of risk probabilities.

Worked Example: Updating From an Ordered Outcome Record

A hypothetical small contractor is assigned a claim probability $\theta$ for each of six comparable policy periods. Before observing its experience, an analyst uses a Beta(2, 5) prior. The recorded outcomes are claim, no claim, no claim, claim, no claim, no claim. Find the likelihood of this ordered record as a function of $\theta$, the posterior distribution, and the prior mean and variance.

Method. Count the claims and no-claims, use the Bernoulli likelihood for the specified sequence, and apply the Beta-Binomial conjugate update. This avoids treating the ordered sequence as though it included the binomial coefficient for a count.

There are $n=6$ trials, $k=2$ claims, and $6-2=4$ no-claims. The probability of this particular ordered sequence conditional on $\theta$ is

\theta\cdot(1-\theta)\cdot(1-\theta)\cdot\theta\cdot(1-\theta)\cdot(1-\theta) =\theta^2(1-\theta)^4.

The prior mean is

E[\theta]=\frac{2}{2+5}=\frac{2}{7}\approx0.2857.

The prior variance is

\operatorname{Var}(\theta)=\frac{(2)(5)}{(2+5)^2(2+5+1)} =\frac{10}{49\cdot8}=\frac{5}{196}\approx0.02551.

Updating the parameters gives $\alpha+k=2+2=4$ and $\beta+n-k=5+4=9$. Therefore, $\theta\mid K=2\sim\operatorname{Beta}(4,9)$. The check is that the first updated parameter increased by the two claim outcomes and the second by the four no-claim outcomes. The common mistake is to add all six trials to the claim parameter, rather than separating claims from no-claims.

Marginal Probabilities Across Risks

A risk’s outcomes conditional on $\theta$ follow a binomial model, but the probability of a count for a randomly selected risk is obtained by averaging over the prior distribution of $\theta$. This distinction matters when interpreting portfolio-level claim counts. The calculation integrates the conditional probability against the Beta density:

P(K=k)=\int_0^1 \binom{n}{k}\theta^k(1-\theta)^{n-k} \frac{\theta^{\alpha-1}(1-\theta)^{\beta-1}}{B(\alpha,\beta)}\,d\theta =\binom{n}{k}\frac{B(\alpha+k,\beta+n-k)}{B(\alpha,\beta)}.

The integral equals the beta-function ratio because the powers of $\theta$ and $1-\theta$ in the integrand are those of $B(\alpha+k,\beta+n-k)$. This formula gives the probability of the count, not of one particular ordering of the outcomes.

Worked Example: Probability of a Claim Count Under a Beta Prior

For a hypothetical equipment coverage class, suppose each risk has a claim probability $\theta\sim\operatorname{Beta}(3,5)$. For one risk, consider four independent policy periods conditional on $\theta$. Calculate the unconditional probability of exactly one claim.

Method. Use the Beta-Binomial formula, including the number of possible positions for the single claim.

Here $\alpha=3$, $\beta=5$, $n=4$, and $k=1$. Hence,

P(K=1)=\binom{4}{1}\frac{B(4,8)}{B(3,5)}.

For positive integers, $B(a,b)=\frac{(a-1)!(b-1)!}{(a+b-1)!}$. Thus,

B(4,8)=\frac{3!\,7!}{11!}=\frac{1}{1320}, \qquad B(3,5)=\frac{2!\,4!}{7!}=\frac{1}{105}.

The ratio is $(1/1320)/(1/105)=105/1320=7/88$. Since $\binom{4}{1}=4$,

P(K=1)=4\left(\frac{7}{88}\right)=\frac{7}{22}\approx0.3182.

The result is a probability, so it lies between zero and one. A common error is to omit the factor of four and report $7/88$; that is the probability of one specified ordering, such as claim followed by three no-claims, not the probability of exactly one claim in any of the four periods. The calculation is checked by distinguishing those two events.

What the Model Assumes—and When It Does Not Apply

The Beta-Binomial model is compact, but its interpretation depends on its assumptions. Conditional on a risk’s $\theta$, the trials are independent and have the same probability of a claim. The prior distribution describes variation in $\theta$ across risks in the modeled population. The number of opportunities $n$ must be meaningful and fixed for the count model being used.

Binary outcomes: Each trial records either a claim outcome or no-claim outcome. If a trial can generate multiple claims and the count itself is the response, a binomial likelihood is not the right likelihood merely because the data are called “claims.”
Comparable trials: The same $\theta$ applies to the risk across the trials in the model. If exposure length or conditions vary so that the trial probabilities change, the equal-probability assumption needs adjustment or a different model.
Conditional independence: Once $\theta$ is fixed, outcomes are assumed independent. Persistent dependence over time that is not represented by $\theta$ can make the likelihood inappropriate.
Population interpretation: The Beta prior represents a distribution of risk-specific probabilities, not one common claim probability for every risk. A fixed-$\theta$ binomial model and a Beta-Binomial population model answer different probability questions.

Worked Example: Same Prior Mean, Different Prior Concentration

Two hypothetical underwriting segments use different prior distributions for a risk’s claim probability. Segment A uses Beta(2, 3), and Segment B uses Beta(4, 6). For a risk in either segment, three comparable periods produce one claim and two no-claims. Compare the prior means and variances, then state the posterior distribution for each segment.

Method. Calculate the moments from each pair of shape parameters, then add the observed claim and no-claim counts to the corresponding posterior parameters. The comparison avoids assuming that equal prior means imply equal prior distributions.

For Segment A, the prior mean is $2/(2+3)=2/5=0.4$, and the variance is

\frac{(2)(3)}{(5)^2(6)}=\frac{6}{150}=0.04.

For Segment B, the prior mean is $4/(4+6)=4/10=0.4$, and the variance is

\frac{(4)(6)}{(10)^2(11)}=\frac{24}{1100}\approx0.02182.

Both priors have mean $0.4$, but Segment B’s prior is more concentrated, as shown by its smaller variance. There is one claim and two no-claims, so the updated distributions are

\text{Segment A: }\operatorname{Beta}(2+1,3+2)=\operatorname{Beta}(3,5), \qquad \text{Segment B: }\operatorname{Beta}(4+1,6+2)=\operatorname{Beta}(5,8).

The posterior parameter sums are $3+5=8$ and $5+8=13$, respectively, each equal to its prior parameter sum plus the three observed trials. The common mistake is to conclude that the two priors are interchangeable because they share a mean; their variances and their resulting posterior distributions differ.

Common Exam Trap

A frequent error is to treat $\alpha$ as the prior number of claims and $\beta$ as the prior number of no-claims, then subtract one from each when updating. For a Beta($\alpha,\beta$) prior in the stated parameterization, the posterior is Beta($\alpha+k,\beta+n-k$); the counts are added to the parameters. The exponents in the density are $\alpha-1$ and $\beta-1$, but the update uses the shape parameters themselves. Another trap is to use $\binom{n}{k}$ when the question gives one specific sequence, or to omit it when the question asks for the probability of a count.

Key takeaway: The Beta prior represents heterogeneity in risk-specific claim probabilities, while the binomial likelihood describes outcomes conditional on a particular probability. Keep those levels distinct: identify the risk probability, count claims and no-claims, and match the combinatorial factor to whether the event is a count or a specified sequence.

Check Your Understanding

For each problem, identify the relevant claim and no-claim counts before applying the model.

A risk has a Beta(3, 7) prior claim probability and records two claims in five comparable trials. What is the posterior distribution? (Answer: Beta(5, 10).)
A risk has a Beta(1, 4) prior and experiences no claims in three trials. Find the posterior distribution and the marginal probability of this zero-claim count. (Answer: Posterior Beta(1, 7); probability $4/7\approx0.5714$.)
Under a Beta(2, 2) prior, a risk records three claims in four trials. Find the posterior distribution and the marginal probability of exactly three claims. (Answer: Posterior Beta(5, 3); probability $8/35\approx0.2286$.)
For each trial, a device is recorded as having at least one claim or no claim. The device has an unknown claim probability with a Beta(2, 5) prior, and the five-trial sequence is no claim, claim, no claim, no claim, claim. Give the likelihood of this ordered sequence and the posterior distribution. (Answer: Likelihood $\theta^2(1-\theta)^3$; posterior Beta(4, 8).)
Two prior distributions are Beta(2, 4) and Beta(5, 10). Compare their means and identify which has smaller variance; then update each after one claim in two trials. (Answer: Both means are $1/3$; Beta(5, 10) has smaller variance; the posteriors are Beta(3, 5) and Beta(6, 11), respectively.)

What You'll Learn