Binomial Credibility Weights from Beta Priors

Why the Number of Trials Matters

When a policyholder’s probability of a particular event is uncertain, a beta prior can represent that uncertainty. After observing binary outcomes, the posterior estimate should reflect both the prior information and the new experience. The key question is how much weight to assign to each. In the beta-binomial model, the posterior mean has a particularly useful form: it is a weighted average of the prior mean and the observed proportion, with a data weight that increases with the number of trials.

This result is useful in Bayesian credibility because it makes the effect of experience explicit. Two groups can have the same observed proportion but receive different amounts of credibility if they have different numbers of trials. The discussion is aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus.

Model and notation: Let \(\Theta\) be an unknown event probability with prior distribution \(\Theta\sim\operatorname{Beta}(\alpha,\beta)\), where \(\alpha>0\) and \(\beta>0\). Conditional on \(\Theta=\theta\), suppose there are \(n\) independent, identically distributed Bernoulli trials, each with event probability \(\theta\). Let \(K\) be the number of events observed, with \(K=k\), and let \(\widehat{p}=k/n\) be the observed proportion when \(n>0\). Define the prior mean \(m=\alpha/(\alpha+\beta)\) and prior parameter sum \(s=\alpha+\beta\).

The Posterior Mean as a Weighted Average

The likelihood of observing \(k\) events in \(n\) trials is proportional to \(\theta^k(1-\theta)^{n-k}\). Multiplying it by the beta prior density adds \(k\) to the first beta parameter and \(n-k\) to the second. Thus, after observing the data, the posterior distribution is \(\operatorname{Beta}(\alpha+k,\beta+n-k)\). Its mean is the posterior estimate of the event probability.

E[\Theta\mid K=k] =\frac{\alpha+k}{\alpha+\beta+n}

To see the credibility form, write \(\alpha=sm\) and, when \(n>0\), write \(k=n\widehat{p}\). Then split the numerator into its prior and data parts:

E[\Theta\mid K=k] =\frac{sm+n\widehat{p}}{s+n} =\frac{s}{s+n}m+\frac{n}{s+n}\widehat{p}

Credibility result: For \(n>0\), the posterior mean is \( (1-Z)m+Z\widehat{p}\), where the data credibility weight is \(Z=n/(s+n)\), the prior weight is \(1-Z=s/(s+n)\), and \(s=\alpha+\beta\). The two weights are nonnegative and sum to one.

The formula is a weighted average, not a rule that replaces the prior estimate with the observed proportion. For finite \(n\) and positive \(\alpha,\beta\), both weights are strictly between zero and one. The posterior mean therefore lies between the prior mean and the observed proportion, inclusive. If the observed proportion is above the prior mean, the posterior mean moves upward; if it is below the prior mean, it moves downward.

Why the Weight Has This Form

The derivation follows directly from the beta posterior. The beta prior density is proportional to \(\theta^{\alpha-1}(1-\theta)^{\beta-1}\). Conditional independence gives a binomial likelihood proportional to \(\theta^k(1-\theta)^{n-k}\). Their product is proportional to

\theta^{\alpha+k-1}(1-\theta)^{\beta+n-k-1}

This is the kernel of a \(\operatorname{Beta}(\alpha+k,\beta+n-k)\) density. Its mean is \((\alpha+k)/(\alpha+\beta+n)\). Substituting \(\alpha=sm\) and \(k=n\widehat{p}\) gives the weighted-average expression above. The appearance of \(n\) in the data weight follows because the likelihood contributes \(n\) trial outcomes, while the prior contribution is controlled by the fixed sum \(s\).

In particular, \(Z=n/(s+n)\) increases with \(n\) for a fixed prior. For example, \(Z\) is close to zero when \(n\) is small compared with \(s\), and it approaches one as \(n\) becomes large. This describes the formula’s weighting behavior; it does not guarantee that a particular observed proportion is close to the unknown \(\Theta\).

Worked Example: Updating a Prior With Moderate Experience

A hypothetical insurer models whether a covered device has a specified repair event during a year. For a device, let the event probability have prior distribution \(\Theta\sim\operatorname{Beta}(2,8)\). Suppose the device’s event status is observed for 60 independent years, with 18 event years. Find the posterior mean and express it as a credibility-weighted average. Name the method and identify the mistake the calculation avoids.

Method: beta-binomial conjugate updating followed by the credibility decomposition. The prior parameters are \(\alpha=2\) and \(\beta=8\), so \(s=2+8=10\), and the prior mean is \(m=2/10=0.2\). The number of trials is \(n=60\), and the observed count is \(k=18\), giving \(\widehat{p}=18/60=0.3\).

The data weight is \(Z=60/(10+60)=60/70=6/7\), and the prior weight is \(1-Z=1/7\). The weighted mean is

\left(\frac{1}{7}\right)(0.2)+\left(\frac{6}{7}\right)(0.3) =\frac{0.2}{7}+\frac{1.8}{7} =\frac{2.0}{7} =0.285714\ldots

Check directly from the posterior parameters: \(\alpha+k=2+18=20\), \(\beta+n-k=8+60-18=50\), and the posterior mean is \(20/(20+50)=20/70=2/7\). Both methods agree. Rounded to four decimal places, the posterior mean is \(0.2857\).

The common mistake avoided is giving the observed proportion full weight and reporting \(0.3\) as the posterior mean. The 60 trials do give the data substantial credibility, but the prior still receives weight \(1/7\).

Interpreting the Prior Parameter Sum

For a fixed prior mean \(m\), the parameter sum \(s=\alpha+\beta\) determines how quickly the data weight grows. A larger \(s\) gives the prior more weight at every fixed \(n\), because \(Z=n/(s+n)\) is smaller. A smaller \(s\) gives the data a larger weight for the same number of trials. The prior mean and prior parameter sum have distinct roles: the mean sets the prior’s central estimate, while the sum controls the weight assigned to that estimate relative to observed experience.

As a basic check, \(0<Z<1\) whenever \(n>0\) and \(s>0\), and the posterior mean must lie between \(m\) and \(\widehat{p}\). If it does not, recheck the parameters, count, and weights. The prior weight is \(s/(s+n)\), not \(n/(s+n)\); reversing them makes the posterior move in the wrong direction when the data and prior disagree.

Worked Example: Small Experience With No Events

For a hypothetical class of electronic components, let the probability of a specified fault be \(\Theta\sim\operatorname{Beta}(4,6)\). Five independent components are tested under the same conditions, and none has the fault. Calculate the posterior mean using credibility weights. Explain why a zero observed proportion does not make the posterior mean zero.

Method: compute the prior and data weights, then average their estimates. The prior parameter sum is \(s=4+6=10\), and the prior mean is \(m=4/10=0.4\). Here \(n=5\), \(k=0\), and \(\widehat{p}=0/5=0\).

The data weight is \(Z=5/(10+5)=1/3\). The prior weight is \(1-Z=2/3\). Therefore,

\left(\frac{2}{3}\right)(0.4)+\left(\frac{1}{3}\right)(0) =\frac{0.8}{3} =\frac{4}{15} =0.266666\ldots

Check by updating the beta parameters: the posterior is \(\operatorname{Beta}(4+0,6+5)=\operatorname{Beta}(4,11)\), whose mean is \(4/(4+11)=4/15\). Rounded to four decimal places, the posterior mean is \(0.2667\).

The common mistake avoided is treating a sample with no observed faults as proof that the underlying event probability is zero. With only five trials, the data weight is \(1/3\), so the prior estimate still receives weight \(2/3\).

Worked Example: Meeting a Target Data Weight

A hypothetical risk analyst uses \(\Theta\sim\operatorname{Beta}(3,7)\) as the prior probability of a binary claim outcome. How many trials are needed for the data credibility weight to be at least \(0.9\)? At that minimum number of trials, suppose the observed proportion is \(0.3\). Find the posterior mean and verify it from the posterior parameters.

Method: solve the credibility-weight inequality, then update. The prior sum is \(s=3+7=10\), so the weight requirement is

\frac{n}{n+10}\geq 0.9 \quad\Longrightarrow\quad n\geq 0.9n+9 \quad\Longrightarrow\quad 0.1n\geq 9 \quad\Longrightarrow\quad n\geq 90.

Thus the minimum integer number of trials is \(n=90\). An observed proportion of \(0.3\) means \(k=90(0.3)=27\) events. The prior mean is \(m=3/10=0.3\), and the data weight at 90 trials is \(90/(90+10)=0.9\). The posterior mean is

$$ (1-0.9)(0.3)+(0.9)(0.3)=0.3. $$

For the direct check, the posterior is \(\operatorname{Beta}(3+27,7+90-27)=\operatorname{Beta}(30,70)\), with mean \(30/(30+70)=0.3\). The agreement is expected because the observed proportion equals the prior mean; changing the weights cannot change an average of two equal values. The common mistake avoided is using \(n/s\) as the credibility weight. The correct weight is \(n/(n+s)\), which is always less than one for finite \(n\).

Conditions, Limits, and Exam Traps

The weighted-average result depends on the stated model. The trials must have a common event probability conditional on \(\Theta\), and they must be conditionally independent given \(\Theta\). The number of trials \(n\) must be known, and the data must be represented by the event count \(k\) out of those \(n\) trials. If trial probabilities differ, outcomes are dependent beyond the shared \(\Theta\), or observations are not a complete set of binary trials, this derivation does not automatically apply.

When \(n=0\), the observed proportion \(k/n\) is undefined, so the weighted-average expression involving \(\widehat{p}\) should not be used. There is no trial information to update the prior: the posterior remains \(\operatorname{Beta}(\alpha,\beta)\), and its mean remains \(m\). The direct posterior formula gives this result because \((\alpha+0)/(\alpha+\beta+0)=m\).

As \(n\) grows, the data weight tends to one. For a fixed observed proportion, the posterior mean consequently moves toward that proportion. This is a statement about the formula, not a claim that the finite-sample observed proportion is certain to equal the true event probability. In a real application, uncertainty remains about \(\Theta\), even when the posterior mean is reported as a single estimate.

Key takeaway: With a beta prior and binomial observations, the posterior mean is a weighted average of the prior mean and the observed proportion. The data weight is \(n/(n+\alpha+\beta)\): it depends on the number of trials and the prior parameter sum, not on the observed count \(k\).

Check Your Understanding

For each question, identify the prior mean, data weight, and posterior mean where requested. Answers are shown at the end of each problem.

A hypothetical coverage has \(\Theta\sim\operatorname{Beta}(1,4)\). In 15 independent trials, 6 events occur. Find the prior mean, data weight, and posterior mean. (Answer: prior mean \(0.2000\); data weight \(0.7500\); posterior mean \(0.3500\).)
For a binary outcome, the prior is \(\operatorname{Beta}(5,5)\), and 1 event is observed in 10 trials. Find the posterior mean both by weighting and by updating the beta parameters. (Answer: \(0.3000\); posterior \(\operatorname{Beta}(6,14)\).)
A hypothetical risk has prior \(\operatorname{Beta}(2,8)\). There are 40 trials and 16 observed events. Calculate the credibility-weighted posterior mean and check it directly. (Answer: prior mean \(0.2000\), data weight \(0.8000\), posterior mean \(0.3600\); posterior \(\operatorname{Beta}(18,32)\).)
A beta prior has parameter sum 12. What is the minimum integer number of trials required for the data weight to be at least \(0.8\)? (Answer: \(n=48\), since \(n/(n+12)\geq0.8\) implies \(n\geq48\).)
A prior is \(\operatorname{Beta}(3,9)\). After 24 trials, the observed proportion is \(0.5\). Calculate the posterior mean and state how it compares with the observed proportion. (Answer: prior mean \(0.2500\), data weight \(24/36=0.6667\), posterior mean \(0.4167\), below the observed proportion.)

What You'll Learn