Credibility Theory: Full Bayesian and Empirical Approaches
Compare full Bayesian and empirical Bayesian credibility theory approaches for Exam MAS-I.
The Credibility Problem
Credibility theory addresses a fundamental actuarial question: how should you combine an individual risk's experience with the collective experience of a larger group? If a policyholder has had no claims for three years, should their premium be based entirely on their personal history, entirely on the class average, or some blend of the two? The answer depends on how much data you have, how variable the data is, and how different risks really are from one another.
The credibility premium takes the general form: Premium = Z * (individual experience) + (1 - Z) * (group average), where Z is the credibility factor between 0 and 1. A larger Z means more weight on individual experience. The key difference between full Bayesian and empirical Bayes approaches lies in how Z, the individual experience, and the group average are determined.
Full Bayesian Credibility
In the full Bayesian approach, you specify a prior distribution for the risk parameter and update it using Bayes' theorem after observing data. The posterior distribution combines prior beliefs with observed evidence. The Bayesian premium is the posterior mean of the risk parameter, which naturally takes the credibility-weighted form.
The most important example for Exam MAS-I is the Poisson-Gamma model. Claim counts X_1, X_2, ..., X_n are conditionally independent Poisson(Lambda) given a risk parameter Lambda, and Lambda itself follows a Gamma(alpha, beta) prior distribution. The prior mean claim rate is alpha/beta and the prior variance is alpha/beta^2.
The Poisson-Gamma Model in Detail
Given n observations x_1, x_2, ..., x_n from Poisson(Lambda), the likelihood is proportional to Lambda^{sum(x_i)} * exp(-n*Lambda). Multiplying by the Gamma(alpha, beta) prior density, which is proportional to Lambda^{alpha-1} * exp(-beta*Lambda), the posterior is proportional to Lambda^{alpha + sum(x_i) - 1} * exp(-(beta + n)*Lambda). This is a Gamma(alpha + sum(x_i), beta + n) distribution.
The posterior mean is E[Lambda | data] = (alpha + sum(x_i)) / (beta + n). This can be rewritten as (n/(n+beta)) * (sum(x_i)/n) + (beta/(n+beta)) * (alpha/beta) = Z * x_bar + (1-Z) * mu_0, where Z = n/(n+beta) is the credibility factor, x_bar is the observed mean, and mu_0 = alpha/beta is the prior mean. As n increases, Z approaches 1 and the premium converges to the individual experience.
Other Conjugate Models
The Normal-Normal model is another conjugate pair. If X | mu ~ Normal(mu, sigma^2) and mu ~ Normal(mu_0, tau^2), the posterior of mu given n observations is Normal with mean Z*x_bar + (1-Z)*mu_0 and variance Z*sigma^2/n, where Z = (tau^2/n) / (tau^2/n + sigma^2/n) = tau^2 / (tau^2 + sigma^2/n). Here the credibility weight depends on the relative magnitude of prior uncertainty (tau^2) and sampling uncertainty (sigma^2/n).
The Binomial-Beta model works similarly. If X | p ~ Binomial(m, p) and p ~ Beta(a, b), the posterior is Beta(a + x, b + m - x). The posterior mean is (a + x) / (a + b + m), which is a credibility-weighted average of the sample proportion x/m and the prior mean a/(a+b). These conjugate models demonstrate a general pattern: Bayesian updating with conjugate priors always yields a credibility-weighted result.
Empirical Bayes: Nonparametric (Buhlmann)
The empirical Bayes approach estimates the prior distribution parameters from the collective data rather than specifying them subjectively. The Buhlmann method, the nonparametric empirical Bayes approach, does not assume any specific parametric form for the prior. Instead, it uses the within-group and between-group variation to estimate the credibility factor directly.
The Buhlmann credibility premium is Z * x_bar_i + (1 - Z) * mu, where Z = n / (n + k) and k = v / a. Here v = E[Var(X | theta)] is the expected process variance (average variability within a single risk) and a = Var(E[X | theta]) is the variance of the hypothetical means (variability between risks). These quantities are estimated from the collective experience of all risk groups. The key advantage is that no parametric prior is assumed; the method works regardless of the true prior distribution.
Empirical Bayes: Semiparametric
The semiparametric empirical Bayes approach assumes a parametric family for the prior distribution (such as Gamma) but estimates its parameters from the data. For the Poisson-Gamma model, this means estimating alpha and beta from the collective experience using method of moments or maximum likelihood, then computing the posterior as if these were the true prior parameters.
For example, if you observe claim counts for 50 risk groups over 5 years each, you can estimate alpha and beta by matching the overall mean and variance of the group means to the theoretical values. The overall mean equals alpha/beta and the overall variance of group means equals alpha/beta^2 + (alpha/beta)/n, where n = 5. Solving these two equations gives estimates of alpha and beta, which are then used to compute credibility premiums for each group.
Connecting Bayesian and Buhlmann
When the Bayesian model is correct, the Buhlmann credibility premium coincides with the Bayesian posterior mean. In the Poisson-Gamma model, v = E[Var(X|Lambda)] = E[Lambda] = alpha/beta and a = Var(E[X|Lambda]) = Var(Lambda) = alpha/beta^2. Therefore k = v/a = beta, and Z = n/(n+beta), exactly matching the Bayesian result. This shows that Buhlmann credibility is optimal (in the least-squares sense) when the Bayesian model holds, and provides a reasonable approximation even when the true prior is not in the assumed family.
Common Pitfalls
A frequent mistake is confusing v and a. Remember: v is the expected within-risk variance (process variance), and a is the between-risk variance (variance of hypothetical means). The credibility parameter k = v/a, not a/v. When risks are very heterogeneous (large a), k is small and Z is close to 1, giving high credibility to individual experience. When within-risk variability dominates (large v), k is large and Z is close to 0.
Another error is forgetting to check that the estimated a is positive. If the between-group variation is smaller than expected given the within-group variation, the estimated a can be negative. By convention, set a = 0 in this case, which gives Z = 0 for all risks (no credibility to individual experience).
Study Tips
Exam MAS-I tests both the full Bayesian and empirical Bayes frameworks and their connections. Practice computing posterior distributions for the Poisson-Gamma and Normal-Normal models. Be able to derive the credibility weight Z from a Bayesian model and verify that it matches the Buhlmann formula. For empirical Bayes problems, practice estimating v and a from grouped data and computing credibility premiums. Know the conceptual distinction between full Bayesian (prior specified) and empirical Bayes (prior estimated from data), and be prepared to explain the advantages of each approach.