Bayesian Estimation and Credibility Theory Fundamentals
Connect Bayesian estimation to credibility theory for Exam STAM.
The Bayesian Framework
In the Bayesian approach, the unknown parameter Theta is treated as a random variable with a prior distribution pi(theta) that reflects prior beliefs or knowledge before observing data. Given Theta = theta, the data X_1, ..., X_n are conditionally independent with density f(x|theta). After observing the data, Bayes' theorem updates the prior to the posterior distribution: pi(theta|x) is proportional to L(x|theta) * pi(theta), where L(x|theta) = product of f(x_i|theta) is the likelihood.
The posterior distribution combines prior knowledge with the information in the data. With more data, the posterior concentrates around the true parameter value, and the influence of the prior diminishes. With little data, the prior has more influence on the posterior.
The Bayesian estimate of the next period's expected claims is the posterior mean: E[X_{n+1}|X_1, ..., X_n] = integral of mu(theta) * pi(theta|x) d(theta), where mu(theta) = E[X|Theta = theta] is the conditional mean given the parameter. This estimate minimizes the expected squared error loss.
Conjugate Prior Models
A prior distribution is conjugate for a given likelihood if the posterior has the same distributional form as the prior, with updated parameters. Conjugate priors make Bayesian calculations tractable because the posterior can be computed in closed form without numerical integration.
Poisson-Gamma model: X|Lambda ~ Poisson(Lambda), Lambda ~ Gamma(alpha, beta). The posterior is Lambda|X ~ Gamma(alpha + sum(x_i), beta/(n*beta + 1)). The posterior mean is (alpha + sum(x_i))/(1/beta + n). This is a weighted average of the prior mean alpha*beta and the sample mean x_bar, with weights proportional to 1/beta and n respectively.
Normal-Normal model: X|Theta ~ Normal(Theta, sigma^2) with sigma^2 known, Theta ~ Normal(mu_0, tau^2). The posterior is Theta|X ~ Normal(posterior mean, posterior variance), where the posterior mean = (mu_0/tau^2 + n*x_bar/sigma^2)/(1/tau^2 + n/sigma^2). Again, this is a weighted average of the prior mean and the sample mean.
Bernoulli-Beta model: X|Q ~ Bernoulli(Q), Q ~ Beta(a, b). The posterior is Q|X ~ Beta(a + sum(x_i), b + n - sum(x_i)). The posterior mean is (a + sum(x_i))/(a + b + n), a weighted average of the prior mean a/(a+b) and the sample proportion.
Connection to Credibility
In all three conjugate models above, the Bayesian estimate of the mean has the form: posterior mean = Z * X_bar + (1 - Z) * mu_0, where Z is the credibility factor (0 <= Z <= 1), X_bar is the sample mean, and mu_0 is the prior mean. This linear credibility formula is exact for these conjugate pairs, not an approximation.
The credibility factor Z represents the weight given to the observed data versus the prior. It increases with sample size n: more data means more weight on the observed experience and less weight on the prior. In the limit as n approaches infinity, Z approaches 1, and the Bayesian estimate converges to the sample mean regardless of the prior.
This connection between Bayesian estimation and credibility theory is fundamental to actuarial science. Credibility theory was developed independently of Bayesian statistics, but the two approaches converge for conjugate models, providing a rigorous theoretical foundation for the credibility formulas used in practice.
Key Structural Quantities
Four quantities define the credibility structure. The hypothetical mean is mu(theta) = E[X|Theta = theta], the expected claim cost for a risk with parameter theta. The process variance is v(theta) = Var(X|Theta = theta), the variability of claims within a single risk class. The expected value of the process variance is v = E[v(Theta)], the average within-risk variability. The variance of the hypothetical means is a = Var(mu(Theta)), measuring how much the expected claim costs differ across risk classes.
The credibility factor is Z = n/(n + k), where k = v/a. The ratio k = v/a compares within-risk variability to between-risk variability. When a is large (risk classes differ substantially), k is small and Z is high even for small n, because even a few observations provide useful information about which risk class the individual belongs to. When v is large (high variability within each risk class), k is large and Z is low, because individual observations are noisy and less informative.
For the Poisson-Gamma model: mu(theta) = theta, v(theta) = theta (since Poisson variance equals the mean), v = E[Theta] = alpha*beta, a = Var(Theta) = alpha*beta^2. Thus k = v/a = 1/beta and Z = n/(n + 1/beta) = n*beta/(n*beta + 1).
Computing the Bayesian Premium
The Bayesian premium (the estimate for the next period's expected claims) can be computed directly from the posterior mean or using the credibility formula. Both approaches give the same answer for conjugate models. The credibility formula is often faster: Premium = Z * X_bar + (1 - Z) * mu_0.
Example: Claim counts follow a Poisson(Lambda) distribution, with Lambda ~ Gamma(4, 0.5). A policyholder has n = 3 years of experience with a total of 8 claims. The prior mean is mu_0 = 4*0.5 = 2.0. k = 1/beta = 1/0.5 = 2.0. Z = 3/(3 + 2) = 0.60. X_bar = 8/3 = 2.667. Premium = 0.60*2.667 + 0.40*2.0 = 1.600 + 0.800 = 2.40.
Verification using the posterior directly: posterior is Gamma(4 + 8, 0.5/(3*0.5 + 1)) = Gamma(12, 0.2). Posterior mean = 12*0.2 = 2.40, confirming the credibility calculation.
Non-Conjugate Models and Numerical Methods
When the prior is not conjugate, the posterior does not have a closed form, and the Bayesian estimate requires numerical integration. On Exam STAM, non-conjugate problems are typically structured so that the integration is manageable (e.g., discrete priors with a small number of values for Theta, or simple continuous priors where the integration can be done analytically).
With a discrete prior (Theta takes values theta_1, ..., theta_m with probabilities p_1, ..., p_m), the posterior is computed by Bayes' theorem: P(Theta = theta_j | X) = L(X|theta_j)*p_j / sum_k(L(X|theta_k)*p_k). The Bayesian premium is then sum of mu(theta_j) * P(Theta = theta_j | X).
Study Tips and Common Pitfalls
The most common mistake is computing the posterior parameters incorrectly for conjugate models. Each conjugate pair has specific update rules. For Poisson-Gamma, the shape parameter increases by the total number of claims and the rate parameter is updated using the sample size. Memorize the update rules for each conjugate pair tested on Exam STAM: Poisson-Gamma, Normal-Normal, and Bernoulli-Beta.
Another frequent error is confusing the prior mean with the process mean. The prior mean E[Theta] represents the average parameter value across the population. The hypothetical mean mu(theta) = E[X|Theta = theta] is the expected claim given a specific parameter value. These may differ (e.g., for the Poisson, mu(theta) = theta, so E[mu(Theta)] = E[Theta]). Practice at least 5 problems computing Bayesian premiums using both the direct posterior approach and the credibility formula to build confidence.