Buhlmann Credibility: Theory and Exam STAM Calculations
Master Buhlmann credibility theory and its application on Exam STAM.
The Buhlmann Model
Buhlmann credibility finds the best linear approximation to the Bayesian estimate without requiring knowledge of the full posterior distribution. Given n observations X_1, ..., X_n from a risk with parameter Theta, the Buhlmann premium is: P_B = Z * X_bar + (1 - Z) * mu, where mu = E[mu(Theta)] is the overall expected claim, X_bar is the sample mean of the observed data, and Z = n/(n + k) is the credibility factor with k = v/a.
Here v = E[Var(X|Theta)] is the expected process variance (average within-risk variability) and a = Var(E[X|Theta]) is the variance of the hypothetical means (between-risk variability). The ratio k = v/a determines how quickly credibility builds with additional observations.
The Buhlmann premium minimizes E[(X_{n+1} - estimate)^2] over all linear functions of X_1, ..., X_n. This means it is the best linear predictor, even when the exact Bayesian estimate is non-linear. For conjugate models (Poisson-Gamma, Normal-Normal, Bernoulli-Beta), the Buhlmann premium equals the exact Bayesian premium.
Step-by-Step Computation
Step 1: Identify the conditional distribution X|Theta and the prior distribution of Theta. Compute mu(theta) = E[X|Theta = theta] and v(theta) = Var(X|Theta = theta) as functions of theta.
Step 2: Compute the structural parameters. The overall mean: mu = E[mu(Theta)]. The expected process variance: v = E[v(Theta)]. The variance of hypothetical means: a = Var(mu(Theta)) = E[mu(Theta)^2] - mu^2. For discrete Theta with known probabilities, these expectations are weighted sums. For continuous Theta, they are integrals.
Step 3: Compute k = v/a. This is the key ratio that determines the speed of credibility buildup.
Step 4: Compute Z = n/(n + k). This is the credibility factor, the weight given to the observed data.
Step 5: Compute the Buhlmann premium P_B = Z * X_bar + (1 - Z) * mu.
Worked Example: Poisson-Gamma
Claims follow a Poisson(Lambda) distribution, and Lambda ~ Gamma(3, 2). A policyholder has 4 years of data with claim counts: 1, 3, 2, 4 (total = 10, X_bar = 2.5).
Step 1: mu(lambda) = E[X|Lambda = lambda] = lambda. v(lambda) = Var(X|Lambda = lambda) = lambda (Poisson variance equals mean).
Step 2: mu = E[Lambda] = 3*2 = 6. v = E[Lambda] = 6. a = Var(Lambda) = 3*2^2 = 12.
Step 3: k = v/a = 6/12 = 0.5.
Step 4: Z = 4/(4 + 0.5) = 4/4.5 = 0.889.
Step 5: P_B = 0.889 * 2.5 + 0.111 * 6 = 2.222 + 0.667 = 2.889.
The credibility premium of 2.889 is much closer to the observed mean (2.5) than the prior mean (6), reflecting the high credibility (Z = 0.889) from four years of data with small k.
Worked Example: Discrete Prior
Claims are exponential with mean Theta, and Theta takes values 100 and 200 with equal probability. Two claims are observed: x_1 = 50 and x_2 = 150 (X_bar = 100).
Step 1: mu(theta) = theta. v(theta) = theta^2 (exponential variance equals mean squared).
Step 2: mu = 0.5*100 + 0.5*200 = 150. v = 0.5*100^2 + 0.5*200^2 = 0.5*10000 + 0.5*40000 = 25,000. a = 0.5*100^2 + 0.5*200^2 - 150^2 = 25,000 - 22,500 = 2,500.
Step 3: k = 25,000/2,500 = 10.
Step 4: Z = 2/(2 + 10) = 2/12 = 0.167.
Step 5: P_B = 0.167 * 100 + 0.833 * 150 = 16.7 + 125.0 = 141.7.
The low credibility (Z = 0.167) reflects the high process variance relative to the between-risk variance. Two observations provide limited information when individual claims are so variable.
Full Credibility and Required Sample Size
Full credibility is the sample size n_0 at which Z reaches a specified level, typically Z = 0.95 or Z = 0.99. Solving Z = n/(n + k) for n: n_0 = Z*k/(1 - Z). For Z = 0.95: n_0 = 0.95*k/0.05 = 19k. For Z = 0.99: n_0 = 0.99*k/0.01 = 99k.
Exam STAM may ask how many years of data are needed to achieve a specified credibility level. The answer depends entirely on k = v/a. Risks with large k (high process variance relative to between-risk variance) require more observations to reach full credibility.
In the Poisson-Gamma example above, k = 0.5, so n_0 for 95% credibility is 19*0.5 = 9.5, meaning 10 years of data. In the discrete prior example, k = 10, so n_0 = 19*10 = 190 observations, which is far more demanding.
Relationship to Bayesian and Limited Fluctuation Credibility
Buhlmann credibility bridges the gap between exact Bayesian estimation and the simpler limited fluctuation (classical) credibility approach. Exact Bayesian estimation requires full specification of the prior and likelihood, which may not be available. Limited fluctuation credibility uses a fixed formula based on the number of claims and a full credibility standard, without requiring a model for the prior.
Buhlmann credibility requires only the structural parameters mu, v, and a, which can be estimated from data without specifying the exact distributions. This makes it more practical than full Bayesian estimation while being more theoretically grounded than limited fluctuation credibility.
The Buhlmann premium always has the linear credibility form P_B = Z*X_bar + (1-Z)*mu, regardless of the underlying distributions. When the distributions happen to be conjugate, this linear form is also the exact Bayesian estimate. When they are not conjugate, the Buhlmann premium is the best linear approximation to the Bayesian estimate.
Study Tips and Common Pitfalls
The most common mistake is computing a = Var(mu(Theta)) incorrectly. Remember that a = E[mu(Theta)^2] - (E[mu(Theta)])^2, which is the variance formula applied to the random variable mu(Theta). For a discrete prior, a = sum of mu(theta_j)^2 * p_j - mu^2. Forgetting to square mu in the subtraction or squaring the wrong quantity are frequent errors.
Another pitfall is confusing v(theta) = Var(X|Theta = theta) with Var(X). The process variance v(theta) is conditional on a specific theta value; v = E[v(Theta)] averages this over the prior. Make sure you compute the conditional variance first, then take the expectation over Theta. Practice at least 5 Buhlmann credibility problems with different model setups to ensure fluency with the computation steps.