Empirical Bayes Methods in Actuarial Credibility
Understand the Buhlmann and Buhlmann-Straub empirical Bayes credibility models for actuarial exams.
Buhlmann Credibility Model
The Buhlmann model provides a least-squares credibility framework. For a risk with parameter theta, the credibility premium is Z times the observed mean plus (1-Z) times the overall mean. The credibility factor Z = n/(n+k) where k = v/a. Here v is the expected value of the process variance (within-risk variance) and a is the variance of the hypothetical means (between-risk variance). The key insight is that as data grows, Z approaches 1 and the estimate converges to the individual risk\'s experience.
The Buhlmann model is sometimes called the "greatest accuracy" credibility approach because it minimizes the expected squared error of the premium estimate. Unlike limited fluctuation credibility, which uses an ad hoc square root formula, Buhlmann credibility arises from a formal optimization problem. This makes it theoretically appealing and widely used in practice.
Understanding v and a
The parameter k = v/a is the ratio of two variance components, and understanding what v and a represent is essential. The expected value of the process variance, v = E[Var(X|theta)], measures the average variability within a single risk class. Think of it as the randomness you would still see even if you knew exactly which risk class an insured belongs to. The variance of the hypothetical means, a = Var(E[X|theta]), measures how much the true mean varies across different risk classes. It captures the heterogeneity in the population.
When a is large relative to v (meaning risks are very different from each other), k is small and Z is close to 1, so individual experience gets heavy weight. When v is large relative to a (meaning most variation is random noise within each class), k is large and Z is close to 0, so the overall mean dominates. This intuition helps you check whether your calculated Z makes sense.
Worked Example
Suppose you have three risk classes observed over four years. Risk A has annual losses of 100, 120, 110, 130. Risk B has 200, 180, 210, 190. Risk C has 150, 140, 160, 155. To find the Buhlmann credibility premium for Risk A, you would first compute the overall mean mu = (mean_A + mean_B + mean_C)/3. Here mean_A = 115, mean_B = 195, mean_C = 151.25, so mu = (115 + 195 + 151.25)/3 = 153.75.
Next, estimate v by computing the sample variance within each risk and averaging: v_hat = (Var_A + Var_B + Var_C)/3. Then estimate a using a = [sum of (mean_i - overall_mean)^2 / (r-1)] - v/(n), where r is the number of risks and n is the number of observations per risk. With n = 4 observations per risk, the credibility factor is Z = 4/(4 + k) where k = v_hat/a_hat. The credibility premium for Risk A is then Z * 115 + (1-Z) * 153.75.
Buhlmann-Straub Extension
The Buhlmann-Straub model extends Buhlmann by allowing varying exposure weights across observations. Each observation has weight w_j, and the credibility factor becomes Z = W/(W+k) where W is total weight. This is important in practice because different years or different risks may have different exposure levels. A risk with 10,000 policies deserves more credibility than one with 100 policies, even if both are observed for the same number of years.
In the Buhlmann-Straub framework, the weighted average for risk i is X_bar_i = (sum of w_ij * X_ij) / W_i, where W_i = sum of w_ij. The estimation formulas for v and a become more complex because they must account for unequal weights. The within-risk estimate of v uses v_hat = sum over all risks of sum_j w_ij * (X_ij - X_bar_i)^2, divided by (total observations minus number of risks). The between-risk estimate requires an adjustment factor involving the sum of weights and sum of squared weights.
Estimation Formulas for v and a
For the Buhlmann-Straub model, the unbiased estimator of v is v_hat = (1/(n_total - r)) * sum_i sum_j w_ij * (X_ij - X_bar_i)^2, where n_total is the total number of observations and r is the number of risk classes. The estimator of a uses the between-class variation: a_hat = [sum_i W_i * (X_bar_i - X_bar)^2 - (r-1)*v_hat] / [W - sum_i(W_i^2)/W], where W is total weight across all risks and X_bar is the overall weighted mean.
A negative estimate of a can occur when within-class variability overwhelms between-class differences. In that case, convention sets a_hat = 0, which means k is infinite and Z = 0 for all risks. This makes intuitive sense: if you cannot detect meaningful differences between risk classes, you should not adjust premiums based on individual experience.
Common Pitfalls
One frequent mistake is confusing the Buhlmann model (equal weights) with Buhlmann-Straub (unequal weights). On Exam STAM, pay close attention to whether the problem provides exposure or weight information. If all observations have equal weight, either model gives the same answer, but the formulas simplify under Buhlmann.
Another pitfall is computing a incorrectly. The between-risk sum of squares must be adjusted by subtracting a term involving v to obtain an unbiased estimate. Simply computing the variance of the risk means does not give the correct a_hat. Also remember that k = v/a, not a/v. Mixing up the ratio flips the credibility in the wrong direction.
Study Tips
Exam STAM problems frequently ask you to compute the credibility premium using sample estimates of v and a from a table of risk class data. Practice the full workflow: compute within-risk and between-risk sums of squares, estimate v and a, compute k and Z, and produce the final credibility premium. Organize your calculations in a table to avoid arithmetic errors. Make sure you can handle both the equal-weight Buhlmann case and the Buhlmann-Straub case with varying exposures. Time yourself on these problems because the computation can be lengthy if you are not efficient.