Buhlmann-Straub Model for Heterogeneous Portfolios
Learn the Buhlmann-Straub credibility model for varying exposure on Exam STAM.
Why the Buhlmann-Straub Extension Is Needed
The standard Buhlmann model assumes that all observations for a given risk have the same variance. In practice, this is rarely the case. An insurance portfolio may have 100 policies in year 1, 150 in year 2, and 200 in year 3. The average loss per policy in each year has different precision because it is based on different volumes of exposure. The Buhlmann-Straub model accounts for this by incorporating exposure weights into the credibility calculation.
The model is essential for actuarial applications because exposure typically varies across observation periods. Premium rate changes, policy count growth, and portfolio restructuring all cause the volume of business to change over time. A credibility model that ignores these differences gives equal weight to a year with 50 policies and a year with 500 policies, which is statistically inefficient.
The Buhlmann-Straub Model
Instead of observing individual claims X_1, ..., X_n, we observe averages X_j = (total claims in period j) / m_j, where m_j is the exposure (number of policies, earned premium, or other volume measure) in period j. The model assumes E[X_j|Theta] = mu(Theta) for all j, and Var(X_j|Theta) = v(Theta)/m_j. The variance of each observation is inversely proportional to its exposure, reflecting the fact that averages based on more data are more precise.
The Buhlmann-Straub premium is: P_BS = Z * X_w + (1 - Z) * mu, where X_w = sum(m_j * X_j) / sum(m_j) is the exposure-weighted average and Z = m/(m + k). Here m = sum(m_j) is the total exposure across all observation periods and k = v/a, the same ratio as in the standard Buhlmann model.
When all exposures are equal (m_j = 1 for all j), the Buhlmann-Straub model reduces to the standard Buhlmann model: X_w becomes X_bar and m becomes n. This confirms that the Buhlmann-Straub model is a proper generalization.
Worked Example
A policyholder has three years of data. Year 1: 50 policies, total claims $12,000 (X_1 = 240). Year 2: 80 policies, total claims $20,000 (X_2 = 250). Year 3: 120 policies, total claims $24,000 (X_3 = 200). From the portfolio, v = 10,000 and a = 500. Then k = v/a = 20.
Total exposure: m = 50 + 80 + 120 = 250. X_w = (50*240 + 80*250 + 120*200)/250 = (12000 + 20000 + 24000)/250 = 56000/250 = 224. Z = 250/(250 + 20) = 250/270 = 0.926. Prior mean mu = 230 (from the overall portfolio).
P_BS = 0.926 * 224 + 0.074 * 230 = 207.4 + 17.0 = 224.4. The high credibility factor (0.926) reflects the substantial total exposure of 250 policies. The premium is pulled close to the observed weighted average.
Estimating Structural Parameters from Data
With multiple risks (r risks), the structural parameters mu, v, and a can be estimated from the data. For risk i observed over n_i periods with exposure m_{ij} in period j and average loss X_{ij}:
The exposure-weighted mean for risk i is X_{w,i} = sum_j(m_{ij} * X_{ij}) / m_{i.}, where m_{i.} = sum_j(m_{ij}) is the total exposure for risk i. The grand weighted mean is mu_hat = sum_i(m_{i.} * X_{w,i}) / m_{..}, where m_{..} = sum_i(m_{i.}).
The estimated process variance is v_hat = sum_i(sum_j(m_{ij} * (X_{ij} - X_{w,i})^2)) / sum_i(n_i - 1). This measures the within-risk variability, weighted by exposure. Each risk contributes sum_j(m_{ij}*(X_{ij} - X_{w,i})^2) with n_i - 1 degrees of freedom.
The estimated between-risk variance is a_hat = [sum_i(m_{i.}*(X_{w,i} - mu_hat)^2) - (r-1)*v_hat] / [m_{..} - sum_i(m_{i.}^2)/m_{..}]. The denominator adjusts for the unequal exposures. If the numerator is negative, set a_hat = 0, which means there is no significant between-risk variability and Z = 0 for all risks.
Multi-Risk Credibility Premiums
Once v_hat and a_hat are estimated, compute k_hat = v_hat/a_hat and the credibility factor for each risk: Z_i = m_{i.}/(m_{i.} + k_hat). The credibility premium for risk i is: P_i = Z_i * X_{w,i} + (1 - Z_i) * mu_hat. Risks with larger total exposure receive higher credibility (larger Z_i), meaning their premiums are weighted more heavily toward their own experience.
Example with two risks: Risk A has m_A = 300, X_{w,A} = 200. Risk B has m_B = 100, X_{w,B} = 350. From the data, v_hat = 5000, a_hat = 1000, so k_hat = 5. mu_hat = (300*200 + 100*350)/400 = 147500/400 = 237.5. Z_A = 300/(300+5) = 0.984. Z_B = 100/(100+5) = 0.952. Premium_A = 0.984*200 + 0.016*237.5 = 200.6. Premium_B = 0.952*350 + 0.048*237.5 = 344.6.
Risk B has lower credibility than Risk A because of its smaller exposure, so its premium is pulled slightly more toward the grand mean.
Comparison with Standard Buhlmann
The key differences between the Buhlmann-Straub model and the standard Buhlmann model are the use of exposure-weighted averages instead of simple averages, total exposure m instead of sample size n in the credibility factor, and the variance assumption Var(X_j|Theta) = v(Theta)/m_j instead of constant variance. The Buhlmann-Straub model is always preferred when exposures vary across observation periods.
Both models produce credibility premiums of the form Z*data + (1-Z)*prior. The structural parameters v and a have the same interpretation in both models: v measures within-risk variability and a measures between-risk variability. The difference is entirely in how the data is weighted and how Z is computed.
Study Tips and Common Pitfalls
The most common mistake is using simple (unweighted) averages instead of exposure-weighted averages. Always compute X_w = sum(m_j*X_j)/sum(m_j) rather than X_bar = sum(X_j)/n. This is the defining feature of the Buhlmann-Straub model and failing to apply it is the most frequent source of errors.
Another pitfall is incorrectly computing the denominator in the a_hat formula. The term m_{..} - sum(m_{i.}^2)/m_{..} corrects for unequal exposures and is analogous to the degrees-of-freedom adjustment in ANOVA. On Exam STAM, the a_hat formula is typically provided, but you must apply it correctly to the given data.
Practice with tabular data presentations (risks in rows, periods in columns, with varying exposures) to build speed and accuracy. At least 4 problems involving parameter estimation from multi-risk data and credibility premium computation will prepare you for the exam. Organize your work carefully, as these problems involve many intermediate calculations where a single arithmetic error can cascade.