Why does Buhlmann-Straub use exposure weights instead of treating each observation equally?

Different observations can have different precision. A rate based on more exposure has lower process variance, so it deserves more weight in the risk's own experience average.

Loss costs are 110, 95, and 120 with exposures 10, 20, and 30. Find the exposure-weighted mean.

The weighted mean is (10(110) + 20(95) + 30(120)) / 60 = 6600 / 60 = 110. Steps: Multiply each rate by its exposure. Sum the weighted amounts. Divide by total exposure.

Using weighted mean 110, collective mean 100, EPV = 400, VHM = 25, and total exposure 60, find the Buhlmann-Straub credibility estimate.

K = EPV / VHM = 16. Z = 60 / (60 + 16) = 0.789. The estimate is 0.789(110) + 0.211(100) = 107.9. Steps: Compute K = 400 / 25. Compute Z = w / (w + K). Blend the weighted mean and collective mean.

In reserving, what two estimates might a credibility method blend for an immature accident year?

It can blend a triangle-driven estimate, such as chain-ladder, with an external or prior estimate, such as an expected loss ratio estimate. The immature year often gives less credibility to the triangle because less development has emerged.

Concept

Buhlmann-Straub Credibility

Buhlmann-Straub is Buhlmann credibility with unequal exposure. It is the right model when different years, classes, or risk cells have different volumes, so each observation does not deserve the same weight.

Why Straub Changes Buhlmann

Plain Buhlmann treats each observation as equally informative. That is rarely true in insurance. A year with 10,000 car-years says more about the risk class than a year with 200 car-years, even when both produce one observed loss ratio.

Buhlmann-Straub keeps the same shrinkage idea but replaces the ordinary sample mean with an exposure-weighted mean. The credibility weight rises with total exposure, not just the number of observations.

Model Setup

For risk i and period j, let X_ij be an observed rate or average loss measure and let w_ij be its exposure. Conditional on the risk parameter Theta_i, the observations share the same hypothetical mean and have process variance inversely proportional to exposure.

The exposure-weighted mean is the statistic that gets credibility. Small cells can still contribute, but they do not move the estimate as much as large cells.

Buhlmann-Straub conditional structure

E[X_{ij}\mid\Theta_i]=\mu(\Theta_i),\qquad \operatorname{Var}(X_{ij}\mid\Theta_i)=\frac{v(\Theta_i)}{w_{ij}}

Exposure-weighted mean

\bar X_{iw}=\frac{\sum_j w_{ij}X_{ij}}{\sum_j w_{ij}},\qquad w_i=\sum_j w_{ij}

Credibility Estimate

Let m be the collective mean, EPV the expected process variance, and VHM the variance of hypothetical means. As in Buhlmann, K = EPV / VHM. The only structural change is that total exposure w_i replaces count n in the credibility weight.

Large exposure produces Z near 1, so the risk's weighted experience dominates. Small exposure produces Z near 0, so the collective mean dominates.

Buhlmann-Straub estimate

\hat\mu_i=Z_i\bar X_{iw}+(1-Z_i)m,\qquad Z_i=\frac{w_i}{w_i+K},\qquad K=\frac{EPV}{VHM}

Empirical Bayes Interpretation

Buhlmann-Straub is empirical Bayes in actuarial notation. The collective portfolio estimates the prior mean and variance components, and each risk's exposure-weighted data updates the estimate toward its own observed experience.

That is why this model fits naturally beside mixed models and hierarchical Bayesian models. The formulas are actuarial, but the statistical action is shrinkage with unequal observation precision.

Worked Example

A territory has observed loss costs of 110, 95, and 120 over exposures 10, 20, and 30. Its exposure-weighted mean is (10(110) + 20(95) + 30(120)) / 60 = 110. Suppose the collective mean is 100, EPV = 400, and VHM = 25. Then K = 16 and Z = 60 / (60 + 16) = 0.789.

The Buhlmann-Straub estimate is 0.789(110) + 0.211(100) = 107.9. The territory has worse-than-average experience, and the total exposure is large enough that most of the estimate follows the territory rather than the collective.

Outstanding Claims Use

Credibility methods for outstanding claims blend a triangle-driven estimate with an external or prior estimate. Buhlmann-Straub is a natural language for that blend when accident years have different exposure, development maturity, or volume.

In an ASTAM reserving answer, do not say only that credibility was applied. State what the two estimates are, what the exposure or maturity weight represents, and why a younger accident year receives less credibility than a mature one.

Common Traps

Trap 1: using the unweighted sample mean when exposures differ materially.

Trap 2: treating Z as a judgment factor instead of deriving it from w_i and K.

Trap 3: mixing units. If X_ij is a loss cost per exposure, w_ij must be the matching exposure measure.

Practice

Original Source-Backed Practice

4 questions built from syllabus outcomes and released-exam patterns. The prompts and answers are original, so they train the skill without copying official exam text.