You could rate every driver as one giant class, or you could give each driver a private rate. Neither works. Partitioning is the art of slicing the portfolio finely enough that each group is alike, but not so finely that the data runs dry.
Ratemaking groups similar risks so one rate can serve all of them. Two forces govern how you draw those groups.
Homogeneity asks whether the risks inside a group truly resemble each other. A group of only sports cars is homogeneous; a group mixing sports cars and minivans is not. Homogeneous groups let one rate fit every member fairly.
Credibility asks whether the group has enough loss data to trust its own experience. A group of five policies produces a noisy, unreliable indication. A group of fifty thousand produces a stable one.
KEY: Homogeneity and credibility trade off directly. Split a class into finer cells and each cell becomes more homogeneous but holds less data, so credibility falls. Merge cells and credibility rises but homogeneity erodes.
Common mistakes
- Splitting for homogeneity while ignoring the vanishing data. A cell with 68 claims is more homogeneous but its Z near 0.25 means its own experience barely counts.
- Using n over n plus k and the square-root rule interchangeably. They are different credibility frameworks. Mixing Bühlmann k into a limited-fluctuation problem, or vice versa, mislabels the method.
- Choosing a complement that is not more stable than the group. The complement must be broader and steadier. Pairing a thin cell with an equally thin sibling as its complement adds no stability.
Bottom line
- Homogeneity means risks in a group share similar expected loss; credibility means the group has enough volume for a stable estimate.
- The two pull in opposite directions: finer partitions raise homogeneity but shrink each cell's data and cut credibility.
- Full credibility standard for frequency is often 1,082 claims (5% error, 90% probability); partial credibility uses the square-root rule Z equals the square root of n over the full-credibility count.
- Credibility Z runs from 0 to 1; the estimate blends observed data at weight Z with a complement at weight 1 minus Z.
Exam shortcut
When a problem gives claims and a full-credibility count, compute Z with the square-root rule first, then blend own indication at Z and complement at 1 minus Z. Do the Z step before anything else. If two proposed cells show nearly equal raw indications, the split fails: no homogeneity gain, only lost credibility. Recommend keeping them together.
The full lesson (about 3,202 words, 21 min read) adds 3 worked examples, all 7 common mistakes, a self-check, free in the app.
Learning objectives
- A5
Browse all free Exam 5 lessons or jump into free Exam 5 practice questions.