Weight an indication for credibility and defend both the method chosen and the complement of credibility it leans on.

A small territory shows a 30% indicated increase off 200 claims. You do not believe it, but you cannot ignore it either. Credibility tells you how much to trust it, and the complement tells you what to trust instead.

A small block of experience swings wildly from year to year. Random noise, not a real shift in risk, drives much of an extreme indication. Credibility is the weight $Z$ you place on your own data. The rest, $1-Z$, goes to a complement of credibility, an outside estimate.

$$\text{Estimate}=Z\cdot \text{Observation}+(1-Z)\cdot \text{Complement}$$

KEY: $Z$ rises toward 1 as your data volume grows. More claims means less random noise, so you trust your own experience more.

KEY: Any credibility measure $Z$ must satisfy three criteria. First, it is bounded: $0\le Z\le 1$, so your own data never carries negative weight and never more than full weight.

Write the volume of data as $n$. The second criterion is the statement $\partial Z/\partial n>0$. The third criterion is stated on credibility per unit of data, $\partial (Z/n)/\partial n<0$.

Common mistakes

• Using classical when the problem is Bühlmann. Seeing capital $K$ as EPV/VHM signals Bühlmann; forcing a square-root rule there is wrong. A lowercase $k$ in a problem is the classical tolerance, not the Bühlmann constant.
• Letting Bühlmann Z hit 1. $Z=n/(n+K)$ only approaches 1. Reporting Z = 1.00 from Bühlmann is a red flag.
• Building the complement from the subject data. Reusing the same 683 claims in the complement violates independence and reimports the noise you meant to escape.

Bottom line

• Credibility-weighted estimate = Z × observation + (1 − Z) × complement, with Z between 0 and 1.
• A credibility measure must be bounded in [0, 1], increase as the volume of data grows, and show a decreasing percentage change in credibility as that volume grows.
• The classical claims standard for full credibility is (z / k) squared; divide by frequency for the exposure standard, scale by (1 + CV²) when severity varies.
• Classical uses the square-root rule; Bühlmann uses Z = n / (n + K) with K = EPV / VHM; Bayesian analysis uses no explicit Z.

Exam shortcut

If the problem gives a full-credibility standard, it is classical: use the square-root rule and cap $Z$ at 1. A lowercase $k$ there is the tolerance inside that standard, not a Bühlmann input. If it gives EPV and VHM (the source abbreviates the numerator EVPV) or a value of capital $K$, it is Bühlmann: use $n/(n+K)$ and never let $Z$ reach 1. The Bühlmann complement is the prior mean.

The full lesson (about 3,611 words, 24 min read) adds 6 worked examples, all 6 common mistakes, a self-check, free in the app.

Learning objectives

• A12

Browse all free Exam 5 lessons or jump into free Exam 5 practice questions.