Understand the framework used for the classical (limited fluctuation), Buhlmann, Buhlmann-Straub, and Bayesian credibility procedures.

A pricing actuary holds three pieces of evidence: a single risk's claim history, a portfolio-wide manual rate, and a model of how risks vary. Each credibility framework decides which pieces to trust and by how much.

What every framework shares. Credibility blends a risk's own data with an external benchmark. The blend weight $Z\in [0,1]$ governs how much the risk's observed mean $\bar{X}$ counts versus the prior or manual mean $\mu$. What separates the four procedures is the justification for choosing $Z$: pure statistical tolerance, MSE-optimal linear blend, exposure-aware MSE-optimal blend, or posterior expectation under a Bayesian model.

Classical credibility was the first procedure to enter actuarial practice. It answers a narrow question: how many observations $n$ make the sample mean trustworthy on its own?

The reliability criterion. The sample mean is fully credible when its random fluctuation around the true mean is bounded with high probability:

$$\mathrm{Pr}\left(|\bar{X}-\mu |\le k\mu \right)\ge 1-\alpha$$

Apply the central limit theorem to $\bar{X}$. The standard error shrinks like $1/\sqrt{n}$, so the inequality becomes $z_{\alpha /2}{\sigma }_{\bar{X}}/\mu \le k$.

Common mistakes

• Treating classical $n_F$ as exposures. $n_F$ is in expected claims. Plug exposure units in directly and you understate $Z$ by a factor of $1/\sqrt{\mu }$. Convert by dividing required claims by $\mu$ or by multiplying observed exposure by $\mu$ first.
• Using the square-root rule for Buhlmann. Square root is classical. Buhlmann uses $Z=n/(n+K)$. A common wrong answer of $\sqrt{n/K}$ appears when candidates blend the two frameworks.
• Failing to convert $v$ to per-exposure for Buhlmann-Straub. If the problem states variance per claim or per period, divide before substituting. Treating a per-period $v$ of 5.0 as a per-exposure $v$ when exposures average 20 inflates $K$ by 20x and crushes $Z$ toward zero.

Bottom line

• Classical (limited fluctuation): trust the sample mean only when it lands within tolerance $k$ of the true mean with probability $1-\alpha$. CLT-driven via the square-root rule, distribution-light, no model of between-risk variation.
• Buhlmann (greatest accuracy): model risks as draws from a distribution. Use $Z=n/(n+K)$ with $K=\text{EPV}/\text{VHM}$. Best linear unbiased predictor.
• Buhlmann-Straub: Buhlmann with unequal exposures. Total exposure $m$ replaces $n$; period averages are exposure-weighted; conditional variance is per unit exposure.
• Bayesian: specify a prior on $\Theta$, compute the posterior, take the posterior mean of the hypothetical mean. Exact, but distribution-heavy.

Exam shortcut

If the problem states a confidence level and a tolerance, default to classical with $n_F=(z_{\alpha /2}/k{)}^2$ and the appropriate severity or aggregate multiplier. If the problem hands you class means with class probabilities (or any setup yielding $\mu$, $v$, $a$) and equal exposures per period, it is a Buhlmann problem: compute the three constants, then $K=v/a$, then $Z=n/(n+K)$.

The full lesson (about 3,088 words, 21 min read) adds 2 worked examples, all 6 common mistakes, a self-check, free in the app.

Learning objectives

• A2

Browse all free MAS-II lessons or jump into free MAS-II practice questions.