Credibility Theory: Limited Fluctuation, Bühlmann, and Bayes

Credibility theory answers a single practical question: how much weight should be placed on an individual risk's own observed experience versus a broader collective (class, industry, prior) estimate? Two traditions — limited fluctuation and greatest-accuracy (Bühlmann/Bayesian) credibility — give different but related answers.

Limited fluctuation credibility

Limited fluctuation credibility asks: how much data is needed so that the observed mean is "close enough" to the true mean with high probability? Full credibility is achieved when the observed aggregate losses are within $\pm k$ of the true expected value with probability $P$, i.e.

$\Pr\big(|\hat X - \mu| \le k\mu\big) \ge P.$

Assuming claim counts are Poisson-distributed (so that the coefficient of variation of aggregate claim counts is $1/\sqrt{n}$) and invoking the normal approximation, this probability statement becomes $z_{\alpha/2}\cdot \sigma/\sqrt n \le k\mu$, i.e. $\sqrt n \ge z_{\alpha/2}/k$, so the full-credibility standard for claim counts is

$n_F = \left(\frac{z_{\alpha/2}}{k}\right)^2.$

For the classic standard $P = 90\%$ ($z_{0.05}=1.645$) and $k=5\%$, $n_F = (1.645/0.05)^2 = 32.9^2 \approx 1{,}082$ expected claims. If severity also varies, the standard is inflated by $1+CV_{severity}^2$ to account for the added variance from claim size, giving $n_F^{severity} = \left(\frac{z_{\alpha/2}}{k}\right)^2\left(1+CV_{sev}^2\right)$.

Below full credibility, the square-root rule assigns partial credibility

$Z = \min\!\left(\sqrt{\frac{n}{n_F}},\, 1\right),$

where $n$ is the actual (expected) claim count observed. This is an ad hoc but historically dominant rule because it is simple, transparent, and matches full credibility at $n=n_F$ while going to zero smoothly as $n\to 0$.

Bühlmann credibility

Greatest-accuracy credibility instead minimizes the mean squared error of a linear estimator of a risk's true mean $\mu(\Theta)$, given observations $\bar X$. The Bühlmann model decomposes total variance into two pieces:

• Expected value of the process variance (EPV), $EPV = E_\Theta[\mathrm{Var}(X\mid\Theta)]$ — the average within-risk variance.
• Variance of hypothetical means (VHM), $VHM = \mathrm{Var}_\Theta(E[X\mid\Theta])$ — the between-risk variance.

The Bühlmann credibility factor is

$Z = \frac{n}{n+K}, \qquad K = \frac{EPV}{VHM},$

and the credibility estimate of an individual risk's mean is $\hat\mu_i = Z\bar X_i + (1-Z)\bar X$, a linear-least-squares compromise between the risk's own mean $\bar X_i$ and the overall collective mean $\bar X$. As $n\to\infty$, $Z\to 1$ (trust own experience); as $VHM\to 0$ (all risks alike), $Z\to0$ (trust the collective).

Bühlmann–Straub extension

Bühlmann–Straub generalizes to unequal exposure per period/risk (e.g., varying payroll or car-years), weighting observations by exposure $m_{ij}$:

$Z_i = \frac{m_i}{m_i + K}, \qquad m_i=\sum_j m_{ij}, \qquad K=\frac{EPV}{VHM}.$

Worked example: three groups, Bühlmann–Straub

Three territories report pure premiums (loss/exposure) over 3 years:

Step 1 — exposure-weighted means per territory ($\bar X_i = \sum m_{ij}X_{ij}/m_i$):

• $A$: $m_A=370$; $\bar X_A = (100(8.0)+150(9.5)+120(7.0))/370 = (800+1425+840)/370 = 3065/370 = 8.284$
• $B$: $m_B=600$; $\bar X_B = (200(5.0)+180(6.0)+220(5.5))/600 = (1000+1080+1210)/600 = 3290/600 = 5.483$
• $C$: $m_C=180$; $\bar X_C = (50(12.0)+60(10.0)+70(13.0))/180 = (600+600+910)/180 = 2110/180 = 11.722$

Step 2 — overall mean: total exposure $m=370+600+180=1150$; total losses $= 3065+3290+2110=8465$; $\bar X = 8465/1150 = 7.361$.

Step 3 — EPV (average, exposure-weighted, of each territory's within-group weighted variance around its own mean):

For territory $A$: $\sum_j m_{Aj}(X_{Aj}-\bar X_A)^2 = 100(8.0-8.284)^2+150(9.5-8.284)^2+120(7.0-8.284)^2$ $= 100(0.0807)+150(1.478)+120(1.649) = 8.07+221.7+197.9 = 427.7$; divided by $(3-1)=2$ degrees of freedom gives $213.8$.

For territory $B$: $200(5.0-5.483)^2+180(6.0-5.483)^2+220(5.5-5.483)^2 = 200(0.2333)+180(0.2672)+220(0.00028)=46.66+48.10+0.06=94.8$; /2 = $47.4$.

For territory $C$: $50(12.0-11.722)^2+60(10.0-11.722)^2+70(13.0-11.722)^2 = 50(0.0772)+60(2.965)+70(1.633)=3.86+177.9+114.3=296.1$; /2 = $148.0$.

$EPV$ (weighted average of the three, using $(n_i-1)$-weighting, a common simplification uses simple average across groups with equal years): $EPV \approx (213.8+47.4+148.0)/3 = 409.2/3 = 136.4$.

Step 4 — VHM: compute weighted variance of $\bar X_i$ around $\bar X$, then subtract an EPV correction term (unbiased ANOVA-type estimator):

$VHM = \frac{\sum_i m_i(\bar X_i-\bar X)^2 - (r-1)\,EPV}{m - \sum_i m_i^2/m}$

with $r=3$ groups. $\sum_i m_i(\bar X_i-\bar X)^2 = 370(8.284-7.361)^2+600(5.483-7.361)^2+180(11.722-7.361)^2$ $=370(0.852)+600(3.527)+180(19.02) = 315.3+2116.2+3423.6 = 5855.1$.

$\sum m_i^2/m = (370^2+600^2+180^2)/1150 = (136900+360000+32400)/1150 = 529300/1150=460.3$; denominator $=1150-460.3=689.7$.

$VHM = (5855.1 - 2(136.4))/689.7 = (5855.1-272.8)/689.7 = 5582.3/689.7 = 8.093$.

Step 5 — K and Z per territory: $K = EPV/VHM = 136.4/8.093 = 16.85$.

• $Z_A = 370/(370+16.85)=0.9564$
• $Z_B = 600/(600+16.85)=0.9727$
• $Z_C = 180/(180+16.85)=0.9143$

Credibility-weighted estimates: $\hat\mu_A = 0.9564(8.284)+0.0436(7.361)=7.925+0.321=8.246$; similarly $B$ and $C$ shrink only slightly toward 7.361 given their large exposure.

Bayesian credibility: Poisson–Gamma

When claim counts $X\mid\lambda \sim \text{Poisson}(\lambda)$ and the prior $\lambda\sim\text{Gamma}(\alpha,\beta)$ (mean $\alpha\beta$), the posterior given $n$ years of data $x_1,\dots,x_n$ is also Gamma, and the posterior mean is exactly a credibility-weighted average:

$E[\lambda\mid \mathbf x] = \frac{\alpha+\sum x_i}{1/\beta+n} = Z\bar x + (1-Z)\,\alpha\beta, \qquad Z=\frac{n}{n+1/\beta}.$

This exact match — the Poisson–Gamma conjugate posterior mean equals the Bühlmann credibility formula — is why Bühlmann credibility is called "exact credibility" for this family; it is one of very few conjugate pairs where linear (Bühlmann) and full Bayesian credibility coincide precisely.

Complement of credibility (Boor)

When $Z<1$, the complement $(1-Z)$ must be assigned to something other than the raw overall mean. Boor's paper on complements of credibility catalogs choices: a larger geographic pool, a trended prior rate, a competitor filing, or a rate derived from a related class, each with tradeoffs in bias vs. stability. The general form is

$\hat P = Z\cdot X + (1-Z)\cdot C,$

and a well-chosen complement $C$ should be unbiased, independent of $X$'s sampling error, and responsive to the same underlying trend as $X$ — using last year's unadjusted rate as $C$, for instance, biases the estimate stale.

Pitfalls

• Confusing $n$ (actual claims) with $n_F$ (the standard) — the standard is a threshold, not a target to be hit exactly.
• Applying claim-count-only full credibility standards to loss ratios without the $(1+CV^2_{sev})$ severity adjustment.
• Using unweighted (equal-year) EPV/VHM formulas on unequal-exposure data, which biases $K$ — Bühlmann–Straub's exposure weighting is not optional when volumes vary materially by period or class.
• Picking a complement of credibility correlated with the direct data, which understates variance of the blended estimate.

Exam relevance

Full-credibility standards, the square-root rule, and Bühlmann/Bühlmann–Straub mechanics are central to CAS Exam MAS-I and Exam 5; the Poisson–Gamma conjugate result is a classic Exam STAM/MAS-I Bayesian credibility question type.

Further reading

• Mahler & Dean, Credibility, CAS Study Note.
• Klugman, Panjer & Willmot, Loss Models, Chapter on Credibility.
• Boor, Credibility Based on Accuracy, PCAS.