advanced 11 min read
Finance & economics · Topic
Credibility Theory
bayes theorem · probability theory · hypothesis testing
Credibility theory provides actuaries with a principled way to weight individual policyholder experience against prior collective data. When a single risk unit has limited history, its own observed losses are statistically noisy; when the history is long, the data become highly informative. The Bühlmann model—a linear approximation to full Bayesian updating—delivers a computationally tractable credibility factor that optimally blends the two sources of information. Extensions to the Bühlmann-Straub model handle varying exposure weights, making credibility the workhorse of commercial lines ratemaking.

Limited Fluctuation Credibility

Limited fluctuation credibility (also called classical or American credibility) assigns full credibility to an insured’s observed data when the observed claim count $n$ is large enough that the estimated pure premium lies within $\pm k\%$ of the true value with probability $p$.

If claims are Poisson-distributed, the full credibility standard for claim frequency is:

\[n \ge n_0 = \left(\frac{z_{(1+p)/2}}{k}\right)^2\]

For $p = 90\%$, $k = 5\%$: $z_{0.95} = 1.645$, so $n_0 = (1.645/0.05)^2 = 1082$.

When $n < n_0$, partial credibility uses:

\[Z = \sqrt{n / n_0}, \quad Z \in [0, 1]\]

The credibility estimate is a blend of observed mean $\bar{X}$ and prior mean $\mu_0$:

\[\hat{\mu} = Z\bar{X} + (1-Z)\mu_0\]

This approach lacks a firm theoretical basis (the square-root rule is ad hoc), which motivates the Bühlmann approach.

Bühlmann Credibility Model

Let $\theta$ be a risk parameter drawn from a prior distribution $\pi(\theta)$. Given $\theta$, the annual losses $X_1, X_2, \ldots, X_n$ are i.i.d. with:

\(\mu(\theta) = E[X_i \mid \theta] \quad\text{(hypothetical mean)}\) \(\sigma^2(\theta) = \text{Var}(X_i \mid \theta) \quad\text{(process variance)}\)

The three structural parameters are:

\[\mu = E[\mu(\theta)], \quad v = E[\sigma^2(\theta)], \quad a = \text{Var}(\mu(\theta))\]

Here $v$ is the expected process variance (within-risk variability) and $a$ is the variance of hypothetical means (between-risk variability).

The Bühlmann premium is the best linear estimator of $\mu(\theta)$ given $X_1, \ldots, X_n$, minimizing the expected squared error among all linear functions of the data:

\[\hat{\mu} = Z\bar{X} + (1-Z)\mu\]

where the credibility factor is:

\[Z = \frac{n}{n + k}, \quad k = \frac{v}{a}\]

The ratio $k = v/a$ is the Bühlmann $k$: high process variance or low between-risk variance makes $k$ large, reducing the weight given to own experience.

Key properties:

  • $Z \to 1$ as $n \to \infty$ (full credibility with infinite data)
  • $Z \to 0$ as $a \to 0$ (all risks are homogeneous; prior always dominates)
  • $Z \to 1$ as $v \to 0$ (observation is noise-free; one year suffices)

Bühlmann–Straub Model

The Bühlmann–Straub extension handles varying exposure. Let $m_i$ be the exposure (e.g., premium, earned units) in year $i$, and let $\bar{X}_i = X_i / m_i$ be the loss ratio. Assume:

\[E\!\left[\bar{X}_i \mid \theta\right] = \mu(\theta), \quad \text{Var}\!\left[\bar{X}_i \mid \theta\right] = \frac{\sigma^2(\theta)}{m_i}\]

The credibility estimate remains:

\[\hat{\mu} = Z\bar{X}_w + (1-Z)\mu\]

where $\bar{X}_w = \sum m_i \bar{X}_i / m$ is the exposure-weighted average, $m = \sum m_i$, and:

\[Z = \frac{m}{m + k}\]

This is formally identical to Bühlmann but with total exposure $m$ replacing count $n$.

Estimation of structural parameters uses the within-group and between-group sum of squares:

\[\hat{v} = \frac{1}{n-r} \sum_{i=1}^r \sum_{j=1}^{n_i} m_{ij}\!\left(\bar{X}_{ij} - \bar{X}_{i\cdot}\right)^2\] \[\hat{a} = \frac{1}{c}\left[\sum_{i=1}^r m_i(\bar{X}_{i\cdot} - \bar{X})^2 - (r-1)\hat{v}\right]\]

where $c$ is a correction factor. This empirical Bayes approach avoids specifying the full prior.

Bayesian Interpretation

The Bühlmann premium is exactly Bayesian when $(X_i \mid \theta)$ is from the exponential family with a conjugate prior. In those cases, the posterior mean is linear in the data.

Example — Poisson-Gamma model: Suppose $(X \mid \lambda) \sim \text{Poisson}(\lambda)$ and $\lambda \sim \text{Gamma}(\alpha, \beta)$ (prior). After observing $n$ years with total claims $s$:

\[(\lambda \mid X_1, \ldots, X_n) \sim \text{Gamma}(\alpha + s, \; \beta + n)\]

The posterior mean is:

\[E[\lambda \mid \text{data}] = \frac{\alpha + s}{\beta + n} = \frac{n}{n + \beta}\cdot \frac{s}{n} + \frac{\beta}{n+\beta}\cdot \frac{\alpha}{\beta} = Z\bar{X} + (1-Z)\mu\]

with $Z = n/(n+\beta)$ and $k = \beta = v/a$. The Bühlmann formula is exact.

Other conjugate pairs used in credibility:

Likelihood Prior $k$
Poisson($\lambda$) Gamma($\alpha, \beta$) $\beta$
Normal($\mu, \sigma^2$) Normal($\mu_0, \tau^2$) $\sigma^2/\tau^2$
Binomial($m, q$) Beta($a, b$) $(a+b+1)/m$

Credibility in Practice

Commercial lines insurers apply the Bühlmann-Straub framework to experience rating and schedule rating:

Experience modification factor for workers’ compensation:

\[\text{Mod} = Z \cdot \frac{\text{Actual losses}}{\text{Expected losses}} + (1-Z)\]

The credibility factor $Z$ for small accounts uses a simplified formula based on expected losses:

\[Z = \frac{E}{E + K}, \quad K = \frac{v}{a}\]

where $K$ is the ballast (NCCI terminology), calibrated periodically from industry data.

Hierarchical credibility (Jewell, 1975) extends the model to multiple levels — e.g., state, industry, individual risk — each with its own credibility weight:

\(\hat{\mu}_{ij} = Z_{ij}\bar{X}_{ij} + (1-Z_{ij})\hat{\mu}_i\) \(\hat{\mu}_i = Z_i \bar{X}_i + (1-Z_i)\mu\)

This recursive structure mirrors empirical Bayes shrinkage estimation and is directly connected to linear mixed models in statistics, where the credibility weights correspond to the optimal BLUP weights.