📊Actuarial Mathematics Unit 6 Review

Credibility theory provides a framework for combining individual risk experience with collective risk experience to produce better estimates of future losses or premiums. The core idea is straightforward: individual data tells you something specific about a risk, but it's noisy when the volume is small. Collective data is more stable but less tailored. Credibility theory gives you a principled way to blend the two.

Three concepts anchor everything that follows:

Credibility premium: the blended estimate that combines individual and collective experience
Credibility factor $Z$ : the weight assigned to individual experience (with $1 - Z$ going to collective experience)
Full vs. partial credibility: whether the individual data is reliable enough to stand on its own or needs supplementation from the group

Classical Credibility Premium

Bühlmann Model for Experience Rating

The Bühlmann model is the foundational credibility model. It assumes a portfolio of risks, each characterized by an unknown risk parameter $\Theta$ and observable outcomes $X$ . The goal is to estimate the pure premium $\mu(\Theta) = E[X \mid \Theta]$ for each risk, using both that risk's own experience and the portfolio's collective experience.

The key result: the Bühlmann credibility premium is a weighted average of the individual mean and the collective mean. This structure recurs throughout credibility theory.

Least Squares Criteria for the Credibility Premium

The credibility premium is derived by minimizing the expected squared error:

$E[(\hat{\mu}(X) - \mu(\Theta))^2]$

Within the class of linear estimators, the optimal credibility premium takes the form:

$\hat{\mu}(X) = \alpha + \beta X$

The coefficients $\alpha$ and $\beta$ are chosen to minimize that squared error. This yields the familiar credibility-weighted formula, where $\hat{\mu}(X) = Z \bar{X} + (1 - Z) \mu_0$ , with $\bar{X}$ being the individual experience mean and $\mu_0$ the collective mean. The restriction to linear estimators is what distinguishes the Bühlmann approach from greatest accuracy credibility.

Full Credibility vs. Partial Credibility

Full credibility ( $Z = 1$ ): the individual experience is large and stable enough to be used on its own. Typically this requires the number of exposures or claims to exceed a threshold derived from a specified confidence level and margin of error.
Partial credibility ( $0 < Z < 1$ ): the individual experience carries some information but isn't sufficient alone, so it's blended with collective experience.

The credibility factor $Z$ controls this blend. As the volume of individual data grows, $Z$ increases toward 1. With very little data, $Z$ is near 0 and the estimate leans heavily on the collective.

Greatest Accuracy Credibility

Conditional Distribution of Risk Parameters

Greatest accuracy credibility drops the restriction to linear estimators. It assumes a conditional distribution of the risk parameter $\Theta$ given the observations $X$ , and targets the Bayesian posterior mean directly.

Common modeling choices:

Normal distribution for continuous loss amounts
Poisson distribution for claim counts

The hyperparameters of these distributions (the parameters governing the prior on $\Theta$ ) are estimated from the collective portfolio experience.

Derivation of the Credibility Premium Formula

The greatest accuracy credibility premium is defined as:

$\hat{\mu}(X) = E[\mu(\Theta) \mid X]$

This is the posterior mean of the pure premium given the observed data. The specific formula depends on the distributional assumptions.

For the normal/normal case (normal likelihood with a normal prior), the result is linear in $X$ and the credibility factor takes the form:

$Z = \frac{n}{n + k}$

where $n$ is the number of exposures and $k$ is the ratio of the process variance to the variance of the hypothetical means ( $k = \sigma^2 / \tau^2$ , often written as $v / a$ in actuarial notation). This is the same form as the Bühlmann result, which is not a coincidence: the Bühlmann model recovers the exact Bayesian answer when the underlying distributions are normal.

Interpretation of the Credibility Factor Z

$Z$ represents the proportion of weight given to individual experience
$Z$ increases as $n$ grows, reflecting the greater statistical reliability of larger samples
$Z \to 1$ as $n \to \infty$ (individual data dominates)
$Z \to 0$ as $n \to 0$ (collective data dominates)
$Z$ also depends on the ratio $k$ : when between-risk variance $\tau^2$ is large relative to within-risk variance $\sigma^2$ , individual data becomes informative faster ( $k$ is smaller, so $Z$ is larger for any given $n$ )

Bühlmann-Straub Model

Assumptions and Notation

The Bühlmann-Straub model generalizes the basic Bühlmann model to handle unequal risk volumes and varying exposure periods. This matters in practice because policyholders differ in size.

Risk $i$ in period $j$ has observable outcome $X_{ij}$ and risk volume (exposure) $P_{ij}$
Each risk has an unknown parameter $\Theta_i$
The process variance for risk $i$ in period $j$ is inversely proportional to $P_{ij}$ : larger exposures produce less noisy observations

Derivation of the Credibility Premium Formula

The credibility premium minimizes the expected squared error, now weighted by the risk volumes $P_{ij}$ . The result is:

$\hat{\mu}_i = Z_i \bar{X}_i + (1 - Z_i) \hat{\mu}_0$

where $\bar{X}_i$ is the volume-weighted average of risk $i$ 's experience, $\hat{\mu}_0$ is the collective mean, and the credibility factor is:

$Z_i = \frac{P_{i \cdot}}{P_{i \cdot} + k}$

Here $P_{i \cdot} = \sum_j P_{ij}$ is the total exposure for risk $i$ , and $k = \sigma^2 / \tau^2$ as before.

Estimation of Model Parameters

The model requires estimates of two variance components:

Within-risk variance $\sigma^2$ (also called process variance $v$ ): captures random fluctuation in outcomes for a given risk
Between-risk variance $\tau^2$ (also called variance of hypothetical means $a$ ): captures how much the true underlying premiums differ across risks

Estimation approaches:

ANOVA-based estimators: unbiased, closed-form, and widely used in practice. These are the standard nonparametric empirical Bayes estimators.
Maximum likelihood estimation (MLE): requires distributional assumptions but can be more efficient.

Once $\hat{\sigma}^2$ and $\hat{\tau}^2$ are obtained, you plug them into the credibility factor formula to compute $Z_i$ and the credibility premium for each risk.

Empirical Bayes Credibility

Relationship to Greatest Accuracy Credibility

Empirical Bayes credibility is a practical implementation of greatest accuracy credibility. The distinction: in a full Bayesian approach, the prior distribution of $\Theta$ is specified in advance. In empirical Bayes, the prior is estimated from the data itself.

The structure is:

Prior distribution (estimated from collective experience) represents the population-level distribution of risk parameters
Likelihood function represents the individual risk's observed data
Posterior mean of $\Theta$ given $X$ serves as the credibility premium

This avoids the need to subjectively specify a prior while still leveraging the Bayesian framework.

Nonparametric Estimation of the Prior Distribution

Nonparametric methods estimate the prior distribution without assuming it belongs to a particular parametric family. This is useful when the true distribution of risk parameters has an unusual shape.

Kernel density estimation (KDE) is the most common approach. It estimates the prior density as:

$\hat{f}(\theta) = \frac{1}{nh} \sum_{i=1}^{n} K\left(\frac{\theta - \theta_i}{h}\right)$

where $K$ is a kernel function (often Gaussian) and $h$ is the bandwidth parameter. The bandwidth controls the smoothness of the estimate: too small and the density is spiky; too large and it's oversmoothed. Selection methods include cross-validation and plug-in rules.

Semiparametric Estimation of the Prior Distribution

Semiparametric methods split the estimation into two parts:

A parametric component captures the broad shape of the prior (e.g., a normal or gamma distribution)
A nonparametric component captures deviations from that parametric form

A common example is a Gaussian mixture model with an unknown number of components. The number of components and their parameters can be estimated using the EM algorithm or Bayesian model selection methods (e.g., reversible jump MCMC). This offers more flexibility than a single parametric family while being more structured than fully nonparametric estimation.

Hierarchical Credibility Models

Motivation for Hierarchical Models

Many insurance portfolios have natural groupings: risks within classes, classes within territories, territories within lines of business. Hierarchical credibility models account for these multiple levels of variation.

The benefit is borrowing strength: a risk with sparse data can draw information not just from the overall portfolio but also from its class or territory. This is especially valuable when data is unbalanced (some classes have many risks, others have few).

Model Assumptions and Notation

Risks are indexed by $i$ within class $j$ , observed over periods $k$
Unknown risk parameters $\Theta_{ij}$ exist at each level of the hierarchy
Observable outcomes $X_{ijk}$ correspond to risk $i$ , class $j$ , period $k$
Hyperparameters at each level capture the variability of risk parameters within and across levels

Derivation of the Credibility Premium Formula

The credibility premium at the individual level is a weighted average of three quantities:

$\hat{\mu}_{ij} = Z_{ij}^{(1)} \bar{X}_{ij} + Z_{ij}^{(2)} \bar{X}_{\cdot j} + (1 - Z_{ij}^{(1)} - Z_{ij}^{(2)}) \bar{X}_{\cdot \cdot}$

where $\bar{X}_{ij}$ is the individual experience, $\bar{X}_{\cdot j}$ is the class-level experience, and $\bar{X}_{\cdot \cdot}$ is the overall portfolio experience. The credibility factors $Z_{ij}^{(1)}$ and $Z_{ij}^{(2)}$ depend on the variance components at each level and the volume of data available. More data at a given level increases the credibility assigned to that level's experience.

Applications of Credibility Theory

Experience Rating in Property/Casualty Insurance

Credibility theory is the backbone of experience rating programs. Individual policyholders (or groups) have their premiums adjusted based on past claims, blended with the class or manual rate.

The Bühlmann-Straub model is the standard tool here. Risk volumes are typically measured by payroll (workers' compensation), revenue, or vehicle-years, depending on the line of business. The credibility-weighted premium ensures that a single bad year doesn't dominate the rate for a small account, while large accounts with stable history get rates that closely track their own experience.

Prospective Premium Calculation in Life Insurance

In life insurance, credibility methods estimate mortality rates by blending the experience of a specific insured group with standard mortality tables. Hierarchical credibility models are natural here because mortality varies across age, gender, smoking status, and underwriting class.

Credibility estimates feed into premium calculations that need to be adequate (cover expected claims), equitable (fair across risk classes), and competitive (not overpriced relative to the market).

Credibility for Excess Loss Coverages

Excess loss coverages (stop-loss, excess-of-loss, umbrella policies) protect against large or catastrophic losses. These losses are rare and highly variable, which means individual experience is thin and volatile.

Empirical Bayes methods are well-suited here because they can accommodate heavy-tailed loss distributions without requiring strong parametric assumptions about the prior. The credibility factor for excess layers tends to be low, reflecting the limited information in sparse large-loss data.

Bayesian Credibility Models

Conjugate Prior Distributions

Conjugate priors are chosen so that the posterior distribution belongs to the same family as the prior. This keeps the math tractable.

Likelihood	Conjugate Prior	Posterior
Normal	Normal	Normal
Poisson	Gamma	Gamma
Binomial	Beta	Beta

The conjugate structure means you can update beliefs analytically without numerical integration, which is a significant computational advantage.

Posterior Distribution of Risk Parameters

The posterior distribution combines the prior (collective experience) with the likelihood (individual experience) via Bayes' theorem:

$f(\theta \mid x) \propto f(x \mid \theta) \cdot \pi(\theta)$

The posterior represents updated beliefs about the risk parameter after observing individual data. As more data accumulates, the posterior concentrates around the true parameter value, and the influence of the prior diminishes.

Point estimates can be taken as the posterior mean, mode, or median. For credibility purposes, the posterior mean is standard because it minimizes squared error loss.

Credibility Premium as Posterior Mean

For conjugate models, the posterior mean takes the credibility-weighted form:

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

where $\mu_0$ is the prior mean and $Z$ depends on the relative precision of the prior and the data. Specifically, $Z$ increases when:

The sample size is larger (more individual data)
The prior is diffuse (less certainty in the collective estimate)
The process variance is small (individual observations are precise)

This connects the Bayesian framework directly back to the Bühlmann credibility formula, showing that the linear credibility form arises naturally from conjugate Bayesian models.

Evaluation of Credibility Estimates

Bias vs. Variance Trade-off

Credibility estimates navigate the classic bias-variance trade-off:

Bias comes from model simplifications: assuming linearity, choosing a particular prior family, or ignoring hierarchical structure
Variance comes from limited data: small samples produce unstable estimates, and estimated variance components add additional uncertainty

The credibility factor $Z$ directly controls this trade-off. A higher $Z$ reduces bias (the estimate tracks individual experience more closely) but increases variance (individual experience is noisier). A lower $Z$ does the opposite.

Mean Squared Error of the Credibility Premium

The mean squared error (MSE) provides a single measure of estimation quality:

$MSE = E[(\hat{\mu} - \mu(\Theta))^2] = \text{Bias}^2 + \text{Variance}$

The optimal credibility factor is the one that minimizes MSE. This is exactly what the Bühlmann and Bühlmann-Straub formulas deliver under their respective assumptions. When those assumptions are violated, the actual MSE may differ from the theoretical optimum.

Empirical Evaluation Using Holdout Samples

To assess how well a credibility model performs in practice, you test it on data that wasn't used to fit the model.

Steps for holdout evaluation:

Split the data into a training set (used to estimate parameters and compute credibility premiums) and a holdout set (used to evaluate predictions)
Fit the credibility model on the training data
Compute predicted premiums for the holdout risks
Compare predictions to actual outcomes using metrics like MSE, mean absolute error (MAE), or predictive log-likelihood

Cross-validation provides more robust performance estimates by repeating this process across multiple splits. Common variants include $k$ -fold cross-validation (split data into $k$ groups, rotate which group is held out) and leave-one-out cross-validation (hold out one observation at a time). Averaging performance across folds reduces the sensitivity to any particular train/test split.

📊Actuarial Mathematics Unit 6 Review