Dependence Between Periods In Credibility Data

Why Dependence Between Periods Matters

Bühlmann-Straub credibility uses repeated period results to estimate a risk’s expected result. Under the usual model, results from different periods are conditionally independent given the risk’s underlying mean. That assumption makes the conditional variance of an exposure-weighted average decrease in a simple way as total exposure grows. If period results are serially correlated, the same calculation generally understates or overstates the average’s process variance, and therefore changes the credibility assigned to the risk’s experience.

The issue is not simply whether periods are “related.” Statistical dependence is broader than covariance: dependent variables can have zero covariance. This tutorial focuses on the second-moment adjustment when conditional covariances are specified, and distinguishes that adjustment from the stronger question of conditional independence.

Write \(W=\sum_{j=1}^n w_j\). The exposure-weighted mean is \(\bar X_w=\sum_j w_jX_j/W\). Conditional on \(\Theta\), it still has mean \(\Theta\) whenever each period has conditional mean \(\Theta\), regardless of whether the periods are independent. Its variance, however, depends on all the conditional covariances:

Using \(\operatorname{Var}(X_j\mid\Theta)=\phi/w_j\), this becomes

The familiar \(\phi/W\) result follows when all the conditional covariances are zero. Conditional independence implies zero conditional covariance when the relevant moments exist, but the converse is not generally true. Thus, zero covariances recover this variance calculation without proving conditional independence.

Equal Exposures and Exchangeable Correlation

A useful special case has \(n\) equally exposed periods, with common conditional variance \(\sigma^2\), and the same conditional correlation \(\rho\) between every pair. This is an exchangeable second-moment structure. Since each pair’s conditional covariance is \(\rho\sigma^2\), the variance of the arithmetic mean is

For this exchangeable correlation matrix to be positive semidefinite, \(-1/(n-1)\leq\rho\leq1\). Positive correlation increases the variance relative to independent periods; negative correlation decreases it. The quantity \(n/[1+(n-1)\rho]\) can be read as an effective number of independent observations for this variance comparison, not as a general replacement for the actual data structure.

Effect on a Scalar Credibility Estimate

Suppose a scalar estimate is formed from the weighted mean. Let the collective mean be \(\mu=E[\Theta]\), with between-risk variance \(a=\operatorname{Var}(\Theta)\). Assume the conditional variance \(V=\operatorname{Var}(\bar X_w\mid\Theta)\) is constant across risks. The best linear predictor of \(\Theta\) based on this mean is \(\mu+Z(\bar X_w-\mu)\), where

This result requires the stated moment assumptions; it does not require a normal distribution. It is a linear credibility result, not in general a statement that the full conditional distribution or posterior mean has this form.

For a linear predictor \(\mu+c(\bar X_w-\mu)\), minimizing mean squared error gives \(c=\operatorname{Cov}(\Theta,\bar X_w)/\operatorname{Var}(\bar X_w)\). Substituting the covariance \(a\) and total variance \(a+V\) yields \(Z=a/(a+V)\). This also explains why serial covariance matters: it changes \(V\), even though the conditional mean remains \(\Theta\).

Unequal Exposures and the Limits of Exposure Weighting

The covariance-sum formula applies directly when exposures differ. If conditional correlations \(\rho_{jk}\) are specified, then \(\operatorname{Cov}(X_j,X_k\mid\Theta)=\rho_{jk}\sqrt{\operatorname{Var}(X_j\mid\Theta)\operatorname{Var}(X_k\mid\Theta)}\). Under the standard marginal variances \(\phi/w_j\), each covariance is \(\rho_{jk}\phi/\sqrt{w_jw_k}\). Substitution gives

This formula adjusts the variance of the usual exposure-weighted mean; it does not prove that the exposure-weighted mean is the minimum-variance estimator under dependence. With a known conditional covariance matrix \(\Sigma\), the minimum-variance linear unbiased estimator of a common conditional mean, when \(\Sigma\) is invertible, uses weights proportional to \(\Sigma^{-1}\mathbf{1}\). Exposure weights are optimal under particular variance structures, including the independent Bühlmann-Straub structure, but need not remain optimal for arbitrary serial covariance. A scalar credibility factor applied to the exposure-weighted mean is therefore a clear adjustment for that chosen summary, not necessarily the best predictor using the entire period vector.

Interpretation and Exam Traps

Positive serial covariance usually means adjacent or otherwise related periods share information, so an exposure-weighted mean contains less independent process information than the standard formula indicates. Negative covariance can have the opposite effect, but it must be consistent with a valid covariance matrix. Neither conclusion means that the risk’s underlying expected result has changed; this discussion holds the conditional mean fixed across periods.

• Do not equate dependence with covariance. Dependence can exist even when covariance is zero. A zero covariance changes the second-moment calculation but does not establish conditional independence.
• Do not omit pair terms. Every distinct pair contributes twice in the variance of a sum, which is why the formula contains \(2\sum_{j<k}\).
• Do not use the independent-period credibility factor automatically. First calculate the conditional variance of the actual summary being used, then use \(Z=a/(a+V)\) if its variance is constant and the scalar linear-prediction assumptions hold.
• Do not assume exposure weighting is always optimal. Under general dependence, the covariance structure can change the minimum-variance weights as well as the variance of the exposure-weighted mean.