Deriving The Bühlmann Experience Modification

Why Derive an Experience Modification?

An experience-rated premium combines a risk’s own observed experience with a collective benchmark. The Bühlmann formula is often introduced as a weighted average, but its weights have a precise justification: among linear estimates based on the observed experience, it minimizes expected squared error in estimating the risk’s hypothetical mean. Writing the premium relative to the collective premium then gives a multiplicative experience modification, which is convenient when the starting rate is expressed as a dollar premium or loss cost.

The derivation concerns equal-exposure observations under the classical Bühlmann model. It does not by itself establish a nonlinear posterior mean, handle unequal exposures, or specify operational rating constraints. Those distinctions matter when deciding whether the formula fits a particular rating problem.

Model and Assumptions

Suppose a risk has \(n\) observations \(X_1,\ldots,X_n\). Conditional on \(\Theta\), assume these observations are independent and identically distributed, with finite second moments, conditional mean \(m(\Theta)\), and conditional variance that may depend on \(\Theta\). Assume also that \(0<\mathrm{VHM}<\infty\) and \(0\leq\mathrm{EPV}<\infty\). Equal exposure and a common measurement basis across observations are implicit in this setup.

The target is \(m(\Theta)\), not the next random observation itself. The distinction is important: an observation has both risk-level variation and process variation, while the hypothetical mean describes the risk level without the new observation’s process noise.

Here, \(\mathrm{EPV}\) measures the average variability of observations around their own risk’s hypothetical mean; \(\mathrm{VHM}\) measures the variability of those hypothetical means across risks. Greater process noise relative to between-risk variation makes individual experience less informative, while more observations increase its weight.

Least-Squares Derivation

Consider a linear predictor of the hypothetical mean of the form \(a+\sum_{j=1}^n b_jX_j\). Choose the coefficients to minimize the unconditional expected squared prediction error \(E[(m(\Theta)-a-\sum_j b_jX_j)^2]\). This is a best linear predictor: it minimizes squared error within the class of linear functions of the observations. Without additional distributional assumptions, it need not equal the full conditional expectation \(E[m(\Theta)\mid X_1,\ldots,X_n]\).

The observations are exchangeable under the model, so the minimizing coefficients on the observations are equal. Write their common value as \(b\). The least-squares residual must have zero covariance with each \(X_j\). This condition follows by differentiating the expected squared error with respect to its coefficient: the derivative is minus twice the covariance of the residual with \(X_j\), so at an interior minimum that covariance is zero.

To see these identities, condition on \(\Theta\). Since \(E[X_j\mid\Theta]=m(\Theta)\), the covariance of \(m(\Theta)\) and \(X_j\) is \(\operatorname{Var}(m(\Theta))\). The law of total variance gives \(\operatorname{Var}(X_j)=E[\operatorname{Var}(X_j\mid\Theta)]+\operatorname{Var}(m(\Theta))\). For two distinct observations, conditional independence makes their conditional covariance zero, leaving only the covariance of their shared conditional mean across risks.

Apply the zero-covariance condition to the residual \(m(\Theta)-a-b\sum_{k=1}^nX_k\) and any \(X_j\). There is one own-observation covariance and \(n-1\) cross-observation covariances:

Thus \(b=\mathrm{VHM}/(\mathrm{EPV}+n\mathrm{VHM})=1/(n+K)\). The least-squares predictor’s intercept makes its expected value equal to \(E[m(\Theta)]=\mu\), because \(E[X_j]=\mu\). Therefore \(a=\mu-bn\mu\). Substituting the value of \(b\) gives the Bühlmann form:

This derivation explains both the weight and the collective complement. It also gives a useful check: if process variance is zero, \(K=0\) and \(Z=1\) for positive \(n\); if process variance is large relative to VHM, \(Z\) is smaller. If VHM is zero, all risks have the same hypothetical mean under the model, \(K\) is not a finite ratio, and the appropriate predictor is the collective mean rather than this formula.

From Premium to Experience Modification

When the collective premium \(\mu\) is positive, divide the Bühlmann premium by it to obtain the multiplicative experience modification \(M\). A modification of \(1.08\) means a premium 8% above the collective benchmark; it is not itself an 8% adjustment.

The percentage adjustment from the collective premium is \(M-1=Z(\bar X-\mu)/\mu\). The experience deviation from the collective is therefore scaled by credibility. If \(\bar X=\mu\), the modification is 1. If experience is above the collective, the modification exceeds 1, provided the model quantities and \(\mu\) are positive as assumed for this interpretation. If \(\mu=0\), the ratio is undefined, although the additive Bühlmann premium may still be defined.

Exam Traps and Model Boundaries

A frequent error is to reverse the variance ratio and use \(\mathrm{VHM}/\mathrm{EPV}\) as \(K\). That makes more process noise increase credibility, which contradicts the model’s interpretation. For example, with \(\mathrm{EPV}=900\), \(\mathrm{VHM}=300\), and \(n=5\), the correct calculation is \(K=900/300=3\) and \(Z=5/(5+3)=0.625\). The reversed ratio gives \(K=1/3\), and the incorrect factor is \(5/(5+1/3)=15/16=0.9375\). The correction is to define \(K\) as process variance divided by between-risk variance before calculating \(Z\).

Other checks help prevent interpretation errors:

• Use the arithmetic mean of the equal-exposure observations, not their total, in \(P=Z\bar X+(1-Z)\mu\).
• Report \(M=P/\mu\) as a factor, and \(M-1\) as the proportional adjustment. Do not label a factor of 1.08 as an 8% factor.
• Keep the target clear: the derivation estimates \(m(\Theta)\), the hypothetical mean. It is not a promise that the next observed loss cost will equal the premium.
• Do not apply the equal-exposure formula unchanged when exposure differs by period or observation. The assumptions used in the covariance derivation would need to be replaced by a model that accounts for the differing exposures.

The premium is a least-squares linear estimate, so its quality depends on the model and its variance components. If a rating application requires bounds, credibility adjustments, or additional constraints, those are separate decisions and do not follow from the unconstrained least-squares derivation.