Class Premium Estimation Versus Individual Risk Prediction

Why the Target of a Credibility Estimate Matters

A credibility estimate is not meaningful until its target is clear. An estimate may be used to set an expected loss cost for a class, or to predict the future loss of a particular risk using that risk’s own experience. Those targets can produce the same numerical estimate under a standard Bühlmann model, but they are not interchangeable interpretations. In particular, a prediction of one risk’s expected loss is not a prediction of the exact amount that risk will claim.

The distinction matters in ratemaking. A class rate is intended to reflect expected losses per exposure for the risks to which the rate applies. A prediction for one insured risk describes that risk’s expected future outcome conditional on its observed experience. It takes additional assumptions to treat a class-wide estimate as the prediction for every member of the class.

The Bühlmann Model and Its Estimate

For the standard Bühlmann model, risks have different underlying parameters, while repeated observations for a given risk share that risk’s parameter. Define \(\mu(\Theta)=E[X_j\mid\Theta]\) and \(\sigma^2(\Theta)=\operatorname{Var}(X_j\mid\Theta)\). Assume that, conditional on \(\Theta\), observations are independent and have the same conditional mean and variance. Across risks, assume the parameters follow a common distribution with finite moments. These assumptions make the observations exchangeable across risks before a particular risk’s experience is known.

Here \(\widehat{\mu}\) is a linear estimate of the risk’s expected loss per exposure, not a forecast of the next observed loss itself. The estimate balances the risk’s own experience, \(\overline{X}\), against the collective mean, \(m\). The factor \(Z\) increases as the risk’s experience volume \(n\) increases. If \(\mathrm{VHM}=0\), all risks have the same conditional mean in this model; the ratio defining \(K\) is not used, and the collective mean is the appropriate estimate of that common mean.

Why the Same Formula Can Serve Two Targets

Under the model, the target for an individual risk can be written as the expected value of its future loss conditional on its past observations. The Bühlmann estimator is a best linear predictor for that target. The derivation shows why the same estimate can also be interpreted as an estimate of that risk’s expected loss per exposure.

This derivation establishes a best linear predictor from the stated moments; it does not require normally distributed losses. Under a fully specified Bayesian model, the exact conditional expectation may use more information than this linear estimator. In the standard credibility application, the formula estimates a risk’s expected loss, and it predicts a future individual loss only as that loss’s conditional mean. Its process variability remains.

From a Risk Estimate to a Class Loss Cost

A class estimate and an individual prediction are not automatically the same quantity for every risk. If the data describe one risk’s experience, the estimate above concerns that risk. A class-wide expected loss cost concerns the risks in the class collectively. It can be estimated from class experience when the observations and exposure base represent that class. Alternatively, it can be obtained by aggregating risk-level predictions over the risks that will be covered.

For a portfolio with \(N\) risks, let \(\widehat{\mu}_r\) be the estimated expected loss per exposure for risk \(r\), and let \(w_r\) be its future exposure. The expected aggregate loss implied by those predictions is \(\sum_{r=1}^N w_r\widehat{\mu}_r\). Dividing by total exposure gives the portfolio’s predicted loss cost per exposure:

This is an exposure-weighted average, not necessarily the estimate for any one risk. If the risks’ expected losses differ, applying one risk’s estimate to the entire class can misstate the class’s expected losses. Conversely, if a single class-level estimate is applied uniformly, that is a pricing choice based on a common class expectation; it does not establish that each insured risk has that same underlying expected loss.

Individual Predictions Need Individual Experience

A risk-level credibility estimate is supported by observations linked to that risk, or by an explicitly defined grouping model. Merely knowing the class average does not provide enough information to distinguish one member’s expected loss from another’s. If only class-level data are available, a class estimate may be appropriate for class ratemaking, but assigning it as a risk-specific prediction amounts to assuming that all members have the same expected loss or that no useful risk-level differentiation is available.

The next example shows how a class average can conceal differences among individually predicted risks. The calculations use the same collective mean and variance components for each risk, and each risk has its own equal-weight history.

What a Credibility Estimate Does—and Does Not—Set

For ratemaking, a credibility estimate commonly supplies an expected loss cost or expected loss rate. It is not, by itself, a complete premium. A premium may also need to account for expenses, contingencies, and profit provisions, as well as the applicable exposure basis and coverage terms. Those additions do not change the interpretation of the credibility estimate: it is still an estimate of expected losses under the model, not a final billed amount.

The interpretation also depends on the data definition. A loss average per policy cannot be used as a loss cost per exposure without reconciling the units. Likewise, observations from different periods or exposure bases must be made comparable before the result can be used as a rate indication. The credibility formula does not repair inconsistent data, a changing risk mix, or an inappropriate model. Its assumptions and the target population must be considered before assigning meaning to the result.