Quantifying Uncertainty In Credibility-Based Premiums

Why Uncertainty in a Credibility Premium Has More Than One Source

A credibility premium is an estimate, not a guarantee of the future loss. Even if the risk’s underlying expected loss were known exactly, actual claims would vary around it. And when the risk’s expected loss is not known, the estimate of that mean is uncertain too. These are distinct sources of uncertainty, and they have different implications for a future-loss forecast.

This tutorial separates process variation, the randomness of claims given a risk’s expected loss, from uncertainty in the risk mean. It also distinguishes the unconditional variance of future losses from the mean squared error of a forecast based on estimated credibility. The calculations use a Bühlmann-Straub model and a normal Bayesian model; all amounts and scenarios are hypothetical.

Definition: Let $\mu$ be a risk’s latent expected loss per exposure unit, and let $S_q$ be its aggregate future loss over $q$ exposure units. Process variance is the variance of $S_q$ conditional on $\mu$. Risk-mean uncertainty is uncertainty about $\mu$. For a forecast $q\hat\mu$, the forecast-error mean squared error is $E[(S_q-q\hat\mu)^2]$, where the expectation is taken over both applicable process randomness and uncertainty in the estimate.

Two Variances That Answer Different Questions

Suppose that, conditional on a risk’s expected loss $\mu$, future losses over $q$ exposure units have mean $q\mu$ and variance $qv$. Here $v$ is the process variance per exposure unit. Suppose also that across risks, $E[\mu]=m$ and $\operatorname{Var}(\mu)=a$. The law of total variance gives the unconditional variance of future loss:

\operatorname{Var}(S_q) =E[\operatorname{Var}(S_q\mid\mu)] +\operatorname{Var}(E[S_q\mid\mu]) =qv+q^2a.

This describes the variation of future losses across risks and claim outcomes. A forecast-error calculation instead asks how far a particular forecast $q\hat\mu$ is likely to be from future loss. Under the same process assumptions, if the estimator uses past data and the future process is independent of that past conditional on $\mu$, then:

E[(S_q-q\hat\mu)^2] =qv+q^2 E[(\mu-\hat\mu)^2].

The first expression uses $a=\operatorname{Var}(\mu)$; the second uses the estimator’s mean squared error for $\mu$. Those quantities are generally not equal. Past experience can reduce uncertainty about a particular risk’s mean, so its estimated mean can have lower MSE than the variation of means across the whole population.

Uncertainty Under Bühlmann-Straub Credibility

For a standard Bühlmann-Straub model, let $X_j$ be the observed loss cost per exposure unit for period or observation $j$, with exposure weight $w_j$. Conditional on the risk parameter, assume observations are independent, have a common conditional mean $\mu$, and have conditional variance $v/w_j$, with a constant process parameter $v$. Across risks, assume $E[\mu]=m$ and $\operatorname{Var}(\mu)=a$. These assumptions give a weighted historical mean with total exposure $W=\sum_j w_j$ and variance $v/W$ conditional on $\mu$.

The Bühlmann-Straub estimate is $\hat\mu=Z\bar X+(1-Z)m$, where $\bar X=\sum_jw_jX_j/W$, $K=v/a$, and $Z=W/(W+K)$. Its MSE for estimating $\mu$ can be derived from the two independent error components: uncertainty in the latent mean around $m$, and process noise in $\bar X$.

\begin{aligned} E[(\hat\mu-\mu)^2] &=(1-Z)^2a+Z^2\frac{v}{W}\\ &=\frac{K^2}{(W+K)^2}a+ \frac{W^2}{(W+K)^2}\frac{v}{W}\\ &=\frac{aK^2+vW}{(W+K)^2} =\frac{aK(W+K)}{(W+K)^2}\\ &=a(1-Z)=\frac{v}{W+K}. \end{aligned}

The second equality uses $Z=W/(W+K)$ and $1-Z=K/(W+K)$; the third uses $v=aK$. The result applies under the stated model and fixed exposures. It is an unconditional MSE over risks and their historical process variation, not a statement that every individual risk has the same conditional uncertainty after observing its own data.

For a future aggregate over $q$ exposure units, the forecast is $q\hat\mu$. Its forecast-error MSE is $qv+q^2a(1-Z)$. By contrast, the unconditional variance of future loss is $qv+q^2a$. The $q^2$ appears because uncertainty in a per-unit mean is multiplied by $q$ in the aggregate, and variance squares that multiplier.

Worked Example: From Unequal Historical Exposure to a Future Risk Load

A hypothetical risk has two historical loss-cost observations: $X_1=\$90$ at exposure $w_1=2$, and $X_2=\$70$ at exposure $w_2=4$. A fitted Bühlmann-Straub model has $m=\$80$, $a=64$ dollars squared, and $v=256$ dollars squared per exposure unit. Forecast aggregate loss for $q=2$ future exposure units. For illustration, add a risk load equal to 10% of the forecast-error standard deviation.

Method and weighted experience. Total historical exposure is $W=2+4=6$. The exposure-weighted mean is $\bar X=[2(90)+4(70)]/6=(180+280)/6=460/6=\$76.6667$ per exposure unit.

Credibility estimate. The parameter $K=v/a=256/64=4$, so $Z=6/(6+4)=0.6$. Therefore $\hat\mu=0.6(76.6667)+0.4(80)=46+32=\$78.00$ per exposure unit. The expected aggregate forecast is $q\hat\mu=2(78)=\$156.00$.

Estimator uncertainty. Its MSE for $\mu$ is $a(1-Z)=64(0.4)=25.6$ dollars squared per exposure unit. Equivalently, the two components are $(1-Z)^2a=0.16(64)=10.24$ and $Z^2v/W=0.36(256/6)=15.36$; their sum is $25.60$.

Forecast-error variance and load. Future process variance is $qv=2(256)=512$. The mean-estimation contribution to aggregate forecast-error variance is $q^2(25.6)=4(25.6)=102.4$. Thus the total is $512+102.4=614.4$ dollars squared, with standard deviation $\sqrt{614.4}=\$24.7871$. The 10% load is $0.10(24.7871)=\$2.4787$, so the indicated amount is $\$156.00+\$2.4787=\$158.4787$, or **\$158.48 to the nearest cent**.

Check and trap avoided. The weighted numerator is $460$, and $460/6=76.6667$, not $76$. Using $76$ would incorrectly produce $\hat\mu=\$77.60$ and a $\$155.20$ forecast. Also, this calculation uses forecast-error MSE; the unconditional variance of future loss would instead be $2(256)+2^2(64)=768$, because it uses population risk-mean variance $a$, not the MSE $25.6$ of this estimator.

Deriving the Forecast-Error Decomposition

The forecast-error formula follows by adding and subtracting the conditional mean $q\mu$. Write the error as a future process deviation plus an estimation error:

S_q-q\hat\mu=(S_q-q\mu)+q(\mu-\hat\mu).

Conditional on $\mu$ and the historical observations, the first term has mean zero and variance $qv$, and the future process is independent of the historical data. Therefore its covariance with the second term is zero in the unconditional squared-error expectation. Squaring and taking expectations gives:

E[(S_q-q\hat\mu)^2] =E[(S_q-q\mu)^2]+q^2E[(\mu-\hat\mu)^2] =qv+q^2\operatorname{MSE}(\hat\mu).

This argument requires the stated conditional process variance and independence of future process from the past, given the risk mean. If future conditions introduce dependence, a trend adjustment, or a different process variance, the corresponding forecast model must account for it; the simple $qv$ term should not be applied automatically.

Worked Example: Separating Future-Loss Variance from Forecast-Error MSE

For a hypothetical Bühlmann-Straub rating, the process variance is $v=200$ dollars squared per exposure unit, the population risk-mean variance is $a=50$ dollars squared per exposure unit, and the fitted credibility estimate has MSE $30$ dollars squared per exposure unit. For $q=2$ future exposure units, calculate both the unconditional future-loss variance and the forecast-error MSE. State why they differ.

Method. Use $qv+q^2a$ for unconditional future-loss variance and $qv+q^2\operatorname{MSE}(\hat\mu)$ for forecast-error MSE.

Future-loss variance. The process contribution is $2(200)=400$. The population risk-mean contribution is $2^2(50)=4(50)=200$. Thus $\operatorname{Var}(S_q)=400+200=\mathbf{600}$ dollars squared.

Forecast-error MSE. The process contribution is again $400$. The estimator-uncertainty contribution is $2^2(30)=4(30)=120$. Thus $E[(S_q-q\hat\mu)^2]=400+120=\mathbf{520}$ dollars squared.

Check and trap avoided. The mean-uncertainty terms are $q^2a=200$ and $q^2\operatorname{MSE}(\hat\mu)=120$, respectively. Substituting $q=2$ gives $2^2=4$, not $8$; the correct totals are 600 and 520, not 800 and 640. The first quantity describes variation of future loss across the population, while the second measures error around a forecast that has already used experience.

Bayesian Posterior Predictive Uncertainty

A Bayesian calculation gives a related but conditional decomposition. After observing data, uncertainty about $\mu$ is represented by its posterior distribution. The posterior predictive variance of future loss is the expected future process variance given the data plus the variance of its conditional mean given the data. For the normal model below, future process variance does not depend on $\mu$, so the expression is $q\sigma^2+q^2\operatorname{Var}(\mu\mid\text{data})$. This is conditional on the observed data, unlike the unconditional Bühlmann-Straub MSE above.

Worked Example: A Bayesian Predictive Variance and Premium Implication

Suppose a hypothetical risk has an unknown annual loss cost $\mu$. Before observing experience, assume $\mu\sim N(100,400)$, where the second parameter is the variance and all amounts are dollars per exposure unit. Conditional on $\mu$, assume three independent historical annual loss costs are normally distributed with mean $\mu$ and variance $\sigma^2=900$ each. The observations are $\$80,\$100,\$120$. For $q=2$ future exposure units, calculate the posterior predictive mean and variance. Then add a 5% loading on the predictive standard deviation.

Method and posterior parameters. The prior precision is $1/400$. The data precision is $3/900=1/300$. The posterior variance is $(1/400+1/300)^{-1}=(7/1200)^{-1}=1200/7=\$171.4286$ squared per exposure unit.

The sample mean is $(80+100+120)/3=\$100$. The posterior mean is the precision-weighted average: \[ \frac{100/400+3(100)/900}{1/400+3/900} =\frac{0.25+0.3333\ldots}{7/1200} =\frac{7/12}{7/1200}=\$100. \] Thus the predicted aggregate mean is $2(100)=\$200$.

Predictive variance. Future process variance is $q\sigma^2=2(900)=1800$. Posterior uncertainty in the future aggregate mean is $q^2(1200/7)=4(1200/7)=4800/7=685.7143$. The total posterior predictive variance is $1800+4800/7=17400/7=2485.7143$ dollars squared. Its standard deviation is $\sqrt{2485.7143}=\$49.8579$.

The 5% loading is $0.05(49.8579)=\$2.4929$, producing an indicated amount of $\$200+\$2.4929=\$202.4929$, or \$202.49 to the nearest cent. The calculation checks because the predictive mean is $q$ times the posterior mean, while the uncertainty in that mean is multiplied by $q^2$. The loading is an illustrative pricing convention, not part of the posterior mean itself.

Interpreting Uncertainty for Premium Decisions

A premium implication depends on what quantity is being priced and how the insurer chooses to reflect uncertainty. The expected-loss estimate $q\hat\mu$ is the central forecast under the model. A risk load can be added using a stated rule, such as a percentage of forecast-error standard deviation, a variance loading, or a capital-based measure. These rules are not interchangeable. A standard-deviation loading, for example, must identify which uncertainty distribution and variance it uses.

The examples used $\lambda$ times the forecast-error standard deviation as an illustrative additive load. Under Bühlmann-Straub assumptions, that standard deviation is $\sqrt{qv+q^2\operatorname{MSE}(\hat\mu)}$. Under a Bayesian posterior predictive model, it is the square root of the posterior predictive variance. A practitioner should not silently substitute the unconditional population variance of future loss when the stated task concerns error around a credibility estimate.

Key takeaway: Separate the future process term from uncertainty in the risk mean. For an aggregate forecast $q\hat\mu$, the general structure is process variance plus $q^2$ times the relevant uncertainty about the mean. Use estimator MSE for an unconditional forecast-error calculation, population variance of risk means for unconditional future-loss variance, and posterior variance for a conditional Bayesian predictive calculation.

Common Exam Traps

Using the wrong mean-uncertainty quantity. Future-loss variance uses $a$; forecast-error MSE uses $E[(\mu-\hat\mu)^2]$. A Bayesian predictive calculation after data uses $\operatorname{Var}(\mu\mid\text{data})$.
Forgetting to square aggregate exposure. Uncertainty in a per-unit mean contributes $q^2$ times its variance, not $q$ times its variance.
Confusing process variance with process standard deviation. Variance terms add; take a square root only after the full variance has been calculated.
Weighting unequal exposures incorrectly. In the Bühlmann-Straub weighted mean, multiply each observation by its exposure before summing, then divide by total exposure.
Treating a risk load as part of the credibility estimate. First calculate the expected-loss forecast; then apply the expressly stated loading rule.

Check Your Understanding

For each problem, identify which uncertainty quantity is requested before calculating.

A Bühlmann-Straub model has $v=144$, $a=36$, and total historical exposure $W=4$. Calculate $Z$, the estimator MSE for $\mu$, and the forecast-error variance for $q=1.5$. (Answer: $Z=0.5$, estimator MSE $=18$, forecast-error variance $=256.5$ dollars squared.)
A future aggregate has $q=3$, process variance $v=100$ per exposure unit, and the estimator MSE for the per-unit mean is $25$. Calculate its forecast-error variance and standard deviation. (Answer: variance $=525$ dollars squared; standard deviation $\approx \$22.9129$.)
For $q=4$, let $v=80$, population risk-mean variance $a=20$, and estimator MSE $=12$. Calculate unconditional future-loss variance and forecast-error MSE. (Answer: future-loss variance $=640$ dollars squared; forecast-error MSE $=512$ dollars squared.)
A Bayesian normal model has prior $\mu\sim N(50,100)$. Two independent observations have conditional distribution $N(\mu,200)$ and sample mean $\$60$. Find the posterior mean and variance, then the predictive mean and variance for $q=2$. (Answer: posterior mean $=\$55$, posterior variance $=50$; predictive mean $=\$110$, predictive variance $=600$ dollars squared.)
A forecasted aggregate mean is $\$300$, and its forecast-error variance is $225+64$ dollars squared. Under a 12% loading on forecast-error standard deviation, calculate the indicated amount. (Answer: standard deviation $=\$17$, load $=\$2.04$, indicated amount $=\$302.04$.)

What You'll Learn