Credibility And Stabilization Of Burning Costs

Why Stabilize a Burning-Cost Indication?

A historical burning cost is an average of observed layer losses on a stated exposure basis. Even after losses have been developed, trended, and adjusted for portfolio changes, that average can be unstable: a few large claims, or a run of years with no layer losses, may materially affect it. Credibility provides a systematic way to recognize this sampling uncertainty. It gives more weight to experience when the experience is more informative and blends it with a specified complement when the experience is limited.

This tutorial uses the Bühlmann model. The experience observations are assumed to be on a consistent basis, such as adjusted annual layer loss per exposure unit. The complement is a specified expected burning cost for the same coverage and basis. It is not a second historical average unless the problem explicitly defines it that way.

Definition: Let $\theta$ represent the risk characteristic underlying a hypothetical insured or portfolio. Let $X_j$ be the adjusted layer burning cost in observation period $j$, with conditional mean $\mu_\theta=E[X_j\mid\theta]$ and conditional variance $\sigma^2_\theta=\operatorname{Var}(X_j\mid\theta)$. The process variance is $v=E[\sigma^2_\theta]$, the average within-risk variance. The variance of hypothetical means is $a=\operatorname{Var}(\mu_\theta)$, the between-risk variance. Write $m=E[\mu_\theta]$ for the collective mean.

The process variance measures random fluctuation around a risk's own expected cost. The variance of hypothetical means measures how much the expected costs differ across risks. These are distinct sources of uncertainty. A layer with volatile annual losses may have a large $v$; a heterogeneous portfolio may have a large $a$. The credibility weight depends on both.

The Bühlmann Estimate

For $n$ equally exposed observations, assume that, conditional on $\theta$, the observations are independent and identically distributed, each with conditional mean $\mu_\theta$ and conditional variance $\sigma^2_\theta$. Also assume the relevant means and variances are finite. The historical average is $\bar X$. The Bühlmann estimate of the risk's expected cost is a weighted average of $\bar X$ and the collective mean $m$.

Z=\frac{n}{n+K} =\frac{na}{na+v}, \qquad K=\frac{v}{a}, \qquad \widehat{\mu}=Z\bar X+(1-Z)m.

Here $Z$ is the credibility assigned to the observed average, and $1-Z$ is the credibility assigned to the complement. $K$ is the number of fully exposed observations for which the complement and experience receive equal weight. The ratio $v/a$ has units of exposure periods when $v$ and $a$ are measured on the same cost-squared scale.

Interpretation: For fixed $a$ and $v$, additional independent exposure increases $Z$. Greater process variance decreases $Z$, because each observation is noisier. Greater variance among hypothetical means increases $Z$, because distinguishing risks using their own experience is more valuable.

Why the Credibility Weight Has This Form

The Bühlmann weight can be justified by finding the linear blend that minimizes mean squared error for estimating $\mu_\theta$. Consider an estimate $\widehat{\mu}_Z=Z\bar X+(1-Z)m$. Conditional on $\theta$, the sample average has mean $\mu_\theta$ and variance $\sigma^2_\theta/n$. Thus $\bar X-\mu_\theta$ has conditional mean zero, while $m-\mu_\theta$ is fixed given $\theta$. The cross-product in the squared error therefore has expectation zero.

\begin{aligned} E[(\widehat{\mu}_Z-\mu_\theta)^2] &= Z^2 E[(\bar X-\mu_\theta)^2] +(1-Z)^2 E[(m-\mu_\theta)^2] \\ &= Z^2\frac{v}{n}+(1-Z)^2a. \end{aligned}

The second line follows because $E[\operatorname{Var}(\bar X\mid\theta)]=v/n$, and $E[(m-\mu_\theta)^2]=\operatorname{Var}(\mu_\theta)=a$. Differentiate this quadratic with respect to $Z$ and set the derivative to zero:

2Z\frac{v}{n}-2(1-Z)a=0 \quad\Longrightarrow\quad Z\left(\frac{v}{n}+a\right)=a \quad\Longrightarrow\quad Z=\frac{na}{na+v}.

Substituting this value into the mean squared error gives the minimum linear-estimator MSE:

\operatorname{MSE}(\widehat{\mu}) =\frac{av}{na+v}.

This MSE is an unconditional measure, averaging over both the risk characteristics and the process variation. Its square root is an RMSE in the same units as the burning cost. It is not the variance of the observed average alone: $ \operatorname{Var}(\bar X)=a+v/n $ under the equal-exposure model. The MSE above instead concerns estimating the underlying expected cost $\mu_\theta$.

Worked Example: Equal-Exposure Experience and RMSE

A hypothetical excess layer has four adjusted annual burning costs of $8,000, $10,000, $12,000, and $14,000 per equal exposure unit. The specified complement is $11,000 per exposure unit. The Bühlmann parameters are $a=64$ and $v=16$, both in thousand-dollars-squared. Calculate the credibility estimate and its RMSE relative to the underlying expected cost. Use the equal-exposure Bühlmann model.

Method. Calculate the arithmetic average, then apply the Bühlmann weight. This model assumes the four observations are conditionally independent and identically distributed for a risk, with the stated $a$ and $v$.

Express the observations in thousands of dollars. Their total is $8+10+12+14=44$, so $\bar X=44/4=11$ thousand dollars. The complement is also $m=11$ thousand dollars.

The credibility constant is $K=v/a=16/64=0.25$ exposure periods. Therefore, $Z=4/(4+0.25)=4/4.25=16/17=0.941176$, and $1-Z=1/17=0.058824$.

The estimate is $ (16/17)(11)+(1/17)(11)=11 $ thousand dollars, or $11,000.00 per exposure unit. Since the experience and complement coincide, the estimate equals both.

For uncertainty, the MSE is $av/(na+v)=(64)(16)/((4)(64)+16)=1,024/272=64/17=3.764706$ thousand-dollars-squared. Thus the RMSE is $\sqrt{64/17}=1.940285$ thousand dollars, which is $1,940.29 to the nearest cent.

Check. Squaring $1.940285$ gives approximately $3.764706$, consistent with the MSE. The common mistake is to report $1,940.33; using the stated MSE and rounding the square root only at the final step gives $1,940.29.

Unequal Exposure Weights

Annual observations need not represent equal amounts of exposure. Under the Bühlmann-Straub extension, assign observation $j$ a positive exposure weight $w_j$. Assume that, conditional on $\theta$, its expected burning cost per exposure unit is $\mu_\theta$, its variance is $v/w_j$, and observations are conditionally independent. The total weight is $W=\sum_j w_j$. The weighted average is used because a higher-weight observation is assumed to have correspondingly lower variance.

\bar X_w=\frac{\sum_j w_j X_j}{W}, \qquad Z=\frac{W}{W+K}=\frac{Wa}{Wa+v}, \qquad \widehat{\mu}=Z\bar X_w+(1-Z)m.

The formula follows from the conditional variance of the weighted average: $\operatorname{Var}(\bar X_w\mid\theta)=\sum_j w_j^2(v/w_j)/W^2=v/W$. Replacing $n$ by total exposure $W$ in the equal-exposure derivation therefore gives the stated weight. The variance assumption is important: the formula is not justified merely because observations have different numeric weights. The weights must represent exposure in a model where variance scales inversely with exposure.

Worked Example: Credibility With Unequal Exposure

A hypothetical layer has three adjusted annual burning costs, in thousands of dollars per exposure unit: $X_1=20$, $X_2=24$, and $X_3=30$. Their exposure weights are $w_1=1$, $w_2=2$, and $w_3=3$. The complement is $m=22$. Let $a=36$ and $v=144$, in thousand-dollars-squared. Calculate the weighted experience, credibility estimate, and RMSE.

Method. Use the exposure-weighted Bühlmann-Straub model, which assumes conditional independence and conditional variance $v/w_j$ for observation $j$.

Total exposure is $W=1+2+3=6$. The weighted loss-cost total is $(1)(20)+(2)(24)+(3)(30)=20+48+90=158$. Hence $\bar X_w=158/6=79/3=26.333333$ thousand dollars.

The credibility constant is $K=144/36=4$. Thus $Z=6/(6+4)=0.6$, and the complement weight is $0.4$. The indicated burning cost is $(0.6)(26.333333)+(0.4)(22)=15.8+8.8=24.6$ thousand dollars, or $24,600.00 per exposure unit.

The MSE is $av/(Wa+v)=(36)(144)/((6)(36)+144)=5,184/360=14.4$ thousand-dollars-squared. The RMSE is $\sqrt{14.4}=3.794733$ thousand dollars, or $3,794.73 to the nearest cent.

Check and trap. The weighted average $26.333333$ lies between the smallest and largest observations, and the final estimate lies between that average and the $22,000 complement. An unweighted average would be $74/3=24.666667$, incorrectly treating the three exposure amounts as equal; using it would produce $0.6(24.666667)+0.4(22)=23.6$ thousand dollars instead of $24,600.

Applying Credibility to Burning Costs

A practical calculation should keep the experience, complement, and variance parameters on the same basis. If the observations are adjusted layer costs per exposure unit, then $m$, $\bar X$ or $\bar X_w$, and the credibility estimate must all be per that same unit. The parameters $a$ and $v$ must be expressed in squared units of that cost. If experience has already been adjusted for development, trend, or portfolio changes, the complement and parameter calibration must be interpreted consistently with those adjustments.

Define the observation.
Specify the layer, period, exposure basis, and adjusted cost measure. Confirm that each observation is comparable before calculating an average.

Choose the exposure measure.
Use $n$ for equal exposures. For unequal exposures, specify weights $w_j$ and calculate $W=\sum_j w_j$, provided the inverse-variance exposure assumption is appropriate.

Set the model inputs.
Identify the complement $m$, process variance $v$, and variance of hypothetical means $a$. These are model inputs, not quantities determined by the observed average alone.

Calculate and check the indication.
Compute $Z$, the relevant experience average, and $\widehat{\mu}=Z\bar X+(1-Z)m$ (or its weighted version). Verify that $0\leq Z\leq1$ and that the estimate lies between experience and complement.

The formula requires a meaningful collective complement and credible estimates of $a$ and $v$. In practice, these parameters may be estimated from a broader body of comparable data or supplied in a problem. A single risk's short history generally cannot identify both within-risk volatility and between-risk heterogeneity reliably. The model also does not remove bias caused by inconsistent loss adjustments, changing coverage, or an exposure measure that fails to reflect the information in the observations.

Situation	Implication under the model
More independent exposure	Increases total weight and raises $Z$
Higher process variance $v$	Reduces the information in each unit of exposure and lowers $Z$
Higher variance of hypothetical means $a$	Makes risk-specific experience more valuable and raises $Z$
Zero between-risk variance $a=0$	The model assigns no credibility to risk-specific deviations from $m$

Common Mistakes and Limits

Confusing the two variances. $v$ is within-risk process variance; $a$ is between-risk variance. Interchanging them reverses the intended effect on credibility.
Counting years instead of exposure. With unequal weights, use $W$, not the number of rows, under the stated Bühlmann-Straub assumptions.
Applying weights twice. If observations are already expressed as costs per exposure unit, form the weighted average as $\sum_j w_jX_j/W$; do not multiply the resulting average by exposure again.
Treating a credibility estimate as an observed burning cost. It is a model-based blend with a complement. Report both its basis and the complement used.
Assuming the formula repairs poor data. Credibility reduces the influence of limited experience; it does not correct inconsistent development, trend, coverage, or exposure adjustments.

Limited layer experience can be particularly variable because a small number of losses may dominate the annual amounts. Credibility makes the effect of that uncertainty explicit rather than relying on an informal judgment about whether the experience “looks stable.” The result is only as useful as its model assumptions and inputs. This treatment is aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus.

Key takeaway: Stabilize a burning cost by blending adjusted experience with a specified complement using a weight based on exposure, process variance, and between-risk variance. For equal exposures, $Z=na/(na+v)$; for unequal exposures under the inverse-variance model, replace $n$ by total exposure $W$.

Check Your Understanding

For each problem, use the stated Bühlmann assumptions and keep all costs in the given units.

A hypothetical layer has five equal-exposure observations with average burning cost $18,000 per unit. The complement is $15,000, $a=9$, and $v=36$, in thousand-dollars-squared. Find $Z$ and the credibility estimate. (Answer: $Z=5/9=0.5556$; $16,666.67 per unit.)
Four equal-exposure adjusted annual costs are $12,000, $18,000, $20,000, and $10,000. The complement is $19,000, $a=25$, and $v=100$, in thousand-dollars-squared. Find the estimate. (Answer: $Z=0.5000$; $17,000.00 per unit.)
Three observations have costs of $8,000, $16,000, and $20,000 with exposure weights 2, 1, and 3. The complement is $12,000, $a=16$, and $v=64$, in thousand-dollars-squared. Find the weighted experience and estimate. (Answer: weighted experience $15,333.33; $Z=0.6000$; estimate $14,000.00 per unit.)
For three equal-exposure observations, $a=49$ and $v=98$, both in thousand-dollars-squared. Calculate the credibility weight and the RMSE for estimating the underlying expected cost. (Answer: $Z=0.6000$; MSE $=19.6$ thousand-dollars-squared; RMSE $=$4,427.19$.)
A weighted experience average is $31,000 per exposure unit, based on total weight $W=8$. The complement is $27,000, $a=25$, and $v=100$, in thousand-dollars-squared. Find the credibility estimate and its RMSE for estimating the underlying expected cost. (Answer: $Z=2/3=0.6667$; estimate $29,666.67 per unit; RMSE $=$2,886.75$.)

Situation	Implication under the model
More independent exposure	Increases total weight and raises \(Z\)
Higher process variance \(v\)	Reduces the information in each unit of exposure and lowers \(Z\)
Higher variance of hypothetical means \(a\)	Makes risk-specific experience more valuable and raises \(Z\)
Zero between-risk variance \(a=0\)	The model assigns no credibility to risk-specific deviations from \(m\)

What You'll Learn