Why the Credibility Factor Has This Form
In Bühlmann-Straub credibility, a risk may be observed over several periods, with a different amount of exposure in each period. More exposure makes the risk’s own experience more informative, but process variability can make that experience noisy. The credibility factor balances these two sources of uncertainty: the risk’s own observed mean and the collective mean.
The factor is \(Z=w/(w+K)\). Here \(w\) is the risk’s total exposure, while \(K\) is the structural constant: the ratio of process variance to variance among the risks’ hypothetical means. Understanding the units and variance components behind these quantities helps prevent common calculation errors.
The credibility estimate combines the risk’s weighted mean with the collective mean: \(Z_i\bar Y_i+(1-Z_i)m\). The factor \(Z_i\) is dimensionless and lies between zero and one when \(w_i\), \(\phi\), and \(a\) are positive. The structural constant \(K\) has the same exposure units as \(w_i\), so the sum \(w_i+K\) is meaningful.
Deriving \(Z=w/(w+K)\)
The weighted mean \(\bar Y_i\) is centered on the risk’s hypothetical mean. Given \(\mu_i\), it has conditional mean \(\mu_i\) and conditional variance \(\phi/w_i\). To see the variance result, use conditional independence and the specified period variances:
Thus a risk’s weighted mean contains between-risk variation, \(a\), as well as process variation, \(\phi/w_i\). The credibility estimate is the best linear predictor of \(\mu_i\) from \(\bar Y_i\). Its slope is the covariance of \(\mu_i\) and \(\bar Y_i\), divided by the variance of \(\bar Y_i\).
Write \(\bar Y_i=\mu_i+\varepsilon_i\), where the process error has mean zero and is uncorrelated with \(\mu_i\). Therefore, \(\operatorname{Cov}(\mu_i,\bar Y_i)=\operatorname{Var}(\mu_i)=a\).
By the variance decomposition, \(\operatorname{Var}(\bar Y_i)=a+\phi/w_i\).
The slope is \(a/(a+\phi/w_i)\). Multiplying the numerator and denominator by \(w_i/a\) gives the credibility factor.
The corresponding linear predictor is \(m+Z_i(\bar Y_i-m)\), which is the same as \(Z_i\bar Y_i+(1-Z_i)m\). If the hypothetical means and conditional observations are jointly normal, this linear predictor is also the conditional expectation of \(\mu_i\) given \(\bar Y_i\). Without normality, the formula remains the best linear predictor under the stated moment assumptions; it need not be the full conditional expectation.
Interpreting the Structural Constant
The numerator \(\phi\) measures process variation: how much observations for the same risk vary around that risk’s hypothetical mean. The denominator \(a\) measures variation in hypothetical means across risks. Therefore, \(K=\phi/a\) compares within-risk noise with between-risk heterogeneity.
- A larger \(\phi\), holding \(a\) fixed, means noisier period experience and a larger \(K\).
- A larger \(a\), holding \(\phi\) fixed, means more variation among risks and a smaller \(K\).
- Total exposure \(w_i\) adds information by reducing the process variance of the weighted mean to \(\phi/w_i\).
These interpretations are relative to the chosen exposure unit. If exposure is re-expressed in a different unit, the numerical value of \(w_i\) and the process parameter \(\phi\) must change consistently. The credibility factor itself should not change merely because the unit label changes.
Worked Example: Combining Period Experience With the Collective Mean
A hypothetical coverage program models a risk’s annual loss cost \(Y_{ij}\) in dollars per exposure unit. Assume \(\mu_i\sim N(1{,}200,200)\), where 200 is the variance, and \(Y_{ij}\mid\mu_i\sim N(\mu_i,800/w_{ij})\), independently across periods. There is no interest or mortality component. The risk has three periods: exposure 1 with loss cost $1,100, exposure 2 with loss cost $1,400, and exposure 1 with loss cost $1,300. Find its credibility factor and estimate, using the collective mean \(m=$1{,}200\).
Method. Calculate the total exposure and exposure-weighted mean, then compute \(K\), \(Z_i\), and the credibility-weighted estimate.
Total exposure is \(w_i=1+2+1=4\). The weighted loss-cost total is \(1(1{,}100)+2(1{,}400)+1(1{,}300)=1{,}100+2{,}800+1{,}300=5{,}200\). Thus \(\bar Y_i=5{,}200/4=$1{,}300\).
The process-variance parameter is \(\phi=800\), and the variance of hypothetical means is \(a=200\). Therefore, \(K=800/200=4\) exposure units, and \(Z_i=4/(4+4)=0.5\).
The estimate is \(0.5($1{,}300)+0.5($1{,}200)=$650+$600=$1{,}250\). The result is halfway between the risk’s weighted mean and the collective mean, as expected when \(Z_i=0.5\).
Check. The estimate differs from the collective mean by $50, which is \(0.5\) times the risk’s $100 difference from that mean. The common mistake to avoid is treating the three periods as total exposure 3; their exposures sum to 4.
Worked Example: Recovering \(K\) and Applying the Factor
A hypothetical equipment policy has loss cost \(Y_{ij}\) in dollars per exposure unit. Assume \(\mu_i\sim N(72,100)\) and \(Y_{ij}\mid\mu_i\sim N(\mu_i,1{,}200/w_{ij})\), independently across periods. There is no interest or mortality component. One risk has exposure 2 at loss cost $70 and exposure 3 at loss cost $82. Find \(K\), \(Z_i\), and the credibility estimate if \(m=$72\).
Method. Use the variance of hypothetical means for \(a\), not its standard deviation, and weight each period’s loss cost by its exposure.
Here, \(\phi=1{,}200\) and \(a=100\), so \(K=1{,}200/100=12\) exposure units. Total exposure is \(w_i=2+3=5\). The weighted loss-cost total is \(2(70)+3(82)=140+246=386\), giving \(\bar Y_i=386/5=$77.20\).
The credibility factor is \(Z_i=5/(5+12)=5/17=0.2941176\ldots\). The estimate is \(Z_i\bar Y_i+(1-Z_i)m=(5/17)(77.20)+(12/17)(72)\). The two contributions are \(386/17=$22.705882\ldots\) and \(864/17=$50.823529\ldots\), for a total of \(1{,}250/17=$73.529411\ldots\). Rounded to the nearest cent, the estimate is $73.53.
Check. The risk’s mean exceeds the collective mean by $5.20, and the estimate exceeds the collective mean by \(5.20(5/17)=$1.529411\ldots\). The common mistake to avoid is using \(a=10\), the square of the standard deviation, instead of the given variance \(a=100\).
Worked Example: Changing Exposure Units Without Changing Credibility
A hypothetical property program initially measures exposure in blocks of 1,000 insured building-months. Assume \(\mu_i\sim N(900,300)\) in dollars per original exposure unit, and \(Y_{ij}\mid\mu_i\sim N(\mu_i,5{,}400/w_{ij})\), independently across periods. There is no interest or mortality component. A risk has total exposure 6 original units and weighted mean loss cost $1,000; the collective mean is $900. Re-express exposure in units one-tenth as large and find the factor and estimate in either unit system.
Method. First calculate the original-unit structural constant and factor. Then transform exposure and process variance together, leaving \(a\) unchanged because the loss-cost scale has not changed.
In the original units, \(\phi=5{,}400\), \(a=300\), and \(K=5{,}400/300=18\). Thus \(w_i=6\) and \(Z_i=6/(6+18)=6/24=0.25\). The estimate is \(0.25($1{,}000)+0.75($900)=$250+$675=$925\).
The new exposure unit is one-tenth as large, so the numerical exposure is ten times the old value: \(w_i'=10(6)=60\). To preserve the same conditional variance, \(\phi'\) must also be ten times as large: \(\phi'=10(5{,}400)=54{,}000\). Indeed, \(\phi'/w_i'=54{,}000/60=900\), equal to \(5{,}400/6=900\). Thus \(K'=54{,}000/300=180\), and \(Z_i'=60/(60+180)=0.25\). The estimate remains $925.
Check. Both systems give the same conditional variance for the risk mean and the same credibility factor. The common mistake to avoid is changing exposure units while leaving \(\phi\) unchanged, which would alter the model rather than merely relabel its exposure scale.
Common Exam Trap: Reversing the Variance Ratio
The structural constant is process variance divided by variance of hypothetical means: \(K=\phi/a\), not \(a/\phi\). For instance, if \(\phi=800\) and \(a=200\), then \(K=4\) exposure units. With \(w=4\), the correct factor is \(4/(4+4)=0.5\).
The wrong calculation \(K=a/\phi=200/800=0.25\) would give \(Z=4/(4+0.25)=0.9412\). That incorrectly gives almost full credibility even though process variance is four times the between-risk variance. Keep the definitions straight: \(a\) is between-risk variance, \(\phi/w\) is process variance in the observed risk mean, and the factor follows from comparing those two components.
This formulation is aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus. In applications, the variance parameters may have to be estimated from data; the factor’s interpretation still depends on what exposure units and variance definitions those estimates use.
Check Your Understanding
For each hypothetical model, calculate the requested quantity. Variances are stated directly; no interest or mortality assumptions are involved.
- A model has \(\phi=500\), \(a=100\), and total exposure \(w=4\). Find \(K\) and \(Z\). (Answer: \(K=5\) exposure units; \(Z=4/9=0.4444\).)
- A risk has \(\phi=1{,}200\), \(a=300\), total exposure 6, weighted mean $80, and collective mean $70. Find the credibility estimate. (Answer: \(K=4\), \(Z=0.6000\), estimate $76.00.)
- A risk has two periods with exposures 2 and 3 and loss costs $60 and $75, respectively. If \(\phi=1{,}000\), \(a=250\), and \(m=$66\), find \(w\), \(\bar Y\), \(K\), \(Z\), and the estimate. (Answer: \(w=5\), \(\bar Y=$69.00\), \(K=4\), \(Z=5/9=0.5556\), estimate $67.67 to the nearest cent.)
- For a model with \(w=9\), an observed credibility factor \(Z=0.75\), and \(a=240\), find \(K\) and \(\phi\). (Answer: \(K=3\) exposure units; \(\phi=720\).)
- A risk’s exposure is 8 units, its weighted mean is $1,080, and the collective mean is $1,000. If \(\phi=1{,}500\) and \(a=250\), calculate \(K\), \(Z\), and the estimate. (Answer: \(K=6\), \(Z=8/14=0.5714\), estimate $1,045.71 to the nearest cent.)