From a Balanced Dataset to a Credibility Estimate
A Bühlmann calculation combines two sources of information: a risk’s own observed experience and the collective mean for the portfolio. In an integrated calculation, the challenge is to carry the data through in the right order. First estimate the process and between-risk variances, then calculate \(K\), then determine each risk’s \(Z\), and finally apply the weights to the risk mean and collective mean.
This tutorial uses a balanced portfolio: every risk has the same number of observations. That structure makes the variance-component estimates especially direct. The calculations are moment-based; they do not require a particular probability distribution for the observations. The limited-fluctuation comparison later will state a separate normal approximation explicitly.
Here, EPV is the expected process variance within a risk, while VHM is the variance of the risk means across the collective. A larger EPV means more sampling noise in a risk’s own experience. A larger VHM means more genuine variation among risks, so an individual risk’s experience carries more information about its own expected value. The ratio \(K\) expresses the amount of exposure, in observation units, needed for the risk’s own experience to have weight one-half.
Estimating EPV and VHM From Balanced Experience
For a balanced dataset, calculate the within-risk sum of squares and the between-risk sum of squares. The within-risk mean square estimates EPV. The between-risk mean square includes both process variation and variation in the underlying risk means; subtracting the within-risk component and dividing by \(r\) isolates the estimated VHM.
These formulas are for the balanced one-way Bühlmann setting described above: the same number of observations per risk, a common conditional process variance, and conditional independence. They should not be applied mechanically to an unbalanced dataset. Also, a sample estimate of VHM can be negative because of random variation. Since a variance component cannot be negative, the boundary choice must be handled explicitly; for this workshop, the examples use data with a positive estimate.
These summarize the individual experience and the collective benchmark.
Use deviations from each risk’s own mean to estimate EPV.
Use deviations of risk means from the grand mean, with the balanced factor \(r\).
Subtract MSW from MSB, divide by \(r\), and take the ratio of estimated EPV to estimated VHM.
Use that risk’s own observation count \(n\), its sample mean, and the collective mean.
Worked Example: Estimating the Variance Components and All Four Risk Estimates
A hypothetical portfolio has four risks, each observed three times. The observations, in thousands of dollars per exposure unit, are: Risk A: 8, 10, 12; Risk B: 10, 12, 14; Risk C: 14, 16, 18; and Risk D: 16, 18, 20. Estimate EPV, VHM, and \(K\), then calculate each credibility-weighted estimate. These are moment-based calculations, so no specific parametric distribution is assumed. Interest and mortality assumptions do not apply to this loss-cost dataset.
Step 1: Find the means. The four risk means are \(10\), \(12\), \(16\), and \(18\). The total of all observations is \(168\), so the grand mean is \(168/12=14\).
Step 2: Calculate MSW. Within each risk, the deviations from its mean are \(-2,0,2\), whose squared deviations sum to \(4+0+4=8\). Across four risks, the within-risk sum of squares is \(4(8)=32\). The within degrees of freedom are \(4(3-1)=8\), so \(\text{MSW}=32/8=4\). Therefore, \(\widehat{\text{EPV}}=4\).
Step 3: Calculate MSB and VHM. The risk-mean deviations from 14 are \(-4,-2,2,4\); their squared values sum to \(16+4+4+16=40\). Thus, \(\text{MSB}=3(40)/(4-1)=40\). The estimated VHM is \((40-4)/3=12\).
Step 4: Calculate \(K\). \(K=4/12=1/3\) observation. Each risk has \(n=3\), so \(Z=3/(3+1/3)=9/10=0.9\).
Step 5: Calculate each estimate. For Risk A, the estimate is \(0.9(10)+0.1(14)=10.4\). For Risk B, it is \(0.9(12)+0.1(14)=12.2\). For Risk C, it is \(0.9(16)+0.1(14)=15.8\). For Risk D, it is \(0.9(18)+0.1(14)=17.6\). Results are in thousands of dollars per exposure unit.
Check. The four estimates sum to \(10.4+12.2+15.8+17.6=56\), so their average is \(56/4=14\), the collective mean. The method is balanced ANOVA variance-component estimation followed by Bühlmann credibility. It avoids using the between-risk variation alone as VHM, which would incorrectly include process variation.
Why the Bühlmann Weight Has This Form
The credibility factor is not an arbitrary percentage. Under the Bühlmann model, it is the coefficient that minimizes mean squared error among linear estimates of the form \(\mu+a(\bar{X}_i-\mu)\), where \(a\) is a weight between zero and one. The following argument gives the reason for the formula.
Write the linear estimate as \(\mu+a(\bar{X}_i-\mu)\). Its error relative to \(\Theta_i\) is \(a(\bar{X}_i-\Theta_i)+(1-a)(\mu-\Theta_i)\). The first term is process noise with variance \(\text{EPV}/n\); the second reflects variation of the risk mean around the collective mean, with variance VHM. Their covariance is zero because the conditional process error has conditional mean zero. Therefore, the mean squared error is
Differentiate with respect to \(a\) and set the derivative equal to zero:
Solving gives \(a(\text{EPV}/n+\text{VHM})=\text{VHM}\), hence
Thus the individual experience receives more weight when \(n\) or VHM increases, and less weight when EPV increases. In practice, the variance components are estimated from data, so the resulting Bühlmann estimate is an empirical credibility estimate rather than a calculation using known population parameters.
Worked Example: Correctly Applying a Given \(K\)
A hypothetical risk has six comparable annual observations with sample mean \(40\). A balanced portfolio analysis gives \(K=3\), and the collective mean is \(34\). Calculate the risk’s credibility factor and credibility-weighted estimate. The values are on the same loss-cost basis; no interest, mortality, or parametric distribution assumption is needed for this calculation.
Method. Use the Bühlmann factor \(Z=n/(n+K)\), then weight the risk mean by \(Z\) and the collective mean by \(1-Z\).
Credibility factor. \(Z=6/(6+3)=6/9=2/3\), so \(1-Z=1/3\).
Weighted estimate. The risk’s experience contributes \((2/3)(40)=80/3\), and the collective mean contributes \((1/3)(34)=34/3\). Their sum is \(114/3=38\).
Check and trap avoided. Since the experience mean exceeds the collective mean, the estimate must lie between 34 and 40; 38 does. A common arithmetic error is to report 36 after setting up the correct weights. Recomputing the weighted sum gives \(26.6667+11.3333=38.0000\), not 36. The estimate is 38 in the original units.
Comparison With a Limited-Fluctuation Standard
A limited-fluctuation standard answers a different question from Bühlmann credibility. It asks how many observations are needed for a sample mean to fall within a specified tolerance of a target with a stated probability, under an assumed sampling model. The Bühlmann factor instead blends observed experience with the collective mean to estimate a risk’s expected value, allowing for both process variation and differences among risks.
For a transparent comparison, suppose observations have process variance EPV, the target mean is \(\mu\), and a normal approximation is used. For a relative tolerance \(q\) and two-sided confidence probability \(p\), the standard full-credibility observation count is
Here \(z_{(1+p)/2}\) is the standard normal quantile. The expression follows by requiring \(z_{(1+p)/2}\sqrt{\text{EPV}/n}\le q\mu\) and solving for \(n\). This is an approximate process-fluctuation criterion; it is not a substitute for estimating VHM or calculating \(K\). Its use depends on the stated normal approximation and on using the process variance and target relevant to the quantity being measured.
Worked Example: Checking a Limited-Fluctuation Benchmark
Use the hypothetical balanced portfolio from the first example, with estimated EPV \(4\) and collective mean \(14\). Compare its three observations per risk with a 90% two-sided limited-fluctuation standard requiring the sample mean to be within 20% of the target. For this benchmark only, assume observations are approximately normal with mean \(14\) and variance \(4\), independently across observations. No interest or mortality assumption applies.
Step 1: Find the absolute tolerance. The relative tolerance is \(q=0.20\), so \(q\mu=0.20(14)=2.8\).
Step 2: Select the normal quantile. For a 90% two-sided probability, \((1+p)/2=0.95\), and \(z_{0.95}=1.6449\).
Step 3: Calculate the required count. The standard deviation of one observation is \(\sqrt{4}=2\). Thus, \(n_0=(1.6449(2)/2.8)^2=(3.2898/2.8)^2\), approximately \(1.3805\). The required integer count is rounded up to 2 observations.
Step 4: Compare with observed volume and Bühlmann credibility. Each risk has 3 observations, which exceeds the approximate full-credibility count of 2 under this limited-fluctuation benchmark. In the earlier Bühlmann calculation, however, its weight was \(Z=0.9\), not 1. The difference is not a contradiction: the limited-fluctuation count addresses process noise around a fixed target, while Bühlmann also shrinks a risk’s experience toward the collective mean because risk means vary across the portfolio.
Check. Substituting \(n=2\) into the standard error gives \(2/\sqrt{2}=1.4142\); multiplying by \(1.6449\) gives about \(2.326\), less than the tolerance \(2.8\). At \(n=1\), the corresponding bound is \(3.2898\), greater than \(2.8\). The count of 2 is therefore the smallest whole number satisfying the stated inequality.
Calculation Checks and Common Traps
A reliable solution keeps the variance estimates, credibility parameter, and risk-specific estimate distinct. The most useful checks are short but catch many errors:
- Use the correct degrees of freedom. The within-risk denominator is \(I(r-1)\); the between-risk denominator is \(I-1\).
- Keep the balanced factor in MSB. The between-risk sum of squared deviations of risk means is multiplied by \(r\).
- Subtract before dividing for VHM. The estimate is \((\text{MSB}-\text{MSW})/r\), not MSB divided by \(r\).
- Use the individual risk’s observation count. In an integrated dataset with balanced counts this is \(r\); in a separate risk calculation, use that risk’s stated \(n\).
- Keep weights attached to the right quantities. The risk mean receives \(Z\); the collective mean receives \(1-Z\). The estimate must lie between the two means when \(0\le Z\le1\).
- Do not equate a limited-fluctuation threshold with \(Z=1\). The two methods answer different questions and may use different assumptions.
Check Your Understanding
For each problem, calculate the requested quantity before checking the answer. Unless a problem states otherwise, treat loss observations as being on a common basis and assume no interest or mortality adjustment is needed.
- A risk has 10 observations, sample mean 30, and the collective mean is 24. If \(K=5\), calculate \(Z\) and the credibility-weighted estimate. (Answer: \(Z=2/3\); estimate \(28\).)
- A balanced dataset has \(r=4\), \(\text{MSW}=6\), and \(\text{MSB}=18\). Estimate EPV and VHM, then calculate \(K\). (Answer: \(\widehat{\text{EPV}}=6\), \(\widehat{\text{VHM}}=3\), and \(K=2\).)
- A risk has five observations with mean 16; the collective mean is 10 and \(K=4\). Find its credibility estimate. (Answer: \(Z=5/9\); estimate \(13.3333\), or \(13.33\) to two decimals.)
- Assume a normal approximation with process variance \(9\), target mean \(30\), relative tolerance 10%, and 95% two-sided probability. With \(z_{0.975}=1.96\), find the minimum whole-number full-credibility count. (Answer: \(n_0=(1.96(3)/3)^2=3.8416\), so round up to \(4\).)
- In one sentence, explain why a risk with enough observations to meet a stated limited-fluctuation count can still have a Bühlmann credibility factor below 1. (Answer: The limited-fluctuation count controls process fluctuation around a target, while Bühlmann also accounts for uncertainty and variation in risk means by shrinking experience toward the collective mean.)