Why Auditing the Calculation Matters
A Bühlmann calculation can look straightforward: find \(K\), calculate \(Z\), then blend experience with a collective mean. But a reversed variance ratio, mismatched exposure units, or omitted complement can materially distort the indicated estimate. A further conceptual error is to treat \(Z\) as a probability rather than as a weight in a linear estimator.
The focus here is not a new credibility method. It is a disciplined way to check the quantities and interpretation used in the Bühlmann method. All amounts and portfolios in the examples are hypothetical.
The quantities in \(K\) are variances on the same observation basis. EPV describes variation within a risk around its own expected value; VHM describes variation among the expected values of different risks. These roles are not interchangeable. If VHM is zero, there is no between-risk variation in the model and this ratio formula does not apply; with positive EPV, the standard credibility estimate then gives no weight to individual experience.
A Short Derivation: Why the Weight Has This Form
The Bühlmann estimate is the best linear predictor of the risk’s expected value \(E[X_j\mid\Theta]\) using \(\bar X\). It is not, in general, the exact conditional expectation for every possible distribution. The derivation makes the direction of the variance ratio easier to remember.
The law of total variance gives \(\operatorname{Var}(\bar X)=\text{VHM}+\text{EPV}/n\): variability in the risk’s mean contributes VHM, while averaging \(n\) conditionally independent observations reduces process variance to \(\text{EPV}/n\). Also, \(\operatorname{Cov}(E[X_j\mid\Theta],\bar X)=\text{VHM}\). The slope of the best linear predictor on \(\bar X\) is therefore covariance divided by variance:
The predictor is \(m+Z(\bar X-m)\). Expanding it gives \(Z\bar X+(1-Z)m\). This explains why the collective complement \(1-Z\) must be present: it shifts the estimate from the collective mean toward the observed experience by the fraction \(Z\), rather than discarding the collective mean.
Check the Variance Order First
The defining ratio is \(K=\text{EPV}/\text{VHM}\), not its reciprocal. A useful diagnostic is to ask whether the resulting credibility behaves sensibly. For fixed \(n\), greater process variation should reduce the weight on an individual’s experience, while greater variation among risk means should increase it. The stated ratio has those properties: increasing EPV increases \(K\) and reduces \(Z\); increasing VHM reduces \(K\) and increases \(Z\).
Worked Example: Detecting a Reversed Variance Ratio
A hypothetical portfolio has EPV \(=400\) dollars squared per annual observation and VHM \(=100\) dollars squared across risks. One risk has three annual observations of \(\$900\), \(\$910\), and \(\$880\). The collective mean is \(\$850\). Calculate the estimate and diagnose what happens if the ratio is reversed.
Method and units. Both variance quantities are expressed per annual observation, so their ratio is dimensionless and \(n=3\) annual observations is consistent with the basis.
Step 1: Calculate \(K\) in the correct order. \(K=400/100=4\) annual observation units.
Step 2: Calculate the experience average and credibility. \(\bar X=(900+910+880)/3=2{,}690/3=\$896.67\), rounded to cents. Then \(Z=3/(3+4)=3/7=0.4286\), to four decimals.
Step 3: Include both estimator components. The estimate is \((3/7)(2{,}690/3)+(4/7)(850)=2{,}690/7+3{,}400/7=6{,}090/7=\$870.00\). The weights sum to one.
Wrong calculation and correction. Reversing the ratio gives \(K=100/400=0.25\) and \(Z=3/(3+0.25)=0.9231\), placing far more weight on the experience. That conflicts with the variance interpretation: the example has substantial process variance relative to variation among risk means. Use EPV/VHM, giving the corrected estimate of \(\$870.00\). The common mistake this avoids is swapping the within-risk and between-risk variance roles.
Keep Exposure and Variance Units Consistent
The observation count \(n\) is not just a number of rows in a data file. It represents the amount of experience on the basis used to define EPV and VHM. For example, if \(K\) is expressed in exposure-years, \(n\) must also be in exposure-years. A set of twelve monthly records might represent one exposure-year, not twelve exposure-years.
When records have equal exposure, their average can be used directly. With unequal exposure, the experience summary should reflect exposure weights; the corresponding Bühlmann-Straub formulation uses total exposure and an exposure-weighted mean. Do not insert a raw record count into the equal-exposure formula when the records represent different amounts of exposure. The following example uses equal monthly exposure and converts the total to years.
Worked Example: Converting Monthly Records to Exposure-Years
A hypothetical insured has 24 equally exposed monthly observations, each representing \(1/12\) of an exposure-year. The annualized loss-cost rate averaged across the months is \(\$112\) per exposure-year. The collective rate is \(\$100\) per exposure-year. EPV and VHM imply \(K=5\) exposure-years. Find the estimate.
Method and unit conversion. Total exposure is \(24(1/12)=2\) exposure-years. Since \(K\) is in exposure-years, use \(n=2\), not 24.
Step 1: Calculate credibility. \(Z=2/(2+5)=2/7=0.2857\), to four decimals. The complement is \(1-Z=5/7=0.7143\).
Step 2: Blend the rates. The estimate is \((2/7)(112)+(5/7)(100)=224/7+500/7=724/7=\$103.43\) per exposure-year, rounded to cents.
Step 3: Check the wrong unit calculation. Using the raw count gives \(Z=24/(24+5)=24/29=0.8276\), which would produce \((24/29)(112)+(5/29)(100)=3{,}188/29=\$109.93\). This incorrectly treats the two years of exposure as 24 exposure-years. Equivalently, express \(K\) in months as \(60\) months: \(24/(24+60)=2/7\), the same correct weight. The common mistake this avoids is combining exposure measures on different time bases.
Use the Collective Complement—and Interpret \(Z\) Correctly
Credibility \(Z\) is the coefficient assigned to the experience average. Its complement \(1-Z\) is assigned to the collective mean. Since \(0\leq Z\leq1\) when \(n\geq0\), \(K>0\), and the model quantities are valid, the estimate is a weighted average of the two inputs. This does not make \(Z\) the probability that the risk’s true mean equals its observed experience, nor the probability of a claim or loss outcome.
Worked Example: Restoring the Omitted Collective Component
A hypothetical risk has an experience average of \(\$1{,}400\) over \(n=2\) observations, and the collective mean is \(\$1{,}000\). Suppose EPV is \(90{,}000\) dollars squared and VHM is \(30{,}000\) dollars squared. Calculate the estimate and explain why \(Z\) is not a probability statement.
Method. Use \(K=\text{EPV}/\text{VHM}\), calculate \(Z\), and apply both weights to their respective dollar amounts.
Step 1: Find \(K\) and \(Z\). \(K=90{,}000/30{,}000=3\). Then \(Z=2/(2+3)=0.4\), and \(1-Z=0.6\).
Step 2: Calculate the complete estimate. The estimate is \((0.4)(1{,}400)+(0.6)(1{,}000)=560+600=\$1{,}160\).
Wrong calculation and correction. Omitting the complement and reporting \((0.4)(1{,}400)=\$560\) does not blend experience with the collective mean; it discards the collective component. The correction is to include \((1-Z)(1{,}000)=\$600\), yielding \(\$1{,}160\). Also, \(Z=0.4\) says that the linear estimate assigns 40% weight to the experience average. It does not say there is a 40% probability that the risk’s expected value is \(\$1{,}400\). The common mistakes avoided are dropping the collective complement and interpreting a coefficient as an event probability.
A Practical Audit Sequence
A short, ordered review catches most calculation errors before an indication is used. It also separates model or unit errors from simple arithmetic errors.
Confirm that EPV is expected conditional variance and VHM is variance among conditional means.
Use \(K=\text{EPV}/\text{VHM}\). Confirm that the observation amount \(n\) and \(K\) use the same exposure and time basis.
Find \(Z=n/(n+K)\) and \(1-Z\), then verify that they add to one.
Multiply \(Z\) by the risk’s experience average and \(1-Z\) by the collective mean.
For valid positive quantities, the estimate should lie between the two input means. Describe \(Z\) as an estimator weight, not as a probability.
The between-input check is especially useful: if the experience average and collective mean are \(\$80\) and \(\$100\), a positive-weight estimate cannot be \(\$120\). Such a result signals a missing complement, a sign or arithmetic error, or an incorrectly formed input.
Common Exam Traps
- Reversing EPV and VHM. The standard factor uses EPV divided by VHM. Check that increasing process variation lowers credibility.
- Counting records instead of exposure. Twelve monthly rows may represent one year of exposure. Convert \(n\) or \(K\) so both use the same unit.
- Using only \(Z\bar X\). A Bühlmann estimate also includes \((1-Z)m\), unless that term is explicitly zero because the collective mean is zero.
- Calling \(Z\) a probability. It is a coefficient in the credibility estimator and does not, by itself, describe the probability of an event.
These checks are aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus. They are useful in practice as well: a precise indication is only as reliable as its variance definitions, exposure basis, and interpretation.
Check Your Understanding
For each problem, check the variance order, exposure basis, and both weights before giving the estimate.
- A risk has EPV \(=180\) and VHM \(=60\), with four observations, an experience average of \(\$75\), and collective mean of \(\$60\). Find \(K\), \(Z\), and the estimate. (Answer: \(K=3\), \(Z=4/7=0.5714\), estimate \(=\$68.57\).)
- EPV is \(1{,}600\) and VHM is \(400\), on an annual exposure basis. A risk has 18 months of experience, average rate \(\$90\), and collective rate \(\$80\). Find credibility and the estimate. (Answer: \(K=4\) years, \(n=1.5\) years, \(Z=3/11=0.2727\), estimate \(=\$82.73\).)
- A calculation uses EPV \(=240\), VHM \(=80\), and \(n=2.5\). The experience average is \(\$140\), and the collective mean is \(\$110\). State the correct \(K\), \(Z\), and estimate; identify the result if the ratio is reversed. (Answer: correct \(K=3\), \(Z=5/11=0.4545\), estimate \(=\$123.64\); reversed \(K=1/3\) gives \(Z=15/17=0.8824\), an incorrect estimate of \(\$136.47\).)
- Twelve quarterly records each represent \(0.25\) exposure-year. The risk’s exposure-weighted average rate is \(\$52\), the collective rate is \(\$40\), and \(K=6\) exposure-years. Find the estimate. (Answer: total exposure \(n=3\) years, \(Z=1/3\), estimate \(=\$44.00\).)
- Ten monthly records each represent one month. The experience average is \(\$126\), the collective mean is \(\$105\), EPV is \(500\), and VHM is \(100\), with both variance quantities defined per month. Find the estimate, and explain why using \(n=10\) with a \(K\) measured in years would be inconsistent. (Answer: \(K=5\) months, \(Z=10/15=0.6667\), estimate \(=\$119.00\); \(n\) and \(K\) must use the same time unit.)