Why Compare the Components?
Observed outcomes vary for more than one reason. Two insureds with the same underlying risk can have different outcomes because of random fluctuation; two insureds can also differ because their underlying expected outcomes are not the same. These sources of variation have different implications for interpreting a portfolio. A large spread in individual observations does not by itself show that risks have substantially different expected outcomes.
The Bühlmann model components put these sources on a common variance scale. The expected process variance describes variation around a risk’s own conditional mean. The variance of hypothetical means describes how much those conditional means differ across risks. Comparing them is a diagnostic: it helps indicate whether observed variation is mostly within-risk randomness or between-risk heterogeneity. The comparison does not, by itself, determine a premium or establish that a particular classification variable is useful.
For a concise comparison, define the process-to-heterogeneity ratio as \(r=\operatorname{EPV}/\operatorname{VHM}\), when \(\operatorname{VHM}>0\). If \(r>1\), expected within-risk process variation is larger than variation among hypothetical means. If \(r<1\), variation among hypothetical means is larger. The ratio is not a percentage and has no universal threshold at which a portfolio changes from “random” to “heterogeneous.” Its meaning depends on the model, the response, and the exposure basis.
A different but related summary is each component’s share of total variance. These shares add to one, whereas \(r\) compares one component directly with the other. Keep the two summaries distinct.
Why the Variance Decomposition Holds
The decomposition follows from the law of total variance. Conditional on a risk parameter, \(X\) has a conditional mean and a conditional variance. Across risks, both may vary. Taking the expected conditional variance accounts for fluctuation within a risk; taking the variance of the conditional mean accounts for differences in expected outcomes between risks.
\(\operatorname{Var}(X)=E[\operatorname{Var}(X\mid\Theta)]+\operatorname{Var}(E[X\mid\Theta])\).
The first term is EPV by definition; the second is VHM by definition.
Therefore, \(\operatorname{Var}(X)=\operatorname{EPV}+\operatorname{VHM}\), provided the stated variances exist and both components refer to the same \(X\).
This also clarifies why averages over repeated periods behave differently from individual observations. Suppose \(X_1,\ldots,X_n\) are conditionally independent and identically distributed given \(\Theta\), with the same conditional mean and variance as \(X\). For \(\bar X=(X_1+\cdots+X_n)/n\), the conditional variance is \(\operatorname{Var}(X\mid\Theta)/n\), while the conditional mean remains \(E[X\mid\Theta]\). Applying the same decomposition gives:
Thus, averaging more periods reduces the process contribution to the variability of a risk average, but does not remove variation in the hypothetical means. This is why a variance measured across observed risk averages should not automatically be interpreted as VHM: those averages still contain process noise.
Worked Examples
Worked Example: A Two-Class Portfolio
Consider a hypothetical portfolio with two equally likely risk types. Conditional on type A, annual claim counts have mean 10 and variance 9. Conditional on type B, they have mean 14 and variance 9. Calculate EPV, VHM, total variance, the process-to-heterogeneity ratio, and both component shares. Interpret the result.
Method and assumptions. Treat the two types as the full risk distribution, with probability \(1/2\) each. The conditional means and variances are on the same one-year exposure basis. The expected process variance is the probability-weighted average of the two conditional variances:
The portfolio mean is \((10+14)/2=12\). The deviations of the two type means from 12 are \(-2\) and \(2\), so:
The total variance is \(9+4=13\). The process-to-heterogeneity ratio is \(9/4=2.25\). The shares of total variance are \(9/13=0.6923\) for process variance and \(4/13=0.3077\) for hypothetical-mean variance, rounded to four decimal places.
Check and interpretation. The shares add to \(0.6923+0.3077=1.0000\), and the total variance agrees with the decomposition. Here, process variation is 2.25 times the variation among the two type means. This does not mean the types have identical expected outcomes; their means differ by 4. It means that, on the stated scale, within-type variance is the larger of these two variance components. Common mistake avoided: comparing the raw means 10 and 14 as though their difference alone measured the amount of between-risk variance.
Worked Example: A Poisson-Gamma Frequency Model
In a hypothetical frequency model, a risk has claim count \(N\) conditional on its claim rate \(\Theta\), with \(N\mid\Theta\sim\operatorname{Poisson}(\Theta)\) for one unit of exposure. Across risks, \(\Theta\) has a gamma distribution with shape 2 and rate 0.5. Calculate the two components, their ratio, and their shares of total variance.
Method and assumptions. Use the gamma shape-rate convention, under which \(E[\Theta]=\alpha/\beta\) and \(\operatorname{Var}(\Theta)=\alpha/\beta^2\). For a Poisson variable, the conditional mean and variance both equal its rate. Therefore:
Since \(\operatorname{Var}(N\mid\Theta)=\Theta\), the expected process variance is \(E[\Theta]=4\). Since \(E[N\mid\Theta]=\Theta\), the variance of hypothetical means is \(\operatorname{Var}(\Theta)=8\). Hence total variance is \(4+8=12\).
The process-to-heterogeneity ratio is \(4/8=0.5\). The process share is \(4/12=1/3=0.3333\), and the hypothetical-mean share is \(8/12=2/3=0.6667\), rounded to four decimal places. The shares sum to \(1.0000\).
Check and interpretation. In this model, the variance among risk-specific expected counts is twice the expected process variance. This comparison concerns one unit of exposure; changing the exposure basis changes the components and must be handled consistently. Common mistake avoided: reversing the ratio. The requested process-to-heterogeneity ratio is \(4/8=0.5\), not \(8/4=2\). The latter is the heterogeneity-to-process ratio, a different comparison.
Worked Example: Correcting Variation Across Risk Averages
A hypothetical balanced portfolio has three risks, each observed for two independent years on the same exposure basis. The claim counts are A: 8 and 12; B: 14 and 18; C: 20 and 24. Estimate EPV and VHM by the balanced method of moments, then calculate the process-to-heterogeneity ratio. Assume a common EPV across risks and years, independent conditional observations, and risks representative of the population of interest.
Step 1: Estimate EPV from within-risk variation. The risk averages are A: \((8+12)/2=10\), B: \((14+18)/2=16\), and C: \((20+24)/2=22\). For each risk, the within-risk sum of squared deviations is \(2^2+(-2)^2=8\). Across three risks, the total within-risk sum of squares is \(8+8+8=24\). There are \(3(2-1)=3\) within-risk degrees of freedom, so the pooled within-risk estimate is:
Step 2: Calculate the observed variance of risk averages. Their grand mean is \((10+16+22)/3=16\). Using the sample variance across the three averages:
Step 3: Remove the process noise in those averages. With two observations per risk, the process contribution to the variance of an average is \(\operatorname{EPV}/2\). Thus the method-of-moments estimate is:
The estimated total variance for a single-period outcome is \(8+32=40\), and the estimated process-to-heterogeneity ratio is \(8/32=0.25\). These estimates use a small hypothetical sample and therefore have sampling uncertainty; the calculations do not make the estimates exact population values.
Check and interpretation. The observed variance of risk averages, 36, exceeds the estimated VHM of 32 by \(8/2=4\), exactly the process contribution expected for two-period averages. Common mistake avoided: treating 36 as VHM without adjustment. Doing so would retain process noise in the between-risk component and overstate estimated heterogeneity.
Interpreting Ratios Carefully
A ratio is most useful when it answers a clearly stated comparison. With \(r=\operatorname{EPV}/\operatorname{VHM}\), a value above one indicates a larger expected process component; a value below one indicates a larger between-risk component. The ratio is unitless if both variances are expressed on the same scale. It does not imply that the larger component is more important for every business decision, nor does it identify the source of heterogeneity.
The component shares answer a different question: what fraction of total variance is assigned to each component? For example, if EPV is 4 and VHM is 8, then the process-to-heterogeneity ratio is \(0.5\), while the process share is \(4/(4+8)=1/3\). These values are not interchangeable. In particular, \(0.5\) is not the process share in this example.
The interpretation also depends on how risks and exposures are defined. Mixing groups with different expected outcomes can increase measured VHM; changing the mix can change the comparison even if each group’s conditional process behavior stays the same. A variance per policy, per vehicle-year, and per unit of payroll are not automatically comparable. Express the two components on the same response and exposure basis before calculating a ratio.
Finally, a high VHM relative to EPV is evidence of meaningful variation in the model’s hypothetical means, not proof that an available rating variable explains that variation. Conversely, a large EPV does not mean the underlying risk means are equal. The components summarize different sources of variation under the assumed model; they do not replace checks of data quality, model assumptions, or the risk population being studied.
Check Your Understanding
For each hypothetical case, calculate the requested comparison on the stated scale. Round ratios to four decimal places where needed.
- A model has EPV 6 and VHM 3. Find total variance, the process-to-heterogeneity ratio, and both shares of total variance. (Answer: total variance 9; ratio 2; process share \(0.6667\), VHM share \(0.3333\).)
- Two equally likely risk types have conditional means 5 and 11 and conditional variances 4 and 4. Find EPV, VHM, total variance, and the process-to-heterogeneity ratio. (Answer: EPV 4; VHM 9; total variance 13; ratio \(0.4444\).)
- For a hypothetical Poisson-Gamma model, \(N\mid\Theta\sim\operatorname{Poisson}(\Theta)\) and \(\Theta\) is gamma with shape 3 and rate 1.5. Find EPV, VHM, total variance, and the ratio. (Answer: EPV 2; VHM \(4/3\); total variance \(10/3\); ratio 1.5.)
- In a balanced study with four risks and four independent periods per risk, the pooled within-risk estimate is 5 and the sample variance of the four risk averages is 8. Use the method of moments to estimate VHM, then calculate estimated total variance and the process-to-heterogeneity ratio. (Answer: estimated VHM 6.75; estimated total variance 11.75; ratio \(0.7407\).)
- A portfolio has EPV 12 and VHM 6. A proposed summary reports “the process share is 2.” Identify the calculation error and give the correct process share and ratio. (Answer: 2 is the process-to-heterogeneity ratio, not a share; process share \(12/18=0.6667\), ratio 2.)