Why Exposure Units Matter
A claim count and a claim rate describe related experience, but their variances are not in the same units. A count generally becomes more variable as the amount of exposure increases. Dividing that count by exposure changes both its scale and its process variance. The variance of hypothetical means changes differently: it reflects differences among risks, so its scaling depends on how the risk mean itself is expressed.
This distinction matters when estimating Bühlmann model components from observations with different exposure amounts, or when reporting components per policy-year, per vehicle-year, or per thousand miles. A numerical comparison is meaningful only when the observations and both variance components use consistent units. The formulas below apply under stated assumptions of stable risk characteristics and additive experience across exposure. This treatment is aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus.
The Scaling Rule
Suppose experience is generated by stable, additive exposure units. Conditional on \(\Theta\), each unit has mean \(m(\Theta)\) and variance \(q(\Theta)\), and the units are conditionally independent. For an integer number \(e\) of such units, the sum has conditional mean \(e\,m(\Theta)\) and conditional variance \(e\,q(\Theta)\). The same relationships extend to continuous exposure under a process model with independent increments, such as a Poisson process.
Applying the total variance decomposition gives the two components for the aggregate observation. If instead the observation is the per-exposure rate \(R_e=X_e/e\), its conditional mean is \(m(\Theta)\), while its conditional variance is divided by \(e^2\). Thus the VHM is unchanged for the rate, but the process variance is divided by \(e\).
Here EPV denotes the expected process variance. The aggregate’s VHM grows as \(e^2\) because its conditional mean is multiplied by \(e\). The rate’s VHM does not change because dividing the aggregate by \(e\) restores the original conditional mean \(m(\Theta)\). This is a direct consequence of \(\operatorname{Var}(cX)=c^2\operatorname{Var}(X)\), applied conditionally and across risks.
Worked Examples
Worked Example: Poisson Claim Counts and Claim Rates
A hypothetical portfolio has a claim frequency parameter \(\Lambda\) per policy-year with \(E[\Lambda]=0.08\) and \(\operatorname{Var}(\Lambda)=0.0016\). For a risk observed over \(e=5\) policy-years, assume \(N\mid\Lambda\sim\operatorname{Poisson}(5\Lambda)\). Find the process variance, VHM, and total variance for both the claim count \(N\) and the annualized rate \(R=N/5\). Use the Poisson frequency model and the total variance decomposition.
Count mean and process variance. Conditional on \(\Lambda\), the Poisson mean and variance are both \(5\Lambda\). Therefore the expected process variance is \(5E[\Lambda]=5(0.08)=0.40\) claims squared. The expected count is also \(5E[\Lambda]=0.40\).
Count VHM. The conditional mean count is \(5\Lambda\), so its variance across risks is \(5^2\operatorname{Var}(\Lambda)=25(0.0016)=0.0400\) claims squared.
Total count variance. The total variance is \(0.40+0.0400=0.4400\) claims squared.
Annualized rate components. Since \(R=N/5\), each variance component for \(R\) is the corresponding count component divided by \(5^2=25\). The process variance is \(0.40/25=0.0160\) claims-squared per policy-year squared. The VHM is \(0.0400/25=0.0016\) in the same rate-squared units. The total variance is \(0.0160+0.0016=0.0176\).
Check and common mistake. The rate mean is \(0.40/5=0.08\), matching \(E[\Lambda]\). A common error is to divide the count process variance by 5, not 25, when forming the variance of \(N/5\). Variance scales by the square of the multiplier, so division by 5 requires division by 25. The component totals also check: \(0.4400/25=0.0176\).
Worked Example: Aggregate Loss and Loss per Exposure
For a hypothetical risk, claim counts follow a Poisson process with rate \(\Lambda\) per exposure unit, where \(E[\Lambda]=0.02\) and \(\operatorname{Var}(\Lambda)=0.0001\). Each claim independently has severity \(Y\), equal to $1,000 with probability 0.75 and $3,000 with probability 0.25. For \(e=10\) exposure units, find the process variance, VHM, and total variance of aggregate loss \(S_e\), then find the corresponding components for loss per exposure \(S_e/10\). Assume severity is independent of frequency and does not vary by risk.
Severity moments. The mean severity is \(E[Y]=0.75(1{,}000)+0.25(3{,}000)=1{,}500\) dollars. The second moment is \(E[Y^2]=0.75(1{,}000^2)+0.25(3{,}000^2)=750{,}000+2{,}250{,}000=3{,}000{,}000\) dollars squared.
Aggregate-loss process variance. In a compound Poisson model, conditional on \(\Lambda\), the aggregate mean is \(10\Lambda E[Y]\) and the aggregate variance is \(10\Lambda E[Y^2]\). Averaging the conditional variance gives \(10(0.02)(3{,}000{,}000)=600{,}000\) dollars squared.
Aggregate-loss VHM. The conditional mean is \(10\Lambda(1{,}500)=15{,}000\Lambda\). Its variance across risks is \(15{,}000^2(0.0001)=225{,}000{,}000(0.0001)=22{,}500\) dollars squared. The expected aggregate loss is \(10(0.02)(1{,}500)=$300\).
Total and per-exposure components. Total aggregate-loss variance is \(600{,}000+22{,}500=622{,}500\) dollars squared. Dividing aggregate loss by 10 divides each variance component by \(10^2=100\). For loss per exposure, the process variance is $6,000 squared, the VHM is $225 squared, and the total is $6,225 squared.
Check and common mistake. The loss-per-exposure mean is $300/10=$30. A common error is to use \(E[Y]^2=1{,}500^2=2{,}250{,}000\) in the compound Poisson process variance. That omits the variability in claim severity; the required second moment is \(E[Y^2]=3{,}000{,}000\). The aggregate components divided by 100 give $6,000 and $225, as calculated.
Worked Example: Converting from Vehicle-Years to Thousands of Vehicle-Years
A hypothetical risk has exposure of 6,000 vehicle-years and an observed count of 18 claims. For the component calculation, assume its underlying frequency parameter \(\Lambda\), measured per vehicle-year, has mean 0.003 and variance \(0.00000144\). Counts are conditionally Poisson with mean equal to exposure times \(\Lambda\). Calculate the count components and the components of the rate per 1,000 vehicle-years. Verify that the result is consistent with expressing frequency per vehicle-year.
Count components in vehicle-year units. For \(e=6{,}000\), expected count and process variance are both \(6{,}000(0.003)=18\). The count VHM is \(6{,}000^2(0.00000144)=36{,}000{,}000(0.00000144)=51.84\) claims squared. Total count variance is \(18+51.84=69.84\) claims squared. The observed count of 18 agrees with the model’s expected count, though the component calculation uses the stated risk-parameter distribution.
Rate per 1,000 vehicle-years. Define the rate observation as \(R=N/6\), because the exposure is six units of 1,000 vehicle-years. Expressing the risk parameter in this unit gives mean \(1{,}000(0.003)=3\) claims per 1,000 vehicle-years and variance \(1{,}000^2(0.00000144)=1.44\). For six such units, the rate process variance is \(3/6=0.5\) rate-squared, and its VHM is \(1.44\). The total is \(0.5+1.44=1.94\) rate-squared.
Verification in the original unit. The observed rate is \(18/6=3\) claims per 1,000 vehicle-years, or \(18/6{,}000=0.003\) claims per vehicle-year. For the rate per vehicle-year, process variance is \(0.003/6{,}000=0.0000005\), VHM is \(0.00000144\), and total variance is \(0.00000194\). Multiplying each of these rate-per-vehicle-year variances by \(1{,}000^2\) gives \(0.5\), \(1.44\), and \(1.94\), respectively.
Check and common mistake. The per-1,000 rate components agree under unit conversion. A common error is to multiply the mean frequency by 1,000 but leave its variance unchanged. A rate multiplied by 1,000 has variance multiplied by \(1{,}000^2\), which is essential for the VHM conversion.
Choosing and Checking the Units
Before estimating or comparing components, specify the observation and its exposure denominator. “Claim variance” is incomplete: it could mean variance of a count over a policy-year, a count over ten policy-years, or a rate per thousand vehicle-years. State the units of both components and of the total variance.
| Observation | Conditional mean under additive exposure | Process variance scaling | VHM scaling |
|---|---|---|---|
| Aggregate count or loss over exposure \(e\) | Proportional to \(e\) | Proportional to \(e\), when increments are conditionally independent | Proportional to \(e^2\) |
| Count or loss per exposure unit | Independent of \(e\) | Proportional to \(1/e\), under the same process assumption | Independent of \(e\), for a stable risk mean per exposure |
| Rate after changing the exposure unit by factor \(c\) | Multiplied by \(c\) | Multiplied by \(c^2\) for the rate’s variance components | Multiplied by \(c^2\) |
The last row concerns a change in how the rate is reported, such as claims per vehicle-year to claims per 1,000 vehicle-years. The rate itself is multiplied by 1,000, so its variance components are multiplied by \(1{,}000^2\). This is distinct from changing the amount of exposure observed for a risk. Always identify which operation is being performed.
Decide whether the variable is a count, aggregate loss, count rate, or loss rate.
For example, use policy-years or thousands of vehicle-years, not an unspecified “year” or “exposure.”
For additive Poisson or compound Poisson experience, the aggregate process variance is proportional to exposure, while the rate process variance is inversely proportional to exposure.
When multiplying a rate by a unit-conversion factor, multiply its variance components by the square of that factor.
Limits of the Rule and Exam Traps
The exposure formulas are not a universal property of every variable called a rate. For example, if \(n\) independent trials each have the same conditional success probability \(p\), the proportion of successes has conditional variance \(p(1-p)/n\). This has the same inverse-trial-count pattern, but it follows from a binomial model, not a Poisson assumption. If trials are dependent, or if the probability changes across trials, the simple scaling may fail.
Likewise, changing exposure while the underlying conditions change is not merely a unit conversion. A longer observation period may contain trend, seasonality, changing limits, or a different mix of covered risks. In those settings, a single stable per-exposure mean may not describe the whole period. A valid calculation must match the model and the exposure definition.
- Do not scale variance like a mean. Dividing a random variable by \(e\) divides its variance by \(e^2\), not by \(e\).
- Do not mix units. A VHM per policy-year rate cannot be added directly to a process variance for a count.
- Do not omit severity variation. Compound Poisson process variance uses \(E[Y^2]\), not just \(E[Y]^2\).
- Do not assume exposure alone guarantees scaling. Conditional independence, stability, and the model’s variance structure are essential.
Check Your Understanding
For each problem, keep the observation and exposure units explicit.
- A hypothetical frequency parameter has mean 0.12 and variance 0.0036 per policy-year squared. Counts are conditionally Poisson over four policy-years. Find the count’s process variance, VHM, and total variance. (Answer: 0.48, 0.0576, and 0.5376 claims squared.)
- Claim frequency has mean 0.01 and variance 0.000025 per exposure-unit squared. Severity is $500 with probability 0.6 and $1,500 with probability 0.4. For eight exposure units, find the aggregate-loss process variance and VHM. (Answer: $84,000 squared and $1,296 squared, respectively.)
- A hypothetical risk has frequency mean 0.004 and variance 0.00000025 per vehicle-year squared, with exposure of 2,400 vehicle-years. Under a Poisson model, find count process variance and VHM. Then express the rate components per 1,000 vehicle-years. (Answer: count components 9.6 and 1.44 claims squared; rate components 1.6667 and 0.25 rate-squared.)
- For a Poisson count rate with fixed frequency parameter 0.12 per exposure unit, compare the process variance of the rate observed over \(e=2\) and \(e=8\). (Answer: 0.06 and 0.015 rate-squared, respectively.)
- A risk’s probability parameter \(P\) has \(E[P]=0.20\) and \(\operatorname{Var}(P)=0.01\). Conditional on \(P\), 50 independent Bernoulli trials are observed. For the observed proportion, find the expected process variance and VHM. Use \(E[P(1-P)]=E[P](1-E[P])-\operatorname{Var}(P)\). (Answer: process variance 0.003 and VHM 0.01 proportion-squared.)