Choosing Exposure Units and Portfolio Granularity

Why Exposure Units and Granularity Matter

A capital model must represent what is exposed to loss. The choice of exposure unit determines what is counted or measured; the model’s granularity determines how finely those exposures remain distinguishable. A policy count, a count of insured locations, and a business unit’s total premium may all describe the same portfolio, but they do not necessarily preserve the same information.

A more detailed representation can distinguish risks with different characteristics and support decisions at a local level. It also requires more detailed, reliable data and can create additional modeling and governance work. A highly aggregated representation may be easier to maintain, but it can conceal concentrations, variation among risks, or shared drivers. The appropriate choice depends on what the model is intended to measure and how its results will be used.

Definition: An exposure unit is the unit used to count, measure, or classify risk-bearing items in a model, such as a policy, contract, insured location, vehicle-year, transaction, or business unit. Granularity is the level of detail at which the model retains information about those units and their risk characteristics. Aggregation combines records or results into broader groups; segmentation separates them into groups that retain relevant differences.

Choosing a Level of Representation

The word “exposure” can refer to either a count of risk-bearing items or a quantity used to scale risk, such as insured values or earned premium. Those measures are not interchangeable. A policy count may be suitable when policies have similar characteristics; insured value may be more informative when potential loss is related to the amount insured. The model should state what is being counted or measured and why it is relevant to the modeled loss.

Representation	Information typically retained	Typical trade-off
Policy level	Policy attributes and policy-level risk differences	Detailed results, with substantial record-level data and control needs
Contract level	Terms and characteristics attached to each contract	Useful where a policy contains multiple distinct contractual exposures; mapping can be complex
Business-unit level	Totals and characteristics organized by operating or reporting unit	Useful for accountability, but can mask variation among the underlying risks
Portfolio level	Overall totals or summary statistics for the whole modeled portfolio	Efficient for enterprise summaries, but weak for identifying local concentrations or differences

These levels are not always a simple hierarchy. One policy can cover several locations; one contract can include multiple coverages; and a business unit can contain many policies and contracts. A sound representation makes those relationships explicit enough to avoid counting the same exposure twice or omitting a material one.

Aggregation need not reduce fidelity when the underlying risks are sufficiently alike and the summary retains all information needed by the model. For example, under a model of identical independent risks, the number of risks and their common loss assumptions may be sufficient to calculate aggregate moments. But a single total, such as portfolio premium, generally cannot reveal how many risks generated it or whether they have different expected loss behavior.

Worked Examples

Worked Example: Homogeneous Policy-Level Exposures

A hypothetical insurer models 120 identical policy-year exposures for a fixed one-year period. Each policy has an independent probability of a claim of $1/60$. A claim produces a fixed payment of $10,000; otherwise, payment is zero. Calculate the expected aggregate payment, its variance and standard deviation, and the probability of at least one claim. Explain what information must be retained if the policies are represented as one portfolio total.

Method and assumptions. Let $I_j$ equal 1 if policy $j$ has a claim and 0 otherwise. The indicators are independent Bernoulli variables with $P(I_j=1)=1/60$. Aggregate payment is $L=10{,}000\sum_{j=1}^{120} I_j$.

Expected payment. The expected number of claims is $120(1/60)=2$. Therefore,

E[L]=10{,}000(2)=\$20{,}000.

Variance and standard deviation. For a Bernoulli variable with probability $p$, the variance is $p(1-p)$. Independence gives the claim-count variance $120(1/60)(59/60)=118/60=1.9666667$. Scaling a random variable by 10,000 multiplies its variance by $10{,}000^2$, so

\operatorname{Var}(L)=120\left(\frac{1}{60}\right)\left(\frac{59}{60}\right)(10{,}000)^2 =196{,}666{,}666.67\ \text{dollars}^2, \qquad \operatorname{SD}(L)=\sqrt{196{,}666{,}666.67} =\$14{,}023.79.

The variance is shown to the nearest cent in dollars squared; the standard deviation is rounded to two decimals. In particular, the square root rounds to $14,023.79, not $14,023.80.

Probability of at least one claim. The probability of no claims is $(59/60)^{120}$, because all 120 policies must have no claim. Thus $P(L>0)=1-(59/60)^{120}=0.8669$, to four decimals.

A portfolio-level representation can preserve these results if it retains the count of 120 and the common claim probability and payment assumptions. If it retains only an unlabeled total with no count or risk assumptions, it cannot support the calculation. Common mistake avoided: treating an aggregate label as though it automatically contains the underlying exposure count and assumptions.

Worked Example: Why Risk Segmentation Can Change Variance

A hypothetical book has 80 policy-year exposures in a lower-risk segment and 40 in a higher-risk segment. Claim indicators are independent within and between segments. Each lower-risk policy has a claim probability of 0.01; each higher-risk policy has a probability of 0.04. Every claim produces a fixed payment of $5,000. Compare the expected payment and variance using the two segments with those from a pooled model that assigns one common probability to all 120 policies.

Method and assumptions. The payment per claim is fixed, the period is the same for all policies, and the claim probabilities are treated as known. For a sum of independent Bernoulli indicators, expected claim count is the sum of their probabilities and its variance is the sum of $p(1-p)$.

Segmented calculation. Expected claim count is $80(0.01)+40(0.04)=0.8+1.6=2.4$. Expected payment is $2.4(\$5{,}000)=\$12{,}000$. Claim-count variance is

$$ 80(0.01)(0.99)+40(0.04)(0.96) =0.792+1.536 =2.328. $$

Therefore, payment variance is $2.328(\$5{,}000)^2=\$58{,}200{,}000$, and the standard deviation is $\sqrt{\$58{,}200{,}000}=\$7{,}628.89$, rounded to two decimals.

Pooled calculation. The probability-weighted average claim probability is $(80(0.01)+40(0.04))/120=2.4/120=0.02$. If the pooled model treats all policies as having this common probability, its claim-count variance is $120(0.02)(0.98)=2.352$. Payment variance is $2.352(\$5{,}000)^2=\$58{,}800{,}000$, and standard deviation is $\$7{,}668.12$.

The expected payment is the same in both models, but the pooled model’s variance is higher by $\$600{,}000$. Segmentation retains the different probabilities; pooling substitutes an average probability and changes the modeled distribution. Common mistake avoided: using the correct pooled mean probability and assuming that this also preserves the variance.

Aggregation, Segmentation, and Model Fidelity

The difference in the preceding example follows from the variance formula. For a sum of random losses $L_j$,

\operatorname{Var}\left(\sum_{j=1}^{n}L_j\right) =\sum_{j=1}^{n}\operatorname{Var}(L_j) +2\sum_{j<k}\operatorname{Cov}(L_j,L_k).

If losses are independent, every covariance term is zero. If the risks have different probabilities or severities, their individual variances still need not equal the variance implied by a pooled average. If losses share drivers, the covariance terms may also matter. This is why retaining the count of exposures alone may be sufficient in a homogeneous model but insufficient when risk characteristics differ or move together.

The practical question is not whether the most detailed data can be stored. It is whether the distinctions removed by aggregation could materially change the modeled result or a decision. A grouping is more defensible when the risks within it have similar relevant characteristics, the group’s exposure measure is appropriate, and the model retains enough information to reproduce its intended calculations.

Derivation: For any random variables $L_1,\ldots,L_n$, expand the variance of their sum using $\operatorname{Var}(Z)=E[Z^2]-E[Z]^2$. The cross terms in $E[(\sum_j L_j)^2]$ are $E[L_jL_k]$. Subtracting $E[\sum_j L_j]^2$, whose cross terms are $E[L_j]E[L_k]$, leaves $E[L_jL_k]-E[L_j]E[L_k]=\operatorname{Cov}(L_j,L_k)$ for each pair. Each pair appears twice, giving the displayed formula. Independence makes each covariance zero, but aggregation alone does not establish independence.

Worked Example: Business-Unit Totals with a Shared Event

A hypothetical insurer reports losses for two business units. A shared event occurs with probability 0.02 and causes a $1,000,000 loss in each unit. Separately, each unit has its own independent event with probability 0.10, causing a $100,000 loss in that unit. The shared-event indicator and both unit-specific indicators are mutually independent. Calculate the variance of total loss and compare it with the result from incorrectly treating the two business-unit losses as independent.

Method and assumptions. Let $C$ be the shared-event indicator, and $I_1,I_2$ the unit-specific indicators. Then $L_1=1{,}000{,}000C+100{,}000I_1$, $L_2=1{,}000{,}000C+100{,}000I_2$, and total loss is $L_1+L_2=2{,}000{,}000C+100{,}000(I_1+I_2)$.

Correct portfolio variance. The shared component has variance $(\$2{,}000{,}000)^2(0.02)(0.98)=\$78{,}400{,}000{,}000$. Each unit-specific component has variance $(\$100{,}000)^2(0.10)(0.90)=\$900{,}000{,}000$. The three indicators are independent, so the components’ variances add:

\operatorname{Var}(L_1+L_2) =78{,}400{,}000{,}000+900{,}000{,}000+900{,}000{,}000 =\$80{,}200{,}000{,}000.

Incorrect independent-unit calculation. The variance of each unit is $(\$1{,}000{,}000)^2(0.02)(0.98)+\$900{,}000{,}000=\$20{,}500{,}000{,}000$. Adding the two unit variances as if the units were independent gives $\$41{,}000{,}000{,}000$, which understates the correct portfolio variance by $\$39{,}200{,}000{,}000$.

The common event is counted once in each unit’s reported loss, so the unit results are not independent. Keeping business-unit detail supports attribution, while retaining the shared event in the portfolio representation avoids losing the common exposure. Common mistake avoided: summing unit variances without checking whether the units have a shared loss driver.

Selecting Granularity for Use and Control

A useful design balances three considerations:

Fidelity. Retain distinctions that could materially affect loss estimates, concentrations, or comparisons. Examples include location, contract terms, risk class, and business unit when those features are relevant to the model.
Data and control burden. More granular records require reliable identifiers, consistent definitions, appropriate reconciliation to totals, and controls against missing or duplicated exposures. Detailed data that are inaccurate can create false precision.
Decision usefulness. Results should be available at a level that supports their intended use. Portfolio totals support an enterprise view; business-unit or contract detail may be needed for accountability, pricing, or concentration review.

When an aggregate model is used, document the information summarized, the reason the grouping is reasonable, and the limitations of the result. When detailed records are used, check that they reconcile to the relevant portfolio totals and that the modeled exposure unit matches the source data’s meaning. A policy count should not silently stand in for a location count, for example, if some policies cover multiple locations.

Common Exam Trap

A frequent error is to assume that equal total exposure implies equal risk. In the segmentation example, both representations had an expected claim count of 2.4, but their variances differed because the individual claim probabilities differed. Another error is to assume business units are independent merely because they are reported separately. A shared event can make their losses dependent even when their unit-specific events are independent.

Key takeaway: Choose an exposure unit that corresponds to the risk being modeled, then retain enough granularity to preserve material differences and shared drivers. Aggregate only when the information lost is not needed for the model’s purpose or the decisions it supports.

Check Your Understanding

For each problem, state the exposure representation or calculation needed and keep money in dollars. Probabilities are rounded to four decimals where requested; standard deviations are rounded to two decimals.

A hypothetical portfolio has 75 independent, identical policy-year exposures. Each has claim probability 0.02 and a fixed claim payment of $4,000. Find expected aggregate payment, payment variance, and standard deviation. (Answer: Expected payment $6,000; variance $23,520,000; standard deviation $4,849.74.)
A portfolio has 50 independent exposures with claim probability 0.01 and 30 with probability 0.05. Each claim pays $2,000. Find the expected payment and variance using the two segments. (Answer: Expected payment $4,000; variance $7,680,000.)
A portfolio has 100 independent exposures in each of two segments, with claim probabilities 0.01 and 0.03, respectively. Each claim pays $5,000. Compare the exact segmented payment variance with the variance from pooling all 200 exposures at the probability-weighted average rate. (Answer: Segmented variance $97,500,000; pooled variance $98,000,000; pooling is higher by $500,000.)
Two business units each incur a $300,000 loss if the same shared event occurs, with probability 0.01. Each unit also has an independent unit-specific event with probability 0.10 causing a $20,000 loss. The shared and unit-specific indicators are mutually independent. Find the correct variance of total loss and the variance obtained by incorrectly treating unit losses as independent. (Answer: Correct variance $3,636,000,000; independence calculation $1,854,000,000; understatement $1,782,000,000.)
A hypothetical portfolio has 40 contracts, each covering two separately identified locations. If the model’s loss event applies independently to each location, state the exposure count that should be retained for that event and identify one limitation of storing only the contract count. (Answer: Retain 80 location exposures; a contract-only count does not show the number of locations and may conceal location-level risk differences.)

What You'll Learn