Why Hierarchical Aggregation Requires Care
An insurer may organize its risks by business unit, product, or risk type, but enterprise capital is ultimately measured for the combined loss. A natural calculation is therefore staged: aggregate risks within each business unit, then combine the business-unit totals. This can reproduce a direct enterprise-wide calculation, but only if the staged calculation retains the information the chosen measure requires.
The central distinction is between aggregating losses and aggregating summaries of losses. Adding random losses is associative: grouping the losses first does not change their total. In contrast, adding standalone capital amounts, or retaining only unit-level means and variances without the cross-unit dependence, generally does not preserve the enterprise result.
For variance calculations, assume all component losses have finite second moments and that the covariance information used is internally consistent. For distributional calculations, the joint distribution or an appropriate dependence model is needed. As in the earlier material on aligning loss bases before aggregation, losses cannot be combined meaningfully if their horizons, valuation bases, or definitions differ.
What Makes Staged Variance Aggregation Exact?
For two business units, the variance of the enterprise loss is the sum of the two unit variances plus twice their covariance. A covariance is positive when the units tend to have losses above their own means at the same time, negative when their deviations tend to offset, and zero when they are uncorrelated.
For a hierarchy with several units, the corresponding formula is
This staged formula is equivalent to calculating the variance from all component risks directly, provided that each unit variance includes the covariances among risks inside that unit and each cross-unit covariance includes all covariances between its component risks.
Write \(T=\sum_g T_g\), and subtract its mean: \(T-E[T]=\sum_g(T_g-E[T_g])\).
The square contains each squared unit deviation and two copies of each product of deviations from distinct units.
The squared terms have expectations \(\operatorname{Var}(T_g)\); the cross-products have expectations \(\operatorname{Cov}(T_g,T_h)\). This gives the staged variance formula.
Since \(T_g=\sum_jX_{g,j}\), \(\operatorname{Cov}(T_g,T_h)=\sum_j\sum_k\operatorname{Cov}(X_{g,j},X_{h,k})\). Thus no cross-unit covariance is lost if the full sum is retained.
The derivation also explains the main limitation: replacing a unit total with only its variance does not supply its covariance with another unit. The cross-unit terms must be calculated or estimated as well. Means, on the other hand, always add: \(E[T]=\sum_gE[T_g]\), assuming the expectations exist.
Worked Example: Checking a Two-Level Variance Calculation
Worked Example: Checking a Two-Level Variance Calculation
A hypothetical insurer models four component losses, in millions of dollars, over one year. Business unit A contains \(X_{A1}\) and \(X_{A2}\); unit B contains \(X_{B1}\) and \(X_{B2}\). The model is jointly normal, with the following construction using independent standard normal variables \(Z_1,Z_2,Z_3,Z_4\):
Calculate the mean and variance of each unit, their covariance, and the enterprise variance. Then verify the enterprise variance by expanding to the component level. The normal model is a mathematical approximation for aggregate loss; it can assign probability to negative losses, so it should not be interpreted as a literal nonnegative severity model.
Method. First obtain component variances and covariances from the independent-factor representation. Then aggregate within units and across units. Because the \(Z\)'s are independent with variance 1, a component's variance is the sum of its squared coefficients, and the covariance of two components is the sum of products of their coefficients on shared \(Z\)'s.
For unit A, the component means are 10 and 6, so \(E[T_A]=16\). The component variances are \(3^2=9\) and \(1^2+(\sqrt{15})^2=16\). Their covariance is \(3(1)=3\), from the shared \(Z_1\). Therefore,
For unit B, the component means are 8 and 4, so \(E[T_B]=12\). The variance of \(X_{B1}\) is \(1+16/15+344/15=25\). The variance of \(X_{B2}\) is \(4+36/15+148/5=36\). Their covariance is \(1(2)+(4/\sqrt{15})(6/\sqrt{15})=2+24/15=18/5=3.6\). Hence,
The cross-unit covariance is the sum of the four cross-unit component covariances: \(\operatorname{Cov}(X_{A1},X_{B1})=3\), \(\operatorname{Cov}(X_{A2},X_{B1})=1+4=5\), \(\operatorname{Cov}(X_{A1},X_{B2})=6\), and \(\operatorname{Cov}(X_{A2},X_{B2})=2+6=8\). Thus \(\operatorname{Cov}(T_A,T_B)=3+5+6+8=22\). The staged enterprise calculation is
For the direct check, the four component variances sum to \(9+16+25+36=86\). The distinct component-pair covariances sum to \(3+3.6+3+5+6+8=28.6\), where the first two terms are within-unit covariances and the last four are cross-unit covariances. Therefore, \(86+2(28.6)=143.2\), matching the staged result. The calculation is checked both by unit totals and by the full component expansion.
Common mistake avoided: Adding the unit variances alone gives \(31+68.2=99.2\), not the enterprise variance. That calculation omits \(2\operatorname{Cov}(T_A,T_B)=44\). The factor representation makes the missing dependence explicit.
Staging a Distributional Measure
Variance aggregation is exact when the relevant means and covariances are correctly carried through. A tail measure such as Value-at-Risk requires more. In general, the distribution of a sum depends on the joint distribution of the unit losses, not just their means and variances.
Under a joint normal model, a sum of unit losses is normal. Its mean and variance determine its distribution, so staged aggregation reproduces the direct enterprise VaR if the unit means, unit variances, and cross-unit covariance terms are all included. Outside such a model, a moment-based approximation may still be useful, but it is an approximation rather than an exact distributional aggregation.
Worked Example: Enterprise VaR From Correlated Unit Totals
Two hypothetical business units have annual losses \(U\sim N(30,64)\) and \(V\sim N(20,36)\), measured in dollars. Their correlation is \(0.25\), and their joint distribution is bivariate normal. Calculate the enterprise 95% VaR by staged aggregation, compare it with the sum of standalone 95% VaRs, and explain the difference.
Method. Calculate the covariance from the correlation and standard deviations, then combine means and variances. The standard deviations are \(\sqrt{64}=8\) and \(\sqrt{36}=6\), so \(\operatorname{Cov}(U,V)=0.25(8)(6)=12\).
The enterprise mean is \(30+20=50\). Its variance is \(64+36+2(12)=124\), and its standard deviation is \(\sqrt{124}=11.1355\) to four decimal places. Using \(z_{0.95}=1.6448536\), the enterprise VaR before final rounding is \(50+1.6448536\sqrt{124}=68.316\) approximately. Rounded to the nearest cent, the enterprise 95% VaR is $68.32.
For comparison, the standalone VaRs are \(30+1.6448536(8)=43.1588\) and \(20+1.6448536(6)=29.8691\). Their sum is approximately $73.03. It is not the enterprise VaR: the sum applies the same upper-tail quantile to both units as if their individual tail outcomes occurred together. The actual joint normal enterprise distribution has standard deviation \(\sqrt{124}\), which reflects the specified correlation.
Common mistake avoided: The incorrect calculation is to add the standalone VaRs and report $73.03 as enterprise VaR. The correction is to aggregate the losses' means and covariance structure first, then apply the normal quantile to the enterprise distribution. In this example the standalone sum is larger, but that observation is not a universal rule for all risk measures and distributions.
When Summaries Are Not Enough
A unit-level summary can conceal important dependence information. The following discrete example shows that even identical unit loss distributions can produce different enterprise VaR when their dependence differs. Define the quantile using the infimum convention above, which matters when a loss distribution has jumps.
Worked Example: Same Unit Marginals, Different Enterprise VaR
Two hypothetical business units each have a loss of $100 with probability 0.10 and a loss of $0 with probability 0.90. Compare the enterprise 95% VaR if the unit losses are independent with the result if they are perfectly positively dependent.
Independent case. The probability that both losses are zero is \(0.90(0.90)=0.81\). The probability of exactly one $100 loss is \(2(0.10)(0.90)=0.18\). The probability of two $100 losses is \(0.10(0.10)=0.01\). Thus, the enterprise loss is $0 with probability 0.81, $100 with probability 0.18, and $200 with probability 0.01. Its cumulative probability at $0 is 0.81, and at $100 it is \(0.81+0.18=0.99\). Therefore, the 95% VaR is $100.
Perfect positive dependence. The two losses occur together: the enterprise loss is $0 with probability 0.90 and $200 with probability 0.10. The cumulative probability at $0 is 0.90, which is below 0.95; at $200 it is 1. Therefore, the 95% VaR is $200.
Each unit has the same marginal mean, \(100(0.10)=10\), and variance \(100^2(0.10)(0.90)=900\) in both cases. Yet the enterprise VaRs differ by $100. The difference is dependence: the independent case has covariance 0, while identical perfectly dependent losses have covariance 900. Unit-level marginal summaries alone cannot identify the enterprise tail.
Common mistake avoided: Treating the same marginal distributions as sufficient to determine the aggregate distribution gives no basis for choosing between these two answers. Retain a joint distribution or dependence model when an exact distributional result is required.
Practical Tests for a Hierarchical Calculation
A staged calculation is most defensible when its intermediate quantities are sufficient for the enterprise measure being calculated. For the mean, unit means suffice. For variance, unit variances and every cross-unit covariance are needed. For an exact distributional measure, the staged method must preserve enough joint-distribution information to determine the distribution of the sum.
| Enterprise quantity | Information to retain at the unit level | When the staged result is exact |
|---|---|---|
| Mean | Each unit's mean | Expectations exist and all units use the same loss definition and horizon |
| Variance | Each unit's variance and every cross-unit covariance | Second moments exist and the covariance terms are correctly included |
| VaR from a joint normal model | Unit means, variances, and cross-unit covariances | The joint normal assumption is appropriate and the covariance structure is valid |
| Exact VaR for a general distribution | The joint distribution, or sufficient information to construct it | The dependence and aggregation model correctly determine the enterprise loss distribution |
When reviewing a result, check that unit totals partition the component losses without omission or duplication, that cross-unit covariance is present, and that the assumptions match the measure being reported. A valid covariance matrix is also required for a covariance-based model: impossible covariance inputs can produce misleading or even negative calculated variances. These checks support consistent enterprise aggregation aligned with the CAS Exam 9 (Risk Management for Actuaries) syllabus.
Check Your Understanding
For each problem, identify the information required by the stated measure and verify whether the calculation uses it.
- Two business units have means 12 and 9, variances 16 and 25, and covariance \(-4\). Find the enterprise mean, variance, and standard deviation. (Answer: Mean \(21\); variance \(33\); standard deviation \(\sqrt{33}\approx5.7446\).)
- Unit A contains \(X_1,X_2\), with variances 4 and 9 and covariance 1. Unit B contains \(X_3\), with variance 16. The cross-unit covariances are \(\operatorname{Cov}(X_1,X_3)=2\) and \(\operatorname{Cov}(X_2,X_3)=-1\). Find the unit variances, their covariance, and the enterprise variance. (Answer: \(\operatorname{Var}(T_A)=15\), \(\operatorname{Var}(T_B)=16\), \(\operatorname{Cov}(T_A,T_B)=1\), and \(\operatorname{Var}(T)=33\).)
- An enterprise loss is modeled as normal with mean $80 and variance \(400\). Calculate its 95% VaR using \(z_{0.95}=1.6448536\), rounded to the nearest cent. (Answer: $112.90.)
- Two units each lose $50 with probability 0.20 and $0 otherwise. Find enterprise 90% VaR under independence, then under perfect positive dependence. (Answer: Independence: $50; perfect positive dependence: $100.)
- Two unit totals have variances 49 and 64 and covariance 14. Find the enterprise variance and standard deviation. Also state the standard deviation that would result if the covariance were incorrectly omitted. Round standard deviations to four decimals. (Answer: Enterprise variance \(141\), standard deviation \(11.8743\); omitting covariance gives variance \(113\) and standard deviation \(10.6301\).)