Deriving The Covariance Diversification Adjustment

Why the Diversification Adjustment Needs a Derivation

A covariance formula combines standalone risk charges into an aggregate charge by recognizing that the risks need not reach their adverse outcomes together. The resulting aggregate is often below the sum of standalone charges. The difference is the diversification adjustment. Deriving the formula makes clear which inputs drive that difference and why a change in a correlation assumption can materially affect reported capital.

This tutorial focuses on the covariance aggregation method and its adjustment. It treats the standalone charges as inputs from the relevant risk categories; it does not determine how those charges were produced. The calculations use monetary amounts in thousands of dollars unless stated otherwise. No interest-rate, mortality, or specific loss-distribution assumption is needed for these covariance calculations: the model input is the set of charges and their correlation matrix.

Definition: Let $c_i\geq 0$ be the standalone capital charge for category $i$, and let $\rho_{ij}$ be the assumed correlation between categories $i$ and $j$. The covariance aggregate is $C=\sqrt{\sum_i c_i^2+2\sum_{i<j}\rho_{ij}c_ic_j}$. The diversification adjustment is $D=\sum_i c_i-C$.

The matrix $R$ whose diagonal entries are 1 and whose off-diagonal entries are $\rho_{ij}$ is the correlation matrix. In vector notation, with $c=(c_1,\ldots,c_n)^\mathsf{T}$, the aggregate is $C=\sqrt{c^\mathsf{T}Rc}$. The two expressions are the same: the diagonal terms give $c_i^2$, while each off-diagonal pair appears twice in the matrix product.

A prescribed regulatory covariance formula is an aggregation rule, not a guarantee that the actual portfolio loss distribution has a particular shape. The variance interpretation below explains the algebra. It should not be read as asserting that a regulatory capital charge is necessarily a standard deviation or that the aggregate is a VaR or TVaR.

Deriving the Covariance Formula

To see why the formula has this form, consider a covariance representation in which each category has a centered, unit-variance risk variable $Z_i$, with $\operatorname{Cov}(Z_i,Z_j)=\rho_{ij}$. Scale each variable by its category charge and form $W=\sum_i c_iZ_i$. The variance of a sum is the sum of the variances plus twice the sum of the pairwise covariances:

\begin{aligned} \operatorname{Var}(W) &=\sum_i c_i^2\operatorname{Var}(Z_i) +2\sum_{i<j}c_ic_j\operatorname{Cov}(Z_i,Z_j)\\ &=\sum_i c_i^2+2\sum_{i<j}\rho_{ij}c_ic_j. \end{aligned}

The square root of this variance is the standard-deviation-scale aggregate, which gives the covariance formula. A prescribed formula can use this same algebra as a capital aggregation convention even when its charges are not literal standard deviations. The derivation therefore justifies the structure of the formula; its regulatory interpretation still depends on the applicable capital framework.

List the standalone charges.
Use the category charges $c_i$ in consistent units, and check that they are nonnegative.

Apply the correlation assumptions.
Include each pairwise term $2\rho_{ij}c_ic_j$ once for each unordered pair $i<j$.

Take the square root and calculate the adjustment.
Set $C=\sqrt{c^\mathsf{T}Rc}$, then calculate $D=\sum_i c_i-C$. Keep unrounded values until the final reporting step.

For nonnegative charges and a valid correlation matrix, the adjustment cannot be negative under this formula. Since every correlation satisfies $\rho_{ij}\leq 1$,

\left(\sum_i c_i\right)^2-C^2 =2\sum_{i<j}(1-\rho_{ij})c_ic_j\geq 0.

Both $\sum_i c_i$ and $C$ are nonnegative, so this inequality implies $C\leq \sum_i c_i$, and hence $D\geq 0$. If every pair of positively charged categories has correlation 1, then $C=\sum_i c_i$ and there is no diversification adjustment. If charges are zero, their pairwise terms contribute zero.

Worked Calculations

Worked Example: Two Categories and the Adjustment

A hypothetical insurer has standalone charges of $18 thousand and $30 thousand for two categories. The prescribed correlation is 0.20. Calculate aggregate capital and the diversification adjustment. Use the covariance method; no distributional, interest, or mortality assumptions are required.

Method. For two categories, $C^2=c_1^2+c_2^2+2\rho_{12}c_1c_2$. The factor of 2 is essential because the covariance appears in both off-diagonal positions of $c^\mathsf{T}Rc$.

Calculate the squared aggregate. In thousands of dollars, $18^2=324$, $30^2=900$, and $2(0.20)(18)(30)=216$. Therefore, $C^2=324+900+216=1{,}440$.

Take the square root. $C=\sqrt{1{,}440}=37.9473319$ thousand dollars. The standalone total is $18+30=48$ thousand dollars, so $D=48-37.9473319=10.0526681$ thousand dollars.

Answer and check. Rounded to the nearest cent, aggregate capital is $37,947.33 and the adjustment is $10,052.67. They sum to $48,000.00, the standalone total. The common mistake avoided is omitting the factor of 2 from the covariance term.

Worked Example: Increasing One Category Charge

A hypothetical insurer has charges of $20 thousand and $30 thousand, with correlation 0.25. A revised exposure assessment increases the first charge to $25 thousand while leaving the second charge and the correlation unchanged. Calculate both aggregates and adjustments, and explain how they change.

Method. Recalculate the covariance formula with the revised charge. The inputs are charges in thousands of dollars and a fixed correlation; the example requires no additional distributional or discounting assumption.

Before the change. The squared aggregate is $20^2+30^2+2(0.25)(20)(30)=400+900+300=1{,}600$. Thus $C=\sqrt{1{,}600}=40$ thousand dollars. The standalone total is $50 thousand, so $D=50-40=10$ thousand dollars.

After the change. The squared aggregate is $25^2+30^2+2(0.25)(25)(30)=625+900+375=1{,}900$. Thus $C=\sqrt{1{,}900}=43.5889894$ thousand dollars. The revised standalone total is $55 thousand, so $D=55-43.5889894=11.4110106$ thousand dollars.

Answer and check. To the nearest cent, aggregate capital rises from $40,000.00 to $43,588.99, an increase of $3,588.99. The adjustment rises from $10,000.00 to $11,411.01, an increase of $1,411.01. Those increases sum to $5,000.00, the increase in the standalone total. The common mistake avoided is holding the standalone total fixed after a category charge changes.

Worked Example: Changing a Correlation Assumption

For two hypothetical risk categories, standalone charges are $30 thousand and $45 thousand. Calculate the aggregate and adjustment at correlation 0.50 and at correlation 0.20, then quantify the changes. Assume each correlation is an admissible two-category correlation; no distribution, interest, or mortality assumption is needed.

Method. Keep the charges fixed and recalculate the pairwise covariance term for each correlation. This isolates the effect of the correlation assumption.

At correlation 0.50. The squared aggregate is $30^2+45^2+2(0.50)(30)(45)=900+2{,}025+1{,}350=4{,}275$. Thus $C=\sqrt{4{,}275}=65.3834839$ thousand dollars. The standalone total is $75 thousand, so $D=75-65.3834839=9.6165161$ thousand dollars.

At correlation 0.20. The squared aggregate is $30^2+45^2+2(0.20)(30)(45)=900+2{,}025+540=3{,}465$. Thus $C=\sqrt{3{,}465}=58.8642453$ thousand dollars, and $D=75-58.8642453=16.1357547$ thousand dollars.

Answer and check. To the nearest cent, aggregate capital is $65,383.48 and the adjustment is $9,616.52 at correlation 0.50; at correlation 0.20, the corresponding amounts are $58,864.25 and $16,135.75. Lowering the correlation reduces the aggregate by $6,519.24 and increases the adjustment by $6,519.24, while the standalone total stays at $75,000.00. The rounded changes reconcile exactly. The common mistake avoided is changing a correlation without recalculating its full pairwise term $2\rho c_1c_2$.

Sensitivity and Validity of the Inputs

The formula also gives a direct sensitivity result. Write $Q=C^2$. Holding all charges fixed, changing a single correlation $\rho_{ij}$ changes $Q$ by $2c_ic_j$ per unit change in correlation. For $C>0$, differentiating $C=\sqrt{Q}$ gives:

\frac{\partial C}{\partial \rho_{ij}}=\frac{c_ic_j}{C}, \qquad \frac{\partial D}{\partial \rho_{ij}}=-\frac{c_ic_j}{C}.

Thus, for positive charges, increasing one correlation while holding the other inputs fixed increases aggregate capital and reduces the adjustment, provided the changed matrix remains valid. This is a local sensitivity: a large change in correlation should be recalculated with the full formula, not estimated by treating the derivative as constant.

For a change in charge $c_i$, again with $C>0$, differentiation gives $\partial C/\partial c_i=(c_i+\sum_{j\ne i}\rho_{ij}c_j)/C$, and therefore $\partial D/\partial c_i=1-\partial C/\partial c_i$. The aggregate response to a higher charge is not necessarily equal to the charge increase, because the category's covariance with the other risks also changes. The worked charge example shows this: a $5,000 standalone increase produces a smaller aggregate increase, with the remainder appearing as an increased adjustment.

Correlation inputs must form a valid correlation matrix: the matrix must be symmetric, have 1s on its diagonal, and be positive semidefinite. In particular, every principal minor must be nonnegative. For three categories, the determinant condition is $1+2\rho_{12}\rho_{13}\rho_{23}-(\rho_{12}^2+\rho_{13}^2+\rho_{23}^2)\geq 0$, in addition to the pairwise bounds $-1\leq\rho_{ij}\leq 1$. Pairwise bounds alone do not establish validity. For example, correlations $0.9,0.9,-0.9$ give determinant $1+2(0.9)(0.9)(-0.9)-(0.9^2+0.9^2+(-0.9)^2)=1-1.458-2.43=-2.888$. The matrix is not positive semidefinite, so it cannot represent a valid joint correlation structure, even though each pairwise value lies between $-1$ and 1.

Common Mistakes and Practical Interpretation

A frequent arithmetic trap is to omit the factor of 2 in a two-category calculation. For charges of 40 and 50 and correlation 0.10, the incorrect calculation $40^2+50^2+0.10(40)(50)=4{,}300$ leads to an aggregate of about 65.57 thousand dollars. It is wrong because it includes only half of the off-diagonal covariance contribution. The correct squared aggregate is $40^2+50^2+2(0.10)(40)(50)=1{,}600+2{,}500+400=4{,}500$, giving $C=\sqrt{4{,}500}=67.0820393$ thousand dollars. The adjustment is $90-67.0820393=22.9179607$ thousand dollars. To the nearest cent, these are $67,082.04 and $22,917.96.

Do not confuse aggregate capital with the adjustment. $C$ is the covariance aggregate; $D$ is the standalone total minus that aggregate.
Keep units consistent. If charges are entered in thousands, the square root is also in thousands. Convert to dollars only after the calculation.
Do not assume every set of pairwise correlations is feasible. Check the correlation matrix, not just each entry's range.
Do not interpret the adjustment as a cash amount automatically available to the insurer. It is the difference implied by the stated aggregation rule; capital fungibility and other prescribed adjustments are separate considerations.

Understanding this derivation is aligned with the CAS Exam 9 (Risk Management for Actuaries) syllabus. In practice, it helps explain whether an aggregate charge moved because a category input changed, because a dependency assumption changed, or because the inputs themselves were not a valid correlation structure.

Check Your Understanding

For each problem, calculate using the covariance formula and retain full precision until the final rounding. Charges are in thousands of dollars unless otherwise stated.

A hypothetical insurer has charges of 16 and 24, with correlation 0.50. Find the aggregate and diversification adjustment, to the nearest hundredth of a thousand dollars. (Answer: aggregate $34.87 thousand; adjustment $5.13 thousand.)
Two categories have charges of 25 and 35, with correlation 0.20. Find the aggregate and adjustment, to the nearest hundredth of a thousand dollars. (Answer: aggregate $46.90 thousand; adjustment $13.10 thousand.)
Three categories have charges of 10, 20, and 30. Their pairwise correlations are $\rho_{12}=0.20$, $\rho_{13}=0.10$, and $\rho_{23}=0.30$. Calculate the aggregate and adjustment to the nearest hundredth of a thousand dollars. (Answer: aggregate $43.59 thousand; adjustment $16.41 thousand.)
Two charges are 30 and 45. Calculate the aggregate at correlation 0.50 and at correlation 0.20, then state the aggregate decrease and adjustment increase, to the nearest hundredth of a thousand dollars. (Answer: at 0.50, aggregate $65.38 thousand and adjustment $9.62 thousand; at 0.20, aggregate $58.86 thousand and adjustment $16.14 thousand; aggregate decreases and adjustment increases by $6.52 thousand.)
A proposed three-category correlation matrix has $\rho_{12}=0.80$, $\rho_{13}=0.80$, and $\rho_{23}=-0.80$. Calculate its determinant using the three-category formula and decide whether it is a valid correlation matrix. (Answer: determinant $1-1.024-1.920=-1.944$; it is invalid because the determinant is negative.)

What You'll Learn