When Rating Models Limit Diversification Benefits

Why Modelled Diversification May Not Be Fully Recognized

Combining risk charges can produce a lower capital estimate than adding the charges individually. That reduction is modelled diversification credit. A rating assessment may nevertheless recognize only part of it. The distinction matters: an aggregation formula gives a result conditional on its assumptions, while a rating judgment also considers whether the assumptions and the insurer’s risk profile support that result.

A concentrated portfolio, risks driven by the same event, or limited evidence about dependence can make a large modelled benefit difficult to rely on. This tutorial develops a disciplined way to compare the modelled result with a deliberately limited recognition of that result. The calculations are hypothetical illustrations, not descriptions of any particular rating agency’s rules. The discussion is aligned with the CAS Exam 9 (Risk Management for Actuaries) syllabus.

Definition: Let $C_i$ be the capital charge for risk category $i$, and let $C_{\mathrm{agg}}$ be the charge after aggregation. The modelled diversification credit is $D_{\mathrm{model}}=\sum_i C_i-C_{\mathrm{agg}}$, when this difference is nonnegative. Recognized diversification credit is the amount allowed in the assessment after any stated limits or judgmental restrictions. It need not equal $D_{\mathrm{model}}$.

A limited-recognition rule can be represented for teaching purposes by $D_{\mathrm{rec}}=\min(\alpha D_{\mathrm{model}},D_{\max})$, where $0\leq\alpha\leq1$ is the recognized fraction and $D_{\max}\geq0$ is a dollar cap. The corresponding capital after this adjustment is $\sum_i C_i-D_{\mathrm{rec}}$. This is a generic illustrative rule, not a universal rating formula. A real assessment may use different procedures, qualitative judgments, or adjustments that cannot be summarized by this expression.

What the Correlation Calculation Does—and Does Not—Establish

One common aggregation structure is the square-root formula. To interpret it, suppose each $C_i$ is proportional to the standard deviation of a centered loss measure over the same horizon, with a common proportionality basis. Let $\rho_{ij}$ be the correlation between the loss measures for categories $i$ and $j$. If the correlation matrix is valid and the model uses a variance-based aggregation, then

C_{\mathrm{agg}} = \sqrt{\sum_i C_i^2+2\sum_{i<j}\rho_{ij}C_iC_j}.

For two risks, this becomes $C_{\mathrm{agg}}=\sqrt{C_1^2+C_2^2+2\rho C_1C_2}$. This formula is not automatically appropriate for every rating charge: charges may be calibrated measures rather than standard deviations, and the formula depends on its specified model and assumptions. A prescribed aggregation rule should be used as given, not interpreted as a universal law of insurer losses.

The reason for the formula is the variance of a sum. If centered losses $L_i$ have standard deviations $\sigma_i$, then $\operatorname{Var}(\sum_i L_i)=\sum_i\operatorname{Var}(L_i)+2\sum_{i<j}\operatorname{Cov}(L_i,L_j)$. Using $\operatorname{Cov}(L_i,L_j)=\rho_{ij}\sigma_i\sigma_j$, and charges proportional to the corresponding standard deviations, gives the square-root aggregation expression. The correlation terms matter because losses that tend to rise together leave less scope for offsetting outcomes.

Calculate the modelled aggregate.
Apply the stated aggregation rule using its specified charges and dependence assumptions.

Find the modelled credit.
Subtract the aggregate from the sum of standalone charges. Check that the result is nonnegative before treating it as a diversification benefit.

Assess whether that credit is supportable.
Consider concentration, common drivers, and the quality of evidence for the dependence assumptions.

Apply the stated recognition limit.
Use only an explicit cap, fraction, or other stated adjustment; do not imply that a hypothetical limit is an industry-wide rule.

Worked Examples

Worked Example: High Correlation Leaves Little Modelled Credit

A hypothetical insurer has two risk charges, $C_1=\$60.00$ million and $C_2=\$80.00$ million. For this example, assume the charges are standard-deviation-based on the same horizon and the stipulated correlation is $\rho=0.75$. First calculate the modelled credit. Then suppose a hypothetical review permits no more than $\$5.00$ million of credit.

Method. Use the two-risk square-root formula, then compare the resulting credit with the stated cap. The example assumes that the supplied correlation and aggregation structure are valid for the exercise; it does not establish that they are suitable for any actual portfolio.

The standalone total is $60.00+80.00=\$140.00$ million. The cross term is $2(0.75)(60.00)(80.00)=\$7{,}200.00$ million squared. Therefore,

C_{\mathrm{agg}} = \sqrt{60.00^2+80.00^2+7{,}200.00} = \sqrt{17{,}200.00} = \$131.15\text{ million}.

The modelled credit is $140.00-131.15=\$8.85$ million, using unrounded aggregate capital for the subtraction. The recognition rule is a maximum of $\$5.00$ million, so recognized credit is $\min(8.85,5.00)=\$5.00$ million. Capital after the limited credit is $140.00-5.00=\$135.00$ million. The calculation checks: the recognized credit is below the modelled credit, and the resulting capital lies between the aggregate charge and the standalone total.

Common mistake avoided: treating the modelled result as the amount automatically recognized. The cap is an additional condition in the problem, so the answer is not $\$131.15$ million of capital after recognition.

Worked Example: Weak Evidence Restricts a Calculated Benefit

A hypothetical insurer has two risk charges of $\$50.00$ million and $\$30.00$ million. A provisional model uses $\rho=0.20$. Because dependence evidence is limited, assume the review recognizes only $40\%$ of modelled diversification credit, with no separate dollar cap. Calculate modelled credit, recognized credit, and capital after recognition.

Method. Calculate the aggregate and credit under the provisional correlation assumption, then multiply the credit—not the aggregate capital—by the stated recognition fraction.

The standalone total is $50.00+30.00=\$80.00$ million. The cross term is $2(0.20)(50.00)(30.00)=\$600.00$ million squared. Thus,

C_{\mathrm{agg}} = \sqrt{50.00^2+30.00^2+600.00} = \sqrt{4{,}000.00} = \$63.25\text{ million}.

Using the unrounded square root, modelled credit is $80.00-\sqrt{4{,}000.00}=\$16.75$ million. Recognized credit is $0.40(80.00-\sqrt{4{,}000.00})=\$6.70$ million. Capital after recognition is $80.00-0.40(80.00-\sqrt{4{,}000.00})=\$73.30$ million. Values are rounded to two decimals in millions only at the final step. As a check, recognized credit is $40\%$ of the modelled credit, and capital after recognition exceeds the modelled aggregate.

The evidence limitation is not a claim that the true correlation must be higher. It is a reason not to rely fully on a benefit whose key assumption is poorly supported. The recognition fraction is an explicit hypothetical rule for this calculation, not a standard prescribed percentage.

Worked Example: Concentrated Exposures and a Credit Cap

A hypothetical insurer has three risk-category charges, each $\$25.00$ million. Assume the stipulated pairwise correlations are all $0.20$, and that the same variance-based assumptions used above apply. The portfolio also has a concentration concern. For this illustration, the review allows at most $40\%$ of modelled credit. Find the modelled aggregate, the allowed credit, and capital after recognition.

Method. Include all three pairwise covariance terms. Then apply the stated percentage limit. The fact that charges are equal does not remove any of the three dependence terms.

The standalone total is $3(25.00)=\$75.00$ million. The sum of squared charges is $3(25.00^2)=\$1{,}875.00$ million squared. There are three pairs, and each pair contributes $2(0.20)(25.00)(25.00)=\$250.00$ million squared, so all cross terms total $3(250.00)=\$750.00$ million squared. Therefore,

C_{\mathrm{agg}} = \sqrt{1{,}875.00+750.00} = \sqrt{2{,}625.00} = \$51.23\text{ million}.

The unrounded modelled credit is $75.00-\sqrt{2{,}625.00}=\$23.77$ million. The allowed credit is $0.40(75.00-\sqrt{2{,}625.00})=\$9.51$ million. Capital after recognition is $75.00-0.40(75.00-\sqrt{2{,}625.00})=\$65.49$ million. Results are rounded to two decimals in millions at the end. The check is $9.51<23.77$, as required by the $40\%$ limit, and $65.49>51.23$, as expected when not all modelled credit is recognized.

A concentration concern can matter even when a mathematical aggregation produces a benefit: the insurer may depend heavily on a small number of exposures, counterparties, or underlying drivers. This example represents that concern through the stated limit; it does not claim that concentration can always be converted into a percentage haircut.

Why a Rating Assessment May Constrain Credit

The calculation identifies the consequence of a chosen aggregation structure. It does not, by itself, establish that the resulting credit is robust. Three considerations commonly explain why a rating assessment may be more cautious than the modelled output:

Concentration. A large share of a risk category may depend on a small set of exposures or on one underlying driver. A category-level charge can conceal that internal concentration. If the exposures are vulnerable to the same adverse development, an aggregate benefit based on broad category labels may overstate the practical scope for offsetting losses.
Dependence. Risks may become more connected in adverse conditions than in ordinary experience. A single correlation estimate may not capture changing relationships, shared event drivers, or dependence in the tail. A lower assumed correlation generally produces more credit in a variance-based formula, so an optimistic dependence input can materially lower the result.
Evidence quality. Sparse data, changing exposure composition, or limited experience with severe events can make a correlation estimate uncertain. Uncertainty does not prove that the estimate is wrong, but it weakens the case for relying on a precise benefit without qualification.

These concerns are distinct. A portfolio can be concentrated even if the average measured correlation is low; risks can be highly dependent without being concentrated in one exposure; and either issue can be difficult to evaluate when evidence is thin. A sound explanation states which concern is present rather than using “limited diversification” as an unsupported conclusion.

Common Exam Trap: Confusing Correlation with a Credit Percentage

Suppose the two charges are $\$60.00$ million and $\$80.00$ million with $\rho=0.75$, as in the first example. A common wrong calculation assumes that a correlation of $0.75$ means “75% of the total is retained,” and calculates capital as $0.75(60.00+80.00)=\$105.00$ million. That is not the square-root aggregation formula: correlation is an input to the cross term, not a direct multiplier on total capital.

The correct calculation is $\sqrt{60.00^2+80.00^2+2(0.75)(60.00)(80.00)}=\$131.15$ million. The resulting modelled credit is $\$8.85$ million, before any further recognition limit. The final rating treatment must then follow the stated recognition conditions. Keeping those steps separate prevents both a misapplied correlation and an assumption that modelled credit is automatically allowable.

Key takeaway: Report the modelled aggregate, the implied modelled credit, and the recognized credit as separate quantities. Explain any limitation by identifying its basis—such as concentration, dependence, or evidence—and apply only the adjustment specified in the problem.

Check Your Understanding

For each hypothetical problem, use the stated aggregation assumptions and give money amounts in millions rounded to two decimals. Do not apply a recognition limit unless the problem states one.

Two risk charges are $C_1=40.00$ and $C_2=30.00$, with $\rho=0.50$. Calculate the aggregate and modelled credit. (Answer: aggregate $=\$60.83$ million; modelled credit $=\$9.17$ million.)
An insurer has standalone charges totaling $\$95.00$ million and a stipulated modelled aggregate of $\$82.00$ million. A hypothetical rule caps recognized credit at $\$6.00$ million. What are recognized credit and capital after recognition? (Answer: recognized credit $=\$6.00$ million; capital after recognition $=\$89.00$ million.)
Two charges are $\$70.00$ million and $\$50.00$ million, with $\rho=0.60$. A hypothetical review recognizes $75\%$ of modelled credit. Calculate the modelled aggregate, recognized credit, and capital after recognition. (Answer: aggregate $=\$107.70$ million; recognized credit $=\$9.22$ million; capital after recognition $=\$110.78$ million.)
Three charges are each $\$25.00$ million, with every pairwise correlation equal to $0.20$. A stated hypothetical rule recognizes no more than $40\%$ of modelled credit. Calculate the modelled aggregate and capital after recognition. (Answer: aggregate $=\$51.23$ million; capital after recognition $=\$65.49$ million.)
Two charges are $\$50.00$ million and $\$40.00$ million. Compare modelled credits under $\rho=0.10$ and $\rho=0.70$. If limited evidence leads a hypothetical review to recognize $25\%$ of the credit under $\rho=0.70$, calculate that recognized credit and resulting capital. (Answer: modelled credits are $\$22.92$ million at $\rho=0.10$ and $\$6.93$ million at $\rho=0.70$; recognized credit $=\$1.73$ million; resulting capital $=\$88.27$ million.)

What You'll Learn