Why Portfolio Mix Matters
A catastrophe portfolio can contain several perils, regions, or peril-region combinations. The portfolio’s expected loss is the sum of the component expected losses, but its uncertainty depends on how those components move together. Two portfolios with identical component losses can therefore have different aggregate volatility when their dependence differs.
Diversification is not a claim that adding exposure must reduce the dollar amount of risk. Adding a region will generally increase expected loss and can increase aggregate standard deviation. The relevant question is whether the added exposure changes risk less than proportionally to portfolio size, and whether the dependence assumptions used to estimate that effect are credible. This distinction is useful in catastrophe portfolio steering and capital planning, aligned with the CAS Exam 9 (Risk Management for Actuaries) syllabus.
Aggregating Volatility Across Components
For two components, correlation summarizes their linear co-movement: \(\rho_{12}=\operatorname{Cov}(X_1,X_2)/(\sigma_1\sigma_2)\), provided both standard deviations are positive. For several components, calculate each pair’s covariance and include it in aggregate variance. A positive covariance increases aggregate variance; a negative covariance decreases it.
The second form applies when each \(\sigma_i\) is positive and \(\rho_{ij}\) is the correlation for that pair. If a component has zero variance, its covariance with every component is zero, and the covariance form remains applicable.
Specify what each modeled loss covers and use the same time period and loss basis for every component.
For each component, identify its expected loss and variance or standard deviation.
Identify the correlations or covariances for every relevant pair. Check that the complete dependence specification is mathematically consistent.
Calculate aggregate variance and standard deviation. Compare with a suitable benchmark, such as the sum of stand-alone standard deviations or a size-adjusted measure.
The variance formula follows by expanding the square around the aggregate mean. Write \(S-E[S]=\sum_i(X_i-E[X_i])\). Squaring this sum produces each squared deviation and, for every distinct pair \(i,j\), two copies of the cross-product. Taking expectations gives each component variance and twice each covariance. That is why the covariance terms enter with a factor of two.
Peril Mix: Compare the Dependence Assumption
Worked Example: Two Perils With Positive Dependence
A hypothetical insurer models one-year losses from a wind peril and a flood peril, in millions of dollars. Under a joint normal approximation, let \(X_W\sim N(14,64)\) and \(X_F\sim N(6,25)\), where the second parameter is the variance. Assume their correlation is \(0.30\). Calculate the aggregate mean, standard deviation, and standard-deviation diversification benefit. Compare the result with an independence assumption.
Method. Use the covariance aggregation formula. The standard deviations are \(\sigma_W=\sqrt{64}=8\) and \(\sigma_F=\sqrt{25}=5\), in millions. Under the stated joint normal approximation, their covariance is \(0.30(8)(5)=12\) million-squared.
The aggregate mean is \(E[S]=14+6=20\) million. Its variance is
Thus \(\operatorname{SD}(S)=\sqrt{113}=10.6301\) million. The sum of stand-alone standard deviations is \(8+5=13\) million, so the diversification benefit on this basis is \(13-10.6301=2.3699\) million. Relative to that sum, the benefit is \(2.3699/13=0.1823\), or \(18.23\%\).
Under independence, covariance is zero, so variance is \(64+25=89\) and standard deviation is \(\sqrt{89}=9.4330\) million. The positive dependence therefore raises aggregate standard deviation by \(10.6301-9.4330=1.1971\) million compared with independence. Check: the aggregate mean remains 20 million under either dependence assumption; dependence changes variance, not the sum of means.
Common mistake avoided: Adding variances and stopping at 89 silently assumes independence. With the stated positive correlation, the covariance term must be included.
A peril mix can improve volatility diversification when its components are less than perfectly positively dependent. But a lower aggregate standard deviation than the sum of component standard deviations does not mean that aggregate risk is small: the result can still be material, and the dependence assumption may dominate the calculation. A useful sensitivity analysis compares defensible dependence assumptions rather than relying on a single correlation estimate.
Geographic Expansion: Absolute and Size-Adjusted Risk
For equally sized regions with common standard deviation \(\sigma\), common mean \(\mu\), and the same pairwise correlation \(\rho\), the variance for \(n\) regions is \(\sigma^2[n+n(n-1)\rho]\). This equicorrelation model is a simplifying assumption, not a default description of real geographic exposure. It requires a valid covariance matrix; for equal pairwise correlations, the admissible range is \(-1/(n-1)\leq\rho\leq1\).
The formula shows why “does expansion reduce risk?” needs a defined measure. For positive \(\rho\), adding an equal-sized region increases total variance and standard deviation. Yet the ratio of standard deviation to expected loss can fall, remain unchanged, or rise depending on how the new region relates to the existing portfolio. If dependence is perfect positive correlation and the regions are identical in scale, the ratio stays unchanged: there is no size-adjusted diversification.
Worked Example: A New Region Under Two Dependence Assumptions
A hypothetical insurer has three regions, each with one-year loss mean \(10\) million and standard deviation \(9\) million. Model the regional losses as jointly normal. Assume the three existing regions have pairwise correlation \(0.20\). A fourth region of the same size is considered. Compare its effect if its correlation with each existing region is \(0.10\), and if instead its correlation with each is \(0.50\). Use standard deviation divided by expected aggregate loss as the size-adjusted measure.
Method. Calculate the original portfolio first, then add the fourth region using its covariances with the existing regions. The original mean is \(3(10)=30\) million. Each variance is \(9^2=81\), and each of the three existing pair covariances is \(0.20(9)(9)=16.2\).
The original standard deviation is \(\sqrt{340.2}=18.4445\) million, and its size-adjusted measure is \(18.4445/30=0.6148\).
Case 1: New-region correlations are \(0.10\). Each of the three new covariances is \(0.10(9)(9)=8.1\). The fourth region adds variance \(81\), plus twice the sum of its three covariances:
The new standard deviation is \(\sqrt{469.8}=21.6749\) million. Expected aggregate loss is \(4(10)=40\) million, so the size-adjusted measure is \(21.6749/40=0.5419\). Absolute volatility rises from \(18.4445\) to \(21.6749\) million, but volatility relative to expected loss falls from \(0.6148\) to \(0.5419\).
Case 2: New-region correlations are \(0.50\). Each new covariance is \(0.50(9)(9)=40.5\). Then
The new standard deviation is \(\sqrt{664.2}=25.7720\) million, and the size-adjusted measure is \(25.7720/40=0.6443\). Here, absolute volatility rises and volatility relative to expected loss also rises. Check: the original three-region variance remains \(340.2\) in both comparisons; only the new region’s assumed dependence changes.
Common mistake avoided: Concluding that expansion reduces dollar volatility because the size-adjusted measure falls in Case 1. It does not: standard deviation increases. The case for diversification is that the increase is relatively modest compared with the increase in expected portfolio size.
The fourth-region dependence specifications in this example are feasible with the stated correlations among the original regions. More generally, a set of individually plausible pairwise correlations can still produce an invalid covariance matrix. A model should check the full matrix rather than selecting each correlation independently.
From Aggregate Volatility to a Loss Quantile
A volatility comparison is not the same as a tail-risk measure. When a normal approximation is appropriate for the aggregate loss, its quantile at probability \(p\) is \(E[S]+z_p\operatorname{SD}(S)\), where \(z_p\) is the \(p\)-quantile of a standard normal distribution. This result depends on the aggregate being modeled as normal. Catastrophe losses may be skewed, heavy-tailed, or zero-inflated, so this approximation should not be treated as a general catastrophe-loss distribution.
Worked Example: Three Components and a 99% Normal Quantile
A hypothetical one-year catastrophe model has three components, each measured in millions. Under a joint normal approximation, \(X_1\sim N(4,36)\), \(X_2\sim N(7,64)\), and \(X_3\sim N(9,100)\). Their correlations are \(\rho_{12}=0.25\), \(\rho_{13}=0\), and \(\rho_{23}=0.50\). Calculate aggregate standard deviation and the 99% normal-approximation quantile. Compare it with the result from incorrectly assuming independence. Use \(z_{0.99}=2.3263\).
Method. The component standard deviations are \(6\), \(8\), and \(10\). The pair covariances are \(0.25(6)(8)=12\), \(0(6)(10)=0\), and \(0.50(8)(10)=40\), respectively. The aggregate mean is \(4+7+9=20\) million.
Therefore \(\operatorname{SD}(S)=\sqrt{304}=17.4356\) million. The 99% normal-approximation quantile is \(20+2.3263(17.4356)=60.56\) million, rounded to the nearest cent.
Under an independence assumption, variance would be \(36+64+100=200\), and standard deviation would be \(\sqrt{200}=14.1421\) million. The corresponding quantile is \(20+2.3263(14.1421)=52.90\) million, rounded to the nearest cent. The independence result understates the quantile by \(60.56-52.90=7.66\) million. Check: both calculations use the same mean and the same normal quantile multiplier; the difference comes from the covariance terms.
Common mistake avoided: Applying the normal quantile formula before aggregating dependence. The multiplier \(z_{0.99}\) does not correct an omitted covariance term.
Common Interpretation Traps
- Equating diversification with a lower expected loss. Expected aggregate loss is the sum of component means; a change in dependence alone does not alter it.
- Confusing absolute and relative risk. A portfolio can have higher aggregate standard deviation after expansion and still have lower standard deviation per dollar of expected loss.
- Assuming independence without saying so. Independence sets all pairwise covariances to zero; it must be supported or clearly identified as a scenario assumption.
- Treating zero correlation as independence in every model. That implication holds for jointly normal variables, not for arbitrary loss distributions.
- Using a single correlation as a complete tail-dependence model. Correlation measures linear association and may not capture how components behave together in extreme outcomes.
- Adding an inconsistent set of correlations. The resulting covariance matrix must be positive semidefinite; otherwise, it cannot describe a valid joint distribution.
Check Your Understanding
Use the stated assumptions for each problem. Unless a question says otherwise, amounts are in millions.
- Two modeled peril losses have standard deviations \(4\) and \(3\), with correlation \(0.50\). Calculate the aggregate standard deviation. (Answer: \(\sqrt{37}=6.0828\) million.)
- Three equally sized regions each have standard deviation \(5\), and every pair has correlation \(0.40\). Calculate aggregate variance and standard deviation. (Answer: Variance \(135\); standard deviation \(11.6190\) million.)
- An existing portfolio has standard deviation \(10\). A proposed component has standard deviation \(6\) and correlation \(0.20\) with the existing aggregate. Calculate the new aggregate standard deviation. (Answer: \(\sqrt{160}=12.6491\) million.)
- An aggregate loss is approximated by a normal distribution with mean \(12\) and standard deviation \(7\). Using \(z_{0.95}=1.6449\), calculate its 95% quantile. (Answer: \(23.51\) million, to the nearest cent.)
- Four equal-mean, equal-standard-deviation regions are modeled with common pairwise correlation \(\rho\). In terms of \(\rho\), does adding regions reduce standard deviation divided by expected aggregate loss when \(\rho<1\)? State the condition on \(\rho\) for the equicorrelation model to be valid for four regions. (Answer: Yes, the ratio decreases as the number of regions increases when \(\rho<1\); validity requires \(-1/3\leq\rho\leq1\).)