Diversification Benefits and Their Pricing Limits

Why Diversification Must Be Measured, Not Assumed

A catastrophe portfolio spread across locations may have lower aggregate volatility than the sum of its risks considered separately. But “spread out” is not a numerical result: the benefit depends on how losses can occur together and on which risk measure is being compared. Two locations can be exposed to the same regional event, while two nearby properties can have different vulnerabilities or policy terms. Geography is relevant to dependence, but it does not determine dependence by itself.

This tutorial compares standalone and aggregated catastrophe risk. The key distinction is that expected losses add, whereas variance and tail measures can change with dependence. That difference matters when interpreting catastrophe model output and when deciding whether a portfolio-level benefit is relevant to an individual risk’s price.

Definition: For a portfolio of \(n\) risks, let \(L_j\) be the annual loss from risk \(j\), and let \(L=\sum_{j=1}^{n}L_j\) be aggregate annual loss. A standalone measure evaluates \(L_j\) separately. An aggregate measure evaluates \(L\). For a risk measure \(R\), the diversification benefit relative to the sum of standalone measures is \(\sum_j R(L_j)-R(L)\). A positive value indicates a lower aggregate measure; a negative value indicates that aggregation increased it.

The comparison must use the same risk measure, confidence level, time horizon, and loss basis on both sides. For example, comparing a one-year aggregate 99% VaR with standalone 95% VaRs does not measure diversification: it changes the confidence level as well as the portfolio.

What Adds and What Depends on Dependence

If each loss has a finite expectation, linearity of expectation gives \(E[L]=\sum_j E[L_j]\), regardless of dependence. Thus, diversification does not, by itself, reduce the portfolio’s expected annual loss (AAL). Dependence does affect variance. For two losses,

\operatorname{Var}(L_1+L_2)=\operatorname{Var}(L_1)+\operatorname{Var}(L_2)+2\operatorname{Cov}(L_1,L_2).

Positive covariance increases aggregate variance relative to the zero-covariance case; negative covariance decreases it. Standard deviation is the square root of variance, so it does not add in the same way as expected loss. Tail measures, including VaR, depend on the full distribution of aggregate loss, not just its mean and variance.

A catastrophe model may represent dependence through shared event footprints, event frequencies, secondary uncertainty, or other modeled relationships. If two properties can be struck by the same event, their losses may be positively associated even when their locations differ. Conversely, geographic separation may reduce exposure to some events without eliminating common drivers such as a broad storm or a correlated hazard. Peril, geography, policy terms, and the modeled event set all affect the result.

Definition: For a loss \(L\) and confidence level \(\alpha\), define \(\operatorname{VaR}_{\alpha}(L)=\inf\{x:P(L\leq x)\geq\alpha\}\). This quantile convention matters when the loss distribution has probability masses. VaR is not always subadditive: it is possible for \(\operatorname{VaR}_{\alpha}(L_1+L_2)\) to exceed \(\operatorname{VaR}_{\alpha}(L_1)+\operatorname{VaR}_{\alpha}(L_2)\).

Worked Examples: Comparing Portfolio Measures

Worked Example: Independent Locations With Different Loss Sizes

A hypothetical insurer models two locations for one year. Location A has loss \(100{,}000\) dollars with probability \(0.10\), and zero otherwise. Location B has loss \(80{,}000\) dollars with probability \(0.10\), and zero otherwise. The two losses are independent. Calculate each expected loss and standard deviation, then the aggregate expected loss, standard deviation, 95% VaR, and the VaR diversification benefit relative to the sum of standalone 95% VaRs.

Method and assumptions. Treat each annual location loss as a Bernoulli loss, with the stated probability applying to the one-year period. Independence means the joint probability of both losses occurring is \(0.10(0.10)=0.01\). For a loss \(a\) occurring with probability \(p\), its expectation is \(pa\), and its variance is \(p(1-p)a^2\).

For A, the expected loss is \(0.10(100{,}000)=10{,}000\). Its variance is \(0.10(0.90)(100{,}000)^2=900{,}000{,}000\), so its standard deviation is \(\sqrt{900{,}000{,}000}=30{,}000\) dollars.

For B, the expected loss is \(0.10(80{,}000)=8{,}000\). Its variance is \(0.10(0.90)(80{,}000)^2=576{,}000{,}000\), so its standard deviation is \(\sqrt{576{,}000{,}000}=24{,}000\) dollars. Aggregate expected loss is \(10{,}000+8{,}000=18{,}000\) dollars.

Independence gives zero covariance, so aggregate variance is \(900{,}000{,}000+576{,}000{,}000=1{,}476{,}000{,}000\). The aggregate standard deviation is \(\sqrt{1{,}476{,}000{,}000}=38{,}418.75\) dollars, rounded to cents. The sum of standalone standard deviations is \(30{,}000+24{,}000=54{,}000\), so the reduction under this comparison is \(54{,}000-38{,}418.75=15{,}581.25\) dollars.

The aggregate loss distribution is: \(0\) with probability \(0.81\), \(80{,}000\) with probability \(0.09\), \(100{,}000\) with probability \(0.09\), and \(180{,}000\) with probability \(0.01\). Its cumulative probability through \(80{,}000\) is \(0.90\), and through \(100{,}000\) is \(0.99\). Therefore aggregate 95% VaR is \(100{,}000\) dollars. Standalone 95% VaRs are \(100{,}000\) and \(80{,}000\) dollars, respectively; their sum is \(180{,}000\). The VaR diversification benefit is \(180{,}000-100{,}000=80{,}000\) dollars.

Check. Aggregate expected loss is exactly the sum of standalone expectations, while both the standard-deviation comparison and the VaR comparison show a reduction relative to summed standalone measures. The result is specific to the stated independence assumption and metrics; it does not imply that expected loss fell.

Common mistake avoided: Adding standalone 95% VaRs and reporting \(180{,}000\) dollars as aggregate 95% VaR. The aggregate quantile must be calculated from the aggregate distribution; here it is \(100{,}000\) dollars.

Worked Example: A Shared Catastrophe Changes the Tail

Two hypothetical locations each incur a loss of \(100{,}000\) dollars if a specified regional event occurs, and zero otherwise. The event has annual probability \(0.10\). First, suppose the losses occur together whenever the event occurs. Then compare with a model in which the two locations have independent event indicators, each with probability \(0.10\). For each case, calculate aggregate expected loss, aggregate variance, and aggregate 95% VaR.

Method and assumptions. Use a one-year horizon and the stated event probabilities. In the shared-event case, one event indicator governs both locations. In the independent case, the indicators are independent. Each location’s expected loss is \(0.10(100{,}000)=10{,}000\) dollars, so expected aggregate loss is \(20{,}000\) dollars in either case.

In the shared-event case, aggregate loss is \(200{,}000\) dollars with probability \(0.10\), and zero with probability \(0.90\). Its variance is \(0.10(0.90)(200{,}000)^2=3{,}600{,}000{,}000\) dollars squared. Since cumulative probability at zero is only \(0.90\), aggregate 95% VaR is \(200{,}000\) dollars.

In the independent case, aggregate loss is zero with probability \(0.90^2=0.81\), \(100{,}000\) dollars with probability \(2(0.10)(0.90)=0.18\), and \(200{,}000\) dollars with probability \(0.10^2=0.01\). Each standalone variance is \(0.10(0.90)(100{,}000)^2=900{,}000{,}000\) dollars squared. Independence makes aggregate variance \(900{,}000{,}000+900{,}000{,}000=1{,}800{,}000{,}000\). The cumulative probability reaches \(0.99\) at \(100{,}000\), so aggregate 95% VaR is \(100{,}000\) dollars.

Check. Both cases have the same expected aggregate loss, but the shared-event model has twice the aggregate variance and a higher 95% VaR. The common event removes much of the modeled diversification. The calculation does not establish that two real locations are independent or perfectly dependent; those are explicit alternative assumptions.

Common mistake avoided: Treating separate location records as separate catastrophe events. The relevant question is whether losses can arise from the same event, not whether they are listed as distinct risks.

Why a Diversification Allocation Must Match Its Measure

An aggregate risk measure can be useful for portfolio decisions, but pricing or reporting often also requires a contribution by risk. The allocation must reconcile to the aggregate measure and must be identified as an allocation of that particular measure. A standard-deviation allocation illustrates the principle.

Result: For \(L=\sum_j L_j\) with finite variances and \(\operatorname{SD}(L)>0\), define the covariance allocation to risk \(j\) as \(\operatorname{Cov}(L_j,L)/\operatorname{SD}(L)\). These allocations sum to the aggregate standard deviation. The result does not say that standard deviation is the only appropriate pricing measure.

To verify the result, use the definition of covariance and the identity \(L=\sum_j L_j\):

\sum_j \operatorname{Cov}(L_j,L) =\operatorname{Cov}\left(\sum_j L_j,L\right) =\operatorname{Cov}(L,L) =\operatorname{Var}(L).

Dividing both sides by \(\operatorname{SD}(L)\) gives \(\sum_j \operatorname{Cov}(L_j,L)/\operatorname{SD}(L)=\operatorname{SD}(L)\). This allocation therefore reconciles exactly. It can differ materially from allocating the aggregate measure in proportion to expected loss or exposure, neither of which is implied by the variance identity.

Worked Example: Allocating Aggregate Standard Deviation

A hypothetical portfolio has independent annual losses \(L_1\) and \(L_2\). Risk 1 loses \(40{,}000\) dollars with probability \(0.50\), otherwise zero. Risk 2 loses \(30{,}000\) dollars with probability \(0.50\), otherwise zero. Calculate the portfolio standard deviation, allocate it using covariance contributions, and compare the result with the sum of standalone standard deviations.

Method and assumptions. Both risks have finite variance and are independent, so their covariance is zero. For risk 1, the expectation is \(0.50(40{,}000)=20{,}000\), and the variance is \(0.50(0.50)(40{,}000)^2=400{,}000{,}000\), giving a standalone standard deviation of \(20{,}000\) dollars.

For risk 2, the expectation is \(0.50(30{,}000)=15{,}000\), and the variance is \(0.50(0.50)(30{,}000)^2=225{,}000{,}000\), giving a standalone standard deviation of \(15{,}000\) dollars. Aggregate expected loss is \(35{,}000\) dollars. Aggregate variance is \(400{,}000{,}000+225{,}000{,}000=625{,}000{,}000\), so aggregate standard deviation is \(25{,}000\) dollars.

Because the risks are independent, \(\operatorname{Cov}(L_1,L)=\operatorname{Var}(L_1)=400{,}000{,}000\), and \(\operatorname{Cov}(L_2,L)=\operatorname{Var}(L_2)=225{,}000{,}000\). The allocated amounts are \(400{,}000{,}000/25{,}000=16{,}000\) dollars for risk 1 and \(225{,}000{,}000/25{,}000=9{,}000\) dollars for risk 2. Their sum is \(25{,}000\) dollars, matching aggregate standard deviation.

The sum of standalone standard deviations is \(20{,}000+15{,}000=35{,}000\) dollars, which exceeds the aggregate standard deviation by \(10{,}000\) dollars. The covariance allocation assigns the aggregate measure consistently; it does not turn the reduction into a lower expected loss. Check: allocated amounts sum to the computed \(25{,}000\)-dollar aggregate standard deviation.

Common mistake avoided: Dividing the portfolio standard deviation equally between the two risks without justification. The covariance method allocates \(16{,}000\) and \(9{,}000\) dollars here because the risks contribute different amounts of variance.

Pricing Limits and Common Exam Traps

A portfolio-level benefit is not automatically a credit of the same amount to every policy or location. Expected loss remains additive, while the amount of any tail-risk reduction depends on the modeled portfolio, its dependence assumptions, and the selected measure. A portfolio standard deviation or VaR is not itself a standalone estimate of the expected claims cost for each risk.

In practice, the relevance of a calculated benefit is limited by several considerations:

Peril and event dependence: A portfolio diversified against one peril may remain concentrated against another. Shared events can produce substantial positive dependence across locations.
Geography and exposure: Distance can alter event overlap, but location alone is not a substitute for a modeled dependence structure. Exposure concentration and vulnerability also matter.
Metric and confidence level: Variance, standard deviation, and tail measures answer different questions. VaR can fail subadditivity, so a lower aggregate VaR than summed standalone VaRs is not guaranteed.
Model and contract basis: Comparisons require consistent event assumptions, policy terms, limits, deductibles, and loss definitions. Otherwise an apparent benefit may reflect mismatched inputs.
Pricing use: A portfolio benefit may inform an aggregate risk view, but a risk-specific price still needs a defensible allocation and must respect applicable underwriting, regulatory, and business constraints. It should not be used to reduce expected loss merely because the portfolio is diversified.

A useful exam check is to write down the exact comparison before calculating: identify the loss variables, dependence assumption, time horizon, risk measure, and confidence level. Then verify whether the conclusion concerns expected loss, volatility, or a tail quantile. These are not interchangeable.

For example, a common incorrect calculation is to infer that two risks have no shared-event exposure because their marginal annual event probabilities are stated separately. Marginal probabilities alone do not establish independence. The correction is to use the stated joint dependence assumption—or to state that the aggregate tail cannot be determined uniquely without it.

Key takeaway: Diversification is metric-specific. Expected losses add; variance depends on covariance; tail measures depend on the aggregate distribution. A pricing implication requires an appropriate allocation and cannot be inferred from geographic spread alone.

Check Your Understanding

For each problem, use the stated one-year loss model and the VaR quantile convention defined above.

Two independent risks each have a loss of \(50{,}000\) dollars with probability \(0.20\), otherwise zero. Find aggregate expected loss, aggregate variance, and aggregate standard deviation. (Answer: Expected loss \(20{,}000\) dollars; variance \(800{,}000{,}000\) dollars squared; standard deviation \(28{,}284.27\) dollars.)
Two risks each lose \(70{,}000\) dollars if one shared event occurs, and zero otherwise. The event probability is \(0.05\). Find aggregate expected loss, variance, and 99% VaR. (Answer: Expected loss \(7{,}000\) dollars; variance \(931{,}000{,}000\) dollars squared; 99% VaR \(140{,}000\) dollars.)
Two independent risks each lose \(40{,}000\) dollars with probability \(0.10\), otherwise zero. Find aggregate 95% VaR and the sum of standalone 95% VaRs. (Answer: Aggregate 95% VaR \(40{,}000\) dollars; sum of standalone 95% VaRs \(80{,}000\) dollars.)
Two independent risks each lose \(100{,}000\) dollars with probability \(0.05\), otherwise zero. Find each standalone 95% VaR and aggregate 95% VaR. What does the result show about VaR subadditivity? (Answer: Each standalone 95% VaR is \(0\); aggregate 95% VaR is \(100{,}000\) dollars, so VaR is not subadditive in this example.)
Independent risks \(L_1\) and \(L_2\) lose \(20{,}000\) and \(10{,}000\) dollars, respectively, each with probability \(0.50\), otherwise zero. Find aggregate standard deviation and the covariance allocation to each risk. (Answer: Aggregate standard deviation \(11{,}180.34\) dollars; allocations \(8{,}944.27\) dollars to \(L_1\) and \(2{,}236.07\) dollars to \(L_2\), summing to the aggregate measure.)

What You'll Learn