Why Marginal Catastrophe Cost Matters
A policy can have a modest expected catastrophe loss and still add materially to a portfolio’s tail risk if its losses tend to occur alongside existing concentrations. Conversely, an exposure’s standalone tail measure does not by itself show how much the portfolio’s tail measure will change. Pricing a portfolio change therefore requires a before-and-after comparison using a consistent loss basis, time horizon, and set of model assumptions.
The same comparison applies when adding a new risk, renewing an existing contract, changing terms, or repricing an exposure. A premium decision is not made by simply assigning the exposure its standalone catastrophe metric. The relevant question is what the change does to the portfolio, and whether the incremental premium covers the resulting expected loss, expenses, and any required risk or capital charge.
The comparison must use consistent definitions. For example, if \(L\) is net of a particular contractual recovery, the proposed portfolio loss \(L+Z\) must be constructed on that same net basis. The exposure contribution \(Z\) should represent the change in annual aggregate loss under the terms being assessed, not a premium or a policy limit. If a renewal changes both the existing portfolio and the contract’s loss contribution, model the complete revised portfolio rather than treating the change as an unrelated new policy.
Portfolio Metrics and Discrete Tail Measures
For an annual loss random variable \(L\), define its cumulative distribution function by \(F_L(x)=P(L\leq x)\). At probability level \(p\), Value at Risk is the lower quantile \(\operatorname{VaR}_p(L)=\inf\{x:F_L(x)\geq p\}\). This convention matters when losses are discrete: VaR is the first loss value at which cumulative probability reaches or exceeds \(p\).
Tail Value at Risk at level \(p\), also called the expected shortfall under a common convention, averages the worst \(1-p\) fraction of outcomes. For an integrable loss with quantile function defined by the lower-quantile convention above,
For a discrete distribution, this integral uses only the fraction of probability mass needed to complete the worst \(1-p\) tail. One must not include an entire boundary probability mass if doing so would make the tail larger than \(1-p\). A practical way to calculate TVaR is to sort outcomes from largest loss to smallest, take probability mass until the required tail probability is filled, and use only the necessary fraction of the last outcome’s probability.
Worked Example: Independent Addition and Fractional Tail Mass
Worked Example: Independent Addition and Fractional Tail Mass
A hypothetical insurer models annual catastrophe losses for its current portfolio as \(L=0\) with probability \(0.90\) and \(L=40\) (in thousands of dollars) with probability \(0.10\). A proposed, independent exposure has annual loss \(Z=0\) with probability \(0.80\) and \(Z=20\) with probability \(0.20\). Find the change in expected loss, 95% VaR, and 95% TVaR. All losses refer to the same annual horizon and are measured in thousands of dollars.
Method. First calculate the current and added expected losses. Then form the joint outcomes using independence, aggregate losses, and calculate each portfolio metric from its distribution.
The current expected loss is \(E[L]=0(0.90)+40(0.10)=4\). The added exposure’s expected loss is \(E[Z]=0(0.80)+20(0.20)=4\). Therefore, the expected-loss increase is \(E[Z]=4\), or \(\$4{,}000\).
Independence gives the following combined distribution: \(L+Z=0\) with probability \(0.90(0.80)=0.72\); \(20\) with probability \(0.90(0.20)=0.18\); \(40\) with probability \(0.10(0.80)=0.08\); and \(60\) with probability \(0.10(0.20)=0.02\). The probabilities sum to \(0.72+0.18+0.08+0.02=1.00\). The combined expected loss is \(0(0.72)+20(0.18)+40(0.08)+60(0.02)=3.6+3.2+1.2=8\), which checks against \(E[L]+E[Z]=4+4=8\).
For the current portfolio, the cumulative probability is \(0.90\) at loss \(0\) and \(1.00\) at loss \(40\), so \(\operatorname{VaR}_{0.95}(L)=40\). For the combined portfolio, cumulative probabilities at losses \(0,20,40,60\) are \(0.72,0.90,0.98,1.00\). Thus \(\operatorname{VaR}_{0.95}(L+Z)=40\), and the marginal 95% VaR is \(40-40=0\).
The worst 5% of current outcomes all have loss \(40\), so \(\operatorname{TVaR}_{0.95}(L)=40\). For the combined portfolio, the worst 5% consists of all \(0.02\) probability at loss \(60\), plus \(0.03\) probability at loss \(40\). Hence \(\operatorname{TVaR}_{0.95}(L+Z)=[0.02(60)+0.03(40)]/0.05=(1.2+1.2)/0.05=48\). The marginal 95% TVaR is \(48-40=8\), or \(\$8{,}000\).
Check and common mistake. The combined TVaR exceeds the current TVaR, but it is not the standalone TVaR of \(Z\), which is \(20\). Nor is the marginal TVaR \(20\): it is the difference \(48-40=8\). The calculation correctly uses only \(0.03\) of the \(0.08\) probability mass at loss \(40\) to complete the 5% tail.
Dependence Determines the Marginal Tail Cost
The expected-loss change is especially straightforward: if the proposed portfolio is exactly \(L+Z\), then linearity gives \(E[L+Z]-E[L]=E[Z]\), regardless of dependence. Tail measures work differently. To calculate a marginal VaR or TVaR, the distribution of \(L+Z\) must reflect whether the proposed exposure’s losses occur in the same years as large existing losses.
The following justification explains the fractional-tail calculation. In a discrete loss distribution, suppose an outcome \(x_j\) has probability mass that crosses the upper-tail boundary. The worst \(1-p\) fraction contains every larger outcome’s probability mass and only enough of \(x_j\)’s mass to make the total exactly \(1-p\). Dividing the loss-weighted mass in that selected tail by \(1-p\) gives the average loss in the tail. This is the discrete calculation implied by the quantile integral: the quantile function remains at each loss value over the corresponding interval of cumulative probabilities.
Here the outcomes \(x_j\) are selected from largest downward, and each selected weight \(w_j\) is no greater than the outcome’s probability. This formula avoids both omitting part of the boundary mass and incorrectly taking all of it.
Worked Example: A Shared Event Changes Tail Cost
A hypothetical insurer models the current portfolio loss \(L\) and the loss \(Z\) from a proposed regional exposure using the following joint annual distribution. Losses are in thousands of dollars. Find the expected-loss increase and marginal 95% TVaR.
| Probability | Current loss \(L\) | Added loss \(Z\) |
|---|---|---|
| 0.90 | 0 | 0 |
| 0.04 | 0 | 50 |
| 0.04 | 100 | 0 |
| 0.02 | 100 | 50 |
Method. Use the joint distribution directly; do not impose independence. The probabilities sum to \(0.90+0.04+0.04+0.02=1.00\). Current expected loss is \(100(0.04+0.02)=6\). The proposed exposure’s expected loss is \(50(0.04+0.02)=3\), so the expected-loss increase is \(3\), or \(\$3{,}000\).
The combined outcomes are \(0\) with probability \(0.90\), \(50\) with probability \(0.04\), \(100\) with probability \(0.04\), and \(150\) with probability \(0.02\). Their expected loss is \(0(0.90)+50(0.04)+100(0.04)+150(0.02)=2+4+3=9\). This agrees with \(E[L]+E[Z]=6+3=9\).
The current portfolio has loss \(100\) with probability \(0.06\), so its worst 5% all have loss \(100\) and \(\operatorname{TVaR}_{0.95}(L)=100\). In the combined portfolio, the worst 5% consists of all \(0.02\) probability at loss \(150\) and \(0.03\) probability at loss \(100\). Therefore, \(\operatorname{TVaR}_{0.95}(L+Z)=[0.02(150)+0.03(100)]/0.05=(3+3)/0.05=120\). The marginal 95% TVaR is \(120-100=20\), or \(\$20{,}000\).
Check and common mistake. The added exposure’s standalone 95% TVaR is \(50\), because its loss is \(50\) with probability \(0.06\). Adding that standalone value to the current TVaR would give \(150\), not the combined TVaR of \(120\). The direct joint calculation shows why that shortcut fails: the exposure loss coincides with a current loss of \(100\) in only \(0.02\) of annual outcomes. Dependence changes the marginal tail cost, even though it does not change the expected-loss increase.
Turning Marginal Cost Into a Rate Test
A marginal catastrophe cost is an input to pricing, not automatically the required premium. A transparent incremental rate test can separate the expected loss, a stated tail-risk charge, and expenses. One simple rule is to charge the expected-loss increase plus a specified proportion of a positive marginal TVaR, plus incremental expenses. The rule’s percentage and expense amount are pricing assumptions, not universal constants; they should be stated and applied consistently.
Here \(\Delta E[L]\) is the revised portfolio’s expected loss minus the current portfolio’s expected loss, \(\Delta\operatorname{TVaR}_p\) is the corresponding change in portfolio TVaR at the selected level, and \(\alpha\) is the stated charge per unit of positive marginal TVaR. Taking the positive part is appropriate only when the stated pricing rule does so; it must not be silently assumed for a different rule. Taxes, commissions, profit provisions, or other loadings should be added only if the stated premium test includes them.
Worked Example: Testing a Renewal’s Incremental Premium
A hypothetical renewal is evaluated with two discrete annual portfolio-loss distributions, in thousands of dollars. Before renewal, loss is \(0\) with probability \(0.98\) and \(90\) with probability \(0.02\). After renewal, loss is \(0.4489795918\) with probability \(0.98\) and \(93\) with probability \(0.02\). The rate rule charges the expected-loss increase, 12% of positive marginal 99% TVaR, and \(0.4\) thousand dollars of incremental expenses. The offered incremental premium is \(1.10\) thousand dollars. Determine the required incremental premium and whether the offer meets the rule.
Method. Calculate the expected loss and 99% TVaR on each side, take their differences, and then apply the stated pricing rule.
Before renewal, expected loss is \(0(0.98)+90(0.02)=1.8\). After renewal, expected loss is \(0.4489795918(0.98)+93(0.02)=0.44+1.86=2.30\), to the displayed precision. The expected-loss increase is \(2.30-1.80=0.50\).
For each distribution, the probability at its lower loss is \(0.98\), which is below \(0.99\). The lower-quantile 99% VaR is therefore the higher loss value: \(90\) before renewal and \(93\) after renewal. In each case the worst 1% lies entirely within the 2% probability mass at that higher loss. Thus, 99% TVaR is \(90\) before and \(93\) after, giving marginal 99% TVaR of \(3\).
The required incremental premium is \(0.50+0.12(3)+0.40=0.50+0.36+0.40=1.26\) thousand dollars, or \(\$1{,}260\). The offer of \(1.10\) thousand dollars is short by \(1.26-1.10=0.16\) thousand dollars, or \(\$160\), so it does not meet the rule.
Check and common mistake. The expected-loss increase is \(0.50\), not the revised portfolio expected loss of \(2.30\). The tail charge is \(0.36\), and expenses are \(0.40\); all three components must be included. Omitting expenses would give \(0.86\), while omitting the expected-loss increase would give \(0.76\); neither is the stated rule.
Common Exam and Pricing Traps
- Using a standalone metric as the portfolio increment. Calculate \(\rho(L+Z)-\rho(L)\) using a combined distribution, not \(\rho(Z)\) alone.
- Adding VaRs or TVaRs. These measures are generally not additive. Form the portfolio loss distribution using the stated dependence assumptions or a supplied joint model.
- Misreading a quantile boundary. If the CDF is below \(p\) at one loss and first reaches \(p\) at the next, the lower-quantile VaR is the next loss, not the earlier one.
- Taking too much discrete tail probability. Use exactly \(1-p\) probability mass for TVaR, including only the necessary fraction of a boundary outcome.
- Confusing rate adequacy with expected-loss adequacy. A premium may cover expected loss and still fail a stated expense and tail-risk rule.
- Changing the loss basis between comparisons. Gross and net losses, annual and event losses, or differently defined terms cannot be compared as though they were the same quantity.
For a portfolio decision, report the expected-loss increase separately from the marginal tail measure. State the selected probability level, the dependence assumption or joint model, and the exact premium rule. This makes clear whether a rate recommendation reflects expected catastrophe cost, tail risk, or both, and makes the calculation auditable.
Check Your Understanding
Use the stated distributions and conventions. Loss amounts are in thousands of dollars unless otherwise specified.
- A current portfolio has loss \(0\) with probability \(0.90\) and \(30\) with probability \(0.10\). An independent addition has loss \(0\) with probability \(0.75\) and \(8\) with probability \(0.25\). Find the expected-loss increase and combined 90% VaR. (Answer: expected-loss increase \(2\); combined 90% VaR \(8\).)
- A joint model has outcomes \((L,Z)=(0,0),(0,40),(80,0),(80,40)\) with probabilities \(0.92,0.03,0.03,0.02\), respectively. Find the expected-loss increase and marginal 95% TVaR. (Answer: expected-loss increase \(2\); marginal 95% TVaR \(16\).)
- A renewal increases expected annual portfolio loss by \(0.7\) and 99% TVaR by \(4\). Under a rule charging 15% of positive marginal 99% TVaR and \(0.2\) of incremental expenses, find the required incremental premium. An offer is \(1.35\). How far is it above or below the rule? (Answer: required premium \(1.50\); the offer is short by \(0.15\).)
- A modeled annual loss is \(0\) with probability \(0.94\), \(25\) with probability \(0.05\), and \(70\) with probability \(0.01\). Find 95% VaR and 95% TVaR using the lower-quantile and fractional-tail conventions. (Answer: 95% VaR \(25\); 95% TVaR \(34\).)
- A current portfolio has loss \(0\) with probability \(0.90\) and \(60\) with probability \(0.10\). An independent addition has loss \(0\) with probability \(0.80\) and \(15\) with probability \(0.20\). Find the expected-loss increase and the combined 90% VaR. (Answer: expected-loss increase \(3\); combined 90% VaR \(15\).)