Calculating Incremental Risk and Capital Effects

Why Incremental Effects Need a Consistent Comparison

A management action can reduce expected loss while increasing the severity of the worst outcomes, or reduce gross risk while adding costs and exposures that weaken the net result. A useful comparison therefore asks the same questions of the same horizon and loss definition for both the baseline portfolio and the proposed portfolio. This calculation framework is aligned with the CAS Exam 9 (Risk Management for Actuaries) syllabus.

The comparison should make its sign convention explicit. In this tutorial, a positive change means the proposed action has more of the quantity being measured; positive relief means it has less. These are different conventions, and confusing them reverses conclusions.

Definition: Let $L_0$ denote the baseline loss over a stated horizon and $L_1^g$ the proposed action's gross loss, before incremental action costs or added exposures. Let $L_1^n$ denote the net proposed loss after those items are included. For any measure $M$, the signed gross change is $M(L_1^g)-M(L_0)$, and the signed net change is $M(L_1^n)-M(L_0)$. A positive relief amount is the negative of the corresponding signed change.

The expected-loss change is linear: it is the proposed expected loss minus the baseline expected loss. Tail measures and capital measures generally are not linear in that way. In particular, calculate the tail measure of each complete portfolio distribution; do not assume that the tail measure of a change or component can simply be added or subtracted.

Tail Risk and Required Capital

For a loss variable $L$, define $\operatorname{VaR}_{\alpha}(L)$ as the smallest loss level whose cumulative probability is at least $\alpha$. For discrete losses, this is the loss at the $\alpha$-quantile, including the effect of probability mass at that point. Define $\operatorname{TVaR}_{\alpha}(L)$ as the average of the upper-tail quantiles. This quantile definition handles discrete distributions without treating all probability mass at the VaR point as if it were entirely above the threshold.

\operatorname{VaR}_{\alpha}(L)=\inf\{x:F_L(x)\geq\alpha\}, \qquad \operatorname{TVaR}_{\alpha}(L)=\frac{1}{1-\alpha}\int_{\alpha}^{1}\operatorname{VaR}_{u}(L)\,du.

For the comparisons below, define required tail capital at confidence level $\alpha$ as TVaR less expected loss. This is one stated capital convention, not a universal regulatory or rating-agency formula. If a decision framework prescribes another capital measure, use that measure consistently for baseline and proposed results.

K_{\alpha}(L)=\operatorname{TVaR}_{\alpha}(L)-E[L].

This definition isolates the tail component above the expected loss. It also means that expected-loss relief is not itself capital relief. A proposed action can have a lower mean but an even more adverse tail, so that required capital increases.

A Discrete-Tail Calculation

For a discrete distribution, order the outcomes from lowest to highest and identify the probability mass above the confidence quantile. If probability at the quantile must be split to fill the upper tail, include only the fraction needed. Equivalently, use the quantile integral in the definition. For example, when the largest 5% of outcomes consist exactly of two complete loss states, TVaR at 95% is their probability-weighted average divided by 0.05.

Derivation: The upper-tail quantile interval from $\alpha$ to 1 has width $1-\alpha$. The integral of the quantile over this interval adds each loss value weighted by the amount of probability assigned to it within that tail. Dividing by $1-\alpha$ converts the tail total into an average. Thus, for a discrete loss distribution, TVaR is the probability-weighted loss in the upper $1-\alpha$ tail, with any boundary mass included only as needed, divided by $1-\alpha$.

Once the expected loss and TVaR are calculated for each portfolio, capital change follows by subtraction. In particular, the change in capital is not generally the change in TVaR alone, because the expected loss also changes:

K_{\alpha}(L_1)-K_{\alpha}(L_0) = \bigl[\operatorname{TVaR}_{\alpha}(L_1)-\operatorname{TVaR}_{\alpha}(L_0)\bigr] - \bigl[E[L_1]-E[L_0]\bigr].

Worked Comparisons

The examples use hypothetical one-year losses in thousands of dollars. There is no mortality component and no discounting; for present-value conversion, $i=0$. Each loss distribution and its probabilities are stated directly. These assumptions keep attention on the comparison of portfolio outcomes rather than on a separate frequency, severity, or discounting model.

Worked Example: Measuring Gross Risk Relief

A hypothetical portfolio has baseline losses of 0, 100, 300, or 800, with probabilities 0.90, 0.05, 0.03, and 0.02, respectively. A proposed action changes the corresponding state losses to 0, 80, 240, and 700, with the same probabilities. Calculate the baseline and proposed expected losses, 95% VaR, 95% TVaR, and tail capital, then find the gross relief. Losses are in thousands of dollars; the horizon is one year, $i=0$, and there is no mortality component. Both distributions are discrete with the stated probabilities.

Method. First calculate the mean from all four states. Then find the 95th percentile and average the largest 5% of outcomes. Finally calculate $K_{0.95}=\operatorname{TVaR}_{0.95}-E[L]$ for each portfolio.

Baseline expected loss: $0.90(0)+0.05(100)+0.03(300)+0.02(800)=0+5+9+16=30$, or $30,000.

The cumulative probabilities at baseline losses 0 and 100 are 0.90 and 0.95, so $\operatorname{VaR}_{0.95}(L_0)=100$. The worst 5% consists of the 3% state at 300 and the 2% state at 800: $\operatorname{TVaR}_{0.95}(L_0)=[0.03(300)+0.02(800)]/0.05=(9+16)/0.05=500$. Thus $K_{0.95}(L_0)=500-30=470$, or $470,000.

Proposed expected loss: $0.90(0)+0.05(80)+0.03(240)+0.02(700)=0+4+7.2+14=25.2$, or $25,200.

The cumulative probability reaches 0.95 at 80, so $\operatorname{VaR}_{0.95}(L_1^g)=80$. Its worst 5% has probability-weighted loss $0.03(240)+0.02(700)=7.2+14=21.2$, giving $\operatorname{TVaR}_{0.95}(L_1^g)=21.2/0.05=424$. Its capital is $424-25.2=398.8$, or $398,800.

The signed expected-loss change is $25.2-30=-4.8$, so expected-loss relief is $4,800. TVaR relief is $500-424=76$, or $76,000. Capital relief is $470-398.8=71.2$, or $71,200. The calculation checks because $76-4.8=71.2$: the reduction in tail capital equals TVaR relief less expected-loss relief.

Common mistake avoided: Reporting the TVaR reduction of $76,000 as capital relief. Under the stated capital definition, the expected-loss change must also be accounted for.

Worked Example: Lower VaR but Higher Tail Capital

Use the same hypothetical baseline distribution as above. A different action produces losses of 0, 80, 250, and 1,000 in the four states, again with probabilities 0.90, 0.05, 0.03, and 0.02. Calculate its expected loss, 95% VaR, 95% TVaR, and capital, then compare with baseline. Losses are in thousands of dollars; this is a one-year discrete model with the stated probabilities, $i=0$, and no mortality component.

Method. Recalculate the full proposed distribution rather than inferring tail risk from the change in ordinary outcomes. In particular, the 1,000 loss in the 2% state belongs to the worst 5%.

The proposed expected loss is $0.90(0)+0.05(80)+0.03(250)+0.02(1{,}000)=0+4+7.5+20=31.5$. Its 95% VaR is 80 because cumulative probability reaches 0.95 there. Its worst 5% has weighted loss $0.03(250)+0.02(1{,}000)=7.5+20=27.5$, so TVaR is $27.5/0.05=550$. Required capital is $550-31.5=518.5$.

Relative to the baseline, expected loss increases by $31.5-30=1.5$, TVaR increases by $550-500=50$, and capital increases by $518.5-470=48.5$, all in thousands of dollars. The action lowers VaR from 100 to 80 but raises both TVaR and capital. The result checks because the capital increase is the TVaR increase of 50 less the expected-loss increase of 1.5.

Common mistake avoided: Treating a lower VaR as proof of lower tail risk. VaR reports a quantile threshold; it does not measure the severity of losses beyond that threshold.

Worked Example: Moving From Gross to Net Effects

Return to the first action, with gross losses 0, 80, 240, and 700. Suppose it also incurs a fixed implementation expense of 3 in every state and creates additional hypothetical operational or settlement losses of 0, 0, 50, and 200 in those states. Calculate its net expected loss, 95% TVaR, and capital, and compare net capital with baseline. Amounts are in thousands of dollars. The model is one-year and discrete with probabilities 0.90, 0.05, 0.03, and 0.02; $i=0$, and there is no mortality component.

Method. Add all action-specific items to each gross state before calculating net measures. This preserves the relationship between the added exposure and the portfolio states.

The net state losses are $0+3+0=3$, $80+3+0=83$, $240+3+50=293$, and $700+3+200=903$. The net expected loss is $0.90(3)+0.05(83)+0.03(293)+0.02(903)=2.7+4.15+8.79+18.06=33.7$.

Cumulative probability reaches 0.95 at 83, so net VaR is 83. The worst 5% weighted loss is $0.03(293)+0.02(903)=8.79+18.06=26.85$. Net TVaR is $26.85/0.05=537$, and net capital is $537-33.7=503.3$.

Baseline capital is 470, so net capital change is $503.3-470=33.3$: capital increases by $33,300. For comparison, the gross action had capital of 398.8, giving gross capital relief of $71,200. Its net expected loss is $33.7-30=3.7$ higher than baseline, and its net TVaR is $537-500=37$ higher. The capital result checks: $37-3.7=33.3$.

Common mistake avoided: Reporting gross capital relief without incorporating the action's added exposures. The net calculation must use net state losses, not just subtract a cost estimate from gross capital relief.

Interpreting the Gross-to-Net Bridge

The gross comparison isolates the direct change to the portfolio loss model. The net comparison includes incremental items attributable to the action, such as implementation expenses, new operational exposures, or other losses that belong in the same horizon and unit of account. State clearly which items are included; an omitted exposure can make a gross result look like a net result.

A deterministic expense has a useful special property under the capital convention used here. If a fixed amount $c$ is added to every possible loss, then each quantile shifts upward by $c$. Therefore TVaR and expected loss each rise by $c$, leaving their difference unchanged:

\operatorname{TVaR}_{\alpha}(L+c)=\operatorname{TVaR}_{\alpha}(L)+c,\qquad E[L+c]=E[L]+c,\qquad K_{\alpha}(L+c)=K_{\alpha}(L).

This does not mean the expense is irrelevant. It increases expected loss and total cost, and may affect a decision criterion that includes expected cost. It means only that a fixed amount added in every state does not change TVaR-minus-mean capital. A state-dependent cost or exposure can change the shape of the distribution and must be recalculated. A capital formula with a floor, adjustment, or prescribed component may also respond differently.

Common Errors in Incremental Comparisons

Mixing sign conventions. State whether a reported number is proposed minus baseline, or baseline minus proposed. Use “increase” and “relief” consistently.
Comparing unmatched quantities. Keep the horizon, currency basis, confidence level, and included loss items consistent. A one-year net result should not be compared with a gross result on another basis.
Substituting VaR for tail severity. A quantile alone says nothing about how large losses are beyond it. Calculate the specified tail measure for both portfolios.
Subtracting component TVaRs as if they were additive. Re-evaluate the complete proposed distribution. Dependence and changes in which outcomes occupy the tail can affect the result.
Calling TVaR relief capital relief. Under $K_{\alpha}=\operatorname{TVaR}_{\alpha}-E[L]$, account for both changes.
Leaving action costs outside the net loss definition. Show whether expenses and added exposures are included, and avoid adding an expected cost a second time if it is already in the loss outcomes.

Key takeaway: Calculate expected loss, tail risk, and capital separately for the baseline, gross action, and net action. Then take differences using a stated sign convention. Gross relief describes the direct modeled risk effect; net results include action-specific costs and exposures.

Check Your Understanding

For each problem, calculate the requested measures using the stated one-year loss distributions. Unless stated otherwise, amounts are in thousands of dollars, $i=0$, and there is no mortality component.

A baseline has losses 0, 50, 200, and 500 with probabilities 0.80, 0.10, 0.06, and 0.04. An action changes these to 0, 40, 180, and 400 with the same probabilities. At 90%, calculate each expected loss, TVaR, and capital using $K_{0.90}=\operatorname{TVaR}_{0.90}-E[L]$, then state capital relief. (Answer: baseline $E[L]=37$, TVaR $=320$, $K=283$; action $E[L]=30.8$, TVaR $=268$, $K=237.2$; capital relief $=45.8$.)
A discrete action-loss model has outcomes 0, 70, 250, and 900 with probabilities 0.90, 0.05, 0.03, and 0.02. Calculate expected loss, 95% VaR, TVaR, and tail capital. (Answer: $E[L]=29$, VaR $=70$, TVaR $=510$, and $K=481$.)
The gross action in Worked Example 1 has expected loss 25.2, TVaR 424, and capital 398.8. A fixed expense of 4 is added in every state. Find net expected loss, TVaR, and capital, and explain the capital change. (Answer: net expected loss $=29.2$, TVaR $=428$, and capital $=398.8$; the fixed expense increases mean and TVaR equally.)
Use the baseline in Worked Example 1. A proposed action has gross state losses 0, 80, 240, and 700 and adds state-dependent losses 0, 0, 100, and 0. Calculate the net expected loss, 95% TVaR, and capital, then compare capital with baseline. (Answer: net losses $=0,80,340,700$; expected loss $=28.2$, TVaR $=484$, capital $=455.8$, for capital relief of $14.2$.)

What You'll Learn