Normalizing Risk, Cost, and Capital Comparisons

Why Comparisons Need a Common Basis

A risk-transfer or capital decision can look attractive simply because two alternatives are being measured differently. One quote may cover one year while another covers two; one capital figure may be a total loss quantile while another is unexpected loss above the mean. A premium may include external expenses, whereas a competing quote may list them separately. Those figures are not directly comparable until their definitions and measurement bases are aligned.

A disciplined comparison states what loss is measured, over what period, at what confidence level, and on what gross or net basis. It also makes the treatment of expenses, taxes, and payment timing explicit. Normalization does not by itself select the preferred action. It makes the quantities being compared interpretable; the decision still depends on the organization's objectives and constraints.

Definition: A comparison is normalized when each alternative is expressed using consistent definitions of loss, exposure, time horizon, confidence level, cost components, tax treatment, and valuation date. Any adjustment must be supported by stated assumptions and sufficient information.

Specify the Risk Measure Before Comparing It

Let $X$ be a loss over a specified horizon, with cumulative distribution function $F_X$. The Value-at-Risk at confidence level $q$, for $0<q<1$, is the $q$-quantile of loss:

\operatorname{VaR}_q(X)=\inf\{x:F_X(x)\ge q\}.

A reported VaR is incomplete for comparison unless its loss definition, horizon, confidence level, and gross or net basis are known. Capital may also mean different things. In a loss-quantile framework, unexpected-loss capital is often defined as $\operatorname{VaR}_q(X)-E[X]$, while total VaR includes the expected loss. A comparison must say which quantity is being used.

The same care applies to the underlying loss. For example, a gross claims loss and a net claims loss after reinsurance are different random variables. A net result may also depend on whether expenses, reinstatement premiums, recoveries, or counterparty nonpayment are included. A label such as “annual risk” does not settle these choices.

Normalize the Time Horizon

Suppose annual losses $X_1,\ldots,X_n$ have the same distribution, with mean $\mu$ and variance $\sigma^2$, and are independent. The aggregate loss over $n$ years is $S_n=\sum_{t=1}^n X_t$. Linearity of expectation gives $E[S_n]=n\mu$. Independence makes all cross-covariances zero, so $\operatorname{Var}(S_n)=n\sigma^2$. Therefore, the aggregate standard deviation is $\sqrt{n}\sigma$, not $n\sigma$.

If annual losses are dependent, then $\operatorname{Var}(S_n)=\sum_t\operatorname{Var}(X_t)+2\sum_{s<t}\operatorname{Cov}(X_s,X_t)$. If the annual distributions change, even the identical-distribution assumption fails. A simple horizon conversion must not be used without checking these conditions.

Worked Example: Comparing One-Year and Two-Year Loss Measures

Two hypothetical alternatives cover the same type of loss, but their models are reported over different horizons. Alternative A has annual losses that are independent and normally distributed as $N(10,16)$ million dollars. Alternative B is reported directly over two years as $N(16,18)$ million dollars. Compare their two-year 99.5% VaR. Use $z_{0.995}=2.576$, where $z_q$ is the standard normal $q$-quantile.

Method and assumptions. A normal distribution is written $N(\mu,\sigma^2)$, so its second parameter is the variance. For A, annual mean is 10 million and annual standard deviation is $\sqrt{16}=4$ million. Assume the two annual losses are independent. The two-year aggregate is then normal with mean $2(10)=20$ million and variance $2(16)=32$ million.

S_{2,A}\sim N(20,32),\qquad \operatorname{SD}(S_{2,A})=\sqrt{32}=4\sqrt{2}\approx5.656854.

For a normal loss, $\operatorname{VaR}_q=\mu+z_q\sigma$. Thus A's two-year VaR is $20+2.576(5.656854)=34.57206$ million, or $34.572 million to three decimal places.

B is already on a two-year basis. Its mean is 16 million, its variance is 18 million, and its standard deviation is $\sqrt{18}=3\sqrt{2}\approx4.242641$ million. Its two-year VaR is $16+2.576(4.242641)=26.92904$ million, or $26.929 million to three decimal places.

On this stated basis, A's VaR exceeds B's by $34.57206-26.92904=7.64302$ million, or $7.643 million to three decimal places. This is a comparison of the specified loss distributions, not a complete decision: the alternatives' costs and other terms have not been assessed.

Common mistake avoided: Multiplying A's annual standard deviation by two would give 8 million. That is not the standard deviation of the sum under independence; the correct value is $\sqrt{2}(4)=5.656854$ million. Likewise, comparing A's one-year VaR directly with B's two-year VaR would mismatch horizons.

Key result: If annual losses are independent and identically distributed normal variables $N(\mu,\sigma^2)$, then their $n$-year sum is $N(n\mu,n\sigma^2)$. The corresponding $q$-VaR is $n\mu+z_q\sqrt{n}\sigma$. This result does not apply unchanged when annual losses are dependent, non-identically distributed, or not adequately represented by the assumed normal model.

Normalize Confidence Levels and Capital Definitions

A 99% VaR and a 99.5% VaR describe different points in a loss distribution. A reported capital number at one confidence level cannot generally be converted to another using a fixed multiplier. The conversion requires information about the distribution or a justified model for it. Even under a normal model, the expected loss must be separated from unexpected-loss capital if the reported measure includes the mean.

The distinction can be shown directly. If $X\sim N(\mu,\sigma^2)$, then $\operatorname{VaR}_q(X)=\mu+z_q\sigma$. If capital is defined as unexpected loss, then $K_q=\operatorname{VaR}_q(X)-\mu=z_q\sigma$. The formula follows because standardizing $X$ gives $(X-\mu)/\sigma\sim N(0,1)$; its $q$-quantile is $z_q$, so the original-scale quantile is $\mu+z_q\sigma$. Subtracting the mean leaves $z_q\sigma$.

Worked Example: Converting Capital to a Common Confidence Level

Two hypothetical portfolios have one-year normally distributed losses on the same net basis. Portfolio C has $X_C\sim N(5,4)$ million dollars, and its capital is reported at 99.5%. Portfolio D has $X_D\sim N(6,9)$ million dollars, but its capital is reported at 99%. Compare unexpected-loss capital at 99.5%. Use $z_{0.99}=2.326$ and $z_{0.995}=2.576$.

Method. Convert each model to the same confidence level, then report unexpected-loss capital as VaR less expected loss. C has standard deviation $\sqrt{4}=2$ million. Its 99.5% VaR is $5+2.576(2)=10.152$ million, so its unexpected-loss capital is $10.152-5=$ $5.152 million.

D has standard deviation $\sqrt{9}=3$ million. Its reported 99% VaR is $6+2.326(3)=12.978$ million, and its 99% unexpected-loss capital is $12.978-6=6.978$ million. At 99.5%, its VaR is $6+2.576(3)=13.728$ million, so its 99.5% unexpected-loss capital is $13.728-6=$ $7.728 million.

At the common 99.5% level, D's unexpected-loss capital exceeds C's by $7.728-5.152=$ $2.576 million. The comparison uses both the same confidence level and the same capital definition.

Common mistake avoided: Comparing D's reported 99% capital of $6.978 million directly with C's 99.5% capital of $5.152 million mixes confidence levels. Nor is it correct to compare total VaR for one portfolio with unexpected-loss capital for the other.

Separate Cost Components, Taxes, and Payment Timing

A cost comparison should identify which amounts are included in each quote. Premium, internal administration, expected retained claims, and other cash expenses may be listed separately or embedded in a quoted amount. Count each component once. For an after-tax comparison, specify the applicable tax rate and which costs are deductible or taxable. The shortcut of multiplying all costs by $1-\tau$ is valid only when every included cost is deductible at tax rate $\tau$, and the tax benefit is usable in the relevant period.

Payment dates matter as well. To compare cash flows at a common valuation date, discount each payment using an explicitly stated effective annual interest rate $i$, with $v=1/(1+i)$. A payment $C_t$ made at the end of year $t$ has present value $C_tv^t$. A present-value comparison requires consistent payment timing and valuation assumptions; it is not a comparison of undiscounted totals.

Worked Example: Comparing Annual Costs After Tax

A hypothetical insurer compares two one-year risk-transfer programs. For A, the annual premium is $1.10 million and includes the provider's expenses; internal administration costs $0.12 million, and expected retained claims cost $0.40 million. For B, the premium is $0.90 million and excludes a $0.20 million program-administration cost; expected retained claims cost $0.55 million. Assume a 25% tax rate, that all listed amounts are deductible in the same year, and that tax benefits can be used. Compare total after-tax costs.

Method. First include each cost once. A's pretax total is $1.10+0.12+0.40=1.62$ million dollars. Its after-tax cost is $1.62(1-0.25)=1.62(0.75)=$ $1.215 million.

B's pretax total is $0.90+0.20+0.55=1.65$ million dollars. Its after-tax cost is $1.65(0.75)=$ $1.2375 million, or $1.238 million to three decimal places. B's after-tax cost exceeds A's by $1.2375-1.215=0.0225$ million, or $0.023 million to three decimal places.

Common mistake avoided: Comparing only the quoted premiums, $1.10 million for A and $0.90 million for B, omits B's separately stated administration and retained-claim costs. Conversely, adding external expenses to A's premium again would double-count them because the problem states that they are included.

Worked Example: Comparing Costs Paid at Different Times

Two hypothetical two-year contracts provide the same stated service, with no tax effects. Contract P requires one payment of $1.90 million immediately. Contract Q requires $1.00 million at the end of each of years 1 and 2. Using an effective annual interest rate of $i=5\%$, compare present values at time 0 and calculate P's level annual equivalent over two years.

Method. The discount factor is $v=1/1.05=0.952381$. P's present value is its immediate payment: $1.900 million. Q's present value is $1.00v+1.00v^2$ million, or $0.952381+0.907029=$ $1.859410 million.

Q has the lower present value by $1.900000-1.859410=$ $0.040590 million. The two-year annuity-immediate factor is $a_{\overline{2}|}=v+v^2=1.859410$. P's level annual equivalent $C$ satisfies $C a_{\overline{2}|}=1.90$, so $C=1.90/1.859410=1.02183$ million dollars per year, or $1.022 million per year to three decimal places.

Common mistake avoided: Adding Q's two payments and comparing the $2.00 million undiscounted total with P's $1.90 million immediate payment ignores their different timing. The present-value comparison uses the stated interest rate and the same valuation date.

A Practical Normalization Sequence

Before ranking alternatives, record the basis of each measure and resolve mismatches in a controlled order. The sequence below is a useful check on a risk, cost, and capital comparison.

Define the exposure and loss.
State what is included, whether the measure is gross or net, and how recoveries, expenses, and other relevant cash flows are treated.

Align the horizon.
Use a common period. If converting horizons, state the assumptions about annual distributions and dependence; do not scale VaR mechanically.

Align the risk and capital measure.
Use a common confidence level and distinguish total loss quantiles from capital measured above expected loss.

Reconcile costs and taxes.
List included and excluded expenses, expected retained loss, tax treatment, and whether tax effects are usable.

Align valuation dates.
Discount cash flows when timing differs, using a stated interest rate and consistent payment assumptions.

Common Comparison Traps

Mixing gross and net loss. A net loss after recoveries is not directly comparable with a gross loss unless the recovery treatment is reconciled.
Scaling a tail measure linearly in time. Under independent, identically distributed normal annual losses, the mean scales with $n$ and the standard deviation with $\sqrt{n}$. Other models or dependencies require their own treatment.
Mixing confidence levels or capital definitions. A total VaR and an unexpected-loss capital amount are not interchangeable.
Counting costs inconsistently. Check whether expenses are embedded in a premium before adding them separately.
Applying a tax factor without checking deductibility. The formula $C(1-\tau)$ does not apply to costs that are not deductible or when the tax benefit cannot be used.
Ignoring cash-flow timing. A premium paid today and equal nominal payments made later do not have the same present value.

Normalization is useful because a decision-maker can then see which differences arise from the alternatives and which arise only from measurement conventions. It does not remove model uncertainty or make unlike risks identical. If key information is unavailable—for example, the dependence needed to aggregate multiyear losses—the responsible conclusion is that the comparison cannot be fully normalized without an additional assumption or analysis.

Key takeaway: A fair risk, cost, and capital comparison starts by fixing the basis, not by ranking the reported numbers. State the loss definition, horizon, confidence level, capital convention, expense and tax treatment, and valuation date; then make only those conversions justified by the available model and assumptions.

Check Your Understanding

For each problem, normalize the stated quantities to the requested basis and retain the assumptions given.

A hypothetical portfolio has independent annual losses $N(7,4)$ million dollars, where the second parameter is variance. Find the three-year 95% VaR using $z_{0.95}=1.645$. (Answer: $26.698 million.)
Annual losses for a hypothetical account are independent and identically distributed $N(4,4)$ million dollars. Find two-year 95% VaR using $z_{0.95}=1.645$. (Answer: $12.653 million.)
Two one-year programs have fully deductible costs at a usable 25% tax rate. Program X has premium $0.72 million, fee $0.08 million, and expected retained claims $0.40 million. Program Y has premium $0.65 million, fee $0.15 million, and expected retained claims $0.50 million. Find each after-tax total and the difference. (Answer: X $0.900 million; Y $0.975 million; Y costs $0.075 million more after tax.)
A one-year loss is $N(3,2.25)$ million dollars. At 99.5%, calculate total VaR and unexpected-loss capital using $z_{0.995}=2.576$. (Answer: VaR $6.864 million; unexpected-loss capital $3.864 million.)
At an effective annual interest rate of 4%, compare an immediate payment of $1.42 million with three payments of $0.50 million at the ends of years 1, 2, and 3. Give the present value of the installments and state which alternative has the lower present value. (Answer: Installments have present value $1.388 million to three decimals; they are lower by approximately $0.032 million.)

What You'll Learn