Calibrating Model Outputs to Capital Targets

Why Calibrate a Model to a Capital Target?

A capital model produces a distribution of losses under specified assumptions; a capital target states the amount the model is expected to support at a particular confidence level and horizon. Calibration connects the two. It should not be treated as a mechanical adjustment that makes one number agree: the model basis, loss definition, horizon, confidence level, and meaning of capital must all match the target.

For example, a target may be stated as a loss quantile, or as unexpected loss above expected loss. These are different quantities. A model can match one and miss the other, especially when an adjustment changes the mean as well as the tail. Calibration also needs a plausibility check: scaling may preserve the distribution's relative shape, but it cannot demonstrate that the original tail shape or component assumptions are appropriate.

Definition: For a loss random variable $X$ with distribution function $F_X$, define its $q$-quantile (Value-at-Risk at confidence level $q$) by $\operatorname{VaR}_q(X)=\inf\{x:F_X(x)\ge q\}$. If the target is unexpected-loss capital, define $K_q(X)=\operatorname{VaR}_q(X)-E[X]$. If the target is gross quantile capital, use $\operatorname{VaR}_q(X)$ itself. State which definition is being calibrated.

The distinction matters in practice. When a model's entire loss distribution is multiplied by a positive constant, both its mean and its quantile are multiplied by that constant, so unexpected-loss capital is multiplied by it as well. But if only the volatility is adjusted while the mean is held fixed, the quantile changes without changing the mean. The calibration method must match the intended target and the intended adjustment.

Horizon and Confidence Level Must Match

A model's time horizon is part of its specification, not a label that can be changed after calculating capital. Under a particular model, losses over multiple years can be aggregated and their quantile recalculated. A common special case assumes annual losses are independent and identically distributed with finite mean $\mu$ and variance $\sigma^2$. For the sum of $T$ annual losses, the mean is $T\mu$, and the variance is $T\sigma^2$. If the annual losses are normal, their sum is also normal; its standard deviation is $\sigma\sqrt{T}$.

X_t\sim N(\mu,\sigma^2)\text{ independently for }t=1,\ldots,T \quad\Longrightarrow\quad \sum_{t=1}^{T}X_t\sim N(T\mu,T\sigma^2)

For this normal model, the $q$-quantile is $T\mu+z_q\sigma\sqrt{T}$, where $z_q$ is the standard normal $q$-quantile. Thus unexpected-loss capital is $z_q\sigma\sqrt{T}$, while gross quantile capital includes the expected loss $T\mu$. This square-root-of-time result depends on the stated assumptions; it is not a general rule for every loss distribution or dependence structure. In particular, serial dependence or changing annual distributions changes the variance aggregation.

Derivation: Independence makes the variance of the sum equal to the sum of the annual variances: $\operatorname{Var}(\sum_{t=1}^T X_t)=\sum_{t=1}^T\operatorname{Var}(X_t)=T\sigma^2$. A sum of independent normal variables is normal, and standardizing that sum gives $(\sum X_t-T\mu)/(\sigma\sqrt T)\sim N(0,1)$. Its $q$-quantile is therefore $z_q$, yielding the stated quantile and unexpected-capital formulas.

Worked Example: Converting an Annual Deviation Model to a Two-Year Target

A hypothetical insurer models annual deviations from expected loss as independent, identically distributed $Z_t\sim N(0,25)$, with amounts in millions of dollars. The two-year target is unexpected-loss capital at $q=0.995$, for which use $z_{0.995}=2.5758$. Calculate the capital. Assume the normal model and independence hold over both years.

Method. Aggregate the annual deviations over the specified horizon, then apply the confidence-level quantile to the aggregate standard deviation. Because these are deviations from expected loss, the capital is the quantile of the deviation sum.

The annual standard deviation is $\sqrt{25}=5$ million. For two independent years, the variance of the summed deviation is $2(25)=50$, so its standard deviation is $\sqrt{50}=7.071068$ million (shown to six decimals).

The two-year unexpected-loss capital is $2.5758\sqrt{50}=18.213656\ldots$ million. Rounded to three decimals, the answer is $18.214 million.

Check. The result uses two years of variance, not two years of standard deviation: $2.5758\sqrt{50}$ is the same as $2.5758(5\sqrt{2})$. Multiplying the annual capital by two would incorrectly assume a different aggregation rule.

Calibrating a Normal Model to a Target

Suppose the modeled portfolio loss is normal and the target is unexpected-loss capital at confidence level $q$. If the portfolio loss has mean $\mu$ and standard deviation $\sigma$, then its capital is $z_q\sigma$. If the target is $K^*$, a simple volatility calibration uses the factor $c=K^*/(z_q\sigma)$. Applying this factor to the modeled standard deviation makes the calibrated capital equal to the target, provided the mean remains unchanged and the normal model is retained.

For a portfolio made up of components, proportional scaling of every component's standard deviation by the same factor preserves their relative standard-deviation proportions. Under a fixed dependence structure, it also scales the aggregate standard deviation by that factor. It does not preserve absolute volatility, and it does not validate the dependence assumptions or the normal tail. Scaling selected components differently can change their relative importance and the aggregate result.

Worked Example: Scaling Component Volatility to a Capital Target

A hypothetical portfolio has two independent normal loss components, measured in millions of dollars: $X_1\sim N(10,9)$ and $X_2\sim N(20,16)$. The second parameter is the variance. The target is unexpected-loss capital of $15 million at $q=0.99$, using $z_{0.99}=2.3263$. Preserve both component means and calibrate both standard deviations by one common factor.

Method. First calculate the modeled aggregate standard deviation and capital. Then divide the target by the modeled capital to obtain the common volatility factor.

Independence gives aggregate mean $10+20=30$ million and aggregate variance $9+16=25$ million. Thus the aggregate standard deviation is $\sqrt{25}=5$ million, and the initial unexpected-loss capital is $2.3263(5)=11.6315$ million.

The required factor is $c=15/11.6315=1.2896015\ldots$. The original component standard deviations are $3$ and $4$ million. After scaling, they are $3c=3.8688045\ldots$ and $4c=5.1584061\ldots$ million. To four decimals, these are 3.8688 and 5.1584 million.

The calibrated aggregate standard deviation is $5c=6.4480076\ldots$ million, or 6.4480 million to four decimals. Using unrounded values, its capital is $2.3263(6.4480076\ldots)=15.0000$ million to four decimals, as required.

Check. The means remain 10 and 20 million; the factor adjusts standard deviations, not means. The common factor preserves the original standard-deviation ratio $3:4$. Rounding the factor or component standard deviations before calculating the final capital can create a small apparent mismatch.

Scaling a Distribution Without Hiding Its Shape

A positive multiplicative adjustment has a precise effect. If $Y=cX$ for $c>0$, then $F_Y(y)=F_X(y/c)$, so $\operatorname{VaR}_q(Y)=c\operatorname{VaR}_q(X)$ and $E[Y]=cE[X]$, when the mean exists. Therefore $K_q(Y)=cK_q(X)$. The distribution's quantiles all scale together: ratios between positive quantiles are unchanged. This is why whole-distribution scaling preserves relative tail shape, even though it changes the level of losses.

That preservation is conditional on multiplying the entire distribution. Adjusting only one tail percentile, changing the mean separately, or scaling some model components but not others is not equivalent. Those changes can alter quantile ratios, the contribution of components, or both. A calibration to one target quantile alone cannot determine whether the rest of the tail is plausible.

Worked Example: Calibrating an Empirical Loss Distribution

A hypothetical model has the following one-period loss distribution, in millions of dollars: loss 0 with probability 0.50, loss 10 with probability 0.30, loss 20 with probability 0.15, and loss 40 with probability 0.05. The target is unexpected-loss capital of $18 million at $q=0.95$. Calibrate by multiplying every loss outcome by the same positive factor.

Method. Find the model's mean and $0.95$-quantile first. Their difference is the uncalibrated unexpected-loss capital; the ratio of target to modeled capital gives the whole-distribution scaling factor.

The cumulative probabilities at losses 0, 10, 20, and 40 are 0.50, 0.80, 0.95, and 1.00. By the quantile definition, $\operatorname{VaR}_{0.95}(X)=20$ million. The mean is $E[X]=0(0.50)+10(0.30)+20(0.15)+40(0.05)=0+3+3+2=8$ million.

The modeled unexpected-loss capital is $20-8=12$ million. Hence $c=18/12=1.5$. The calibrated mean is $1.5(8)=12$ million, and the calibrated $0.95$-quantile is $1.5(20)=30$ million. The resulting capital is $30-12=18$ million, exactly the target.

As a shape check, the original $0.99$-quantile is 40 million, because the cumulative probability remains 0.95 below that outcome. After scaling, the $0.99$-quantile is 60 million. The ratio of the $0.99$-quantile to the $0.95$-quantile is 2 both before and after scaling. The calculation therefore meets the capital target while retaining the empirical distribution's relative quantile shape.

Check. The target is unexpected capital, not a target $0.95$-quantile of 18 million. Setting the latter would calibrate a different quantity.

Calibration Controls and Common Traps

Before applying a factor, document the target's basis. In particular, confirm that the modeled and target losses use the same horizon and loss definition, and that both use the same confidence level and capital convention. A target for net losses cannot be reconciled by changing a gross-loss model's scale without first aligning the bases. Likewise, a target stated as a gross quantile should not be compared directly with unexpected-loss capital.

A common exam trap is to scale a quantile but forget that the mean changes when the entire loss distribution is scaled. For a whole-distribution multiplier $c$, the correct calculation is $K_q(cX)=c\operatorname{VaR}_q(X)-cE[X]=cK_q(X)$. It is wrong to use $c\operatorname{VaR}_q(X)-E[X]$ unless the mean is deliberately held fixed by a different model adjustment. In the normal component example, the adjustment applies to volatility alone, so holding the means fixed is explicit; in the empirical example, every loss outcome is scaled, so the mean scales too.

Key takeaway: A calibration factor can force agreement at a chosen capital target, but agreement is not evidence of model validity. State the target basis and horizon, derive the factor from the defined capital measure, apply it consistently, and assess how it affects the full tail and each material component.

Check Your Understanding

For each problem, use the stated model and capital definition. Answers are provided for checking.

Annual loss deviations are independent $N(0,9)$ random variables, in millions. Find four-year unexpected-loss capital at $q=0.975$, using $z_{0.975}=1.96$. (Answer: $1.96\sqrt{4(9)}=$$11.760 million.)
Annual losses are independent and identically distributed $N(8,16)$ million, with the second parameter denoting variance. At $q=0.95$, use $z_{0.95}=1.645$. For a three-year horizon, calculate unexpected-loss capital and gross quantile capital. (Answer: unexpected capital $=1.645\sqrt{3(16)}=$$11.3969 million; gross quantile capital $=3(8)+11.3969=$$35.3969 million, with intermediate capital rounded to four decimals.)
Two independent normal loss components have standard deviations 3 and 4 million. Their means are fixed, and a common volatility factor is to achieve $15 million of unexpected-loss capital at $q=0.99$, using $z_{0.99}=2.3263$. Find the factor and calibrated component standard deviations. (Answer: $c=15/[2.3263(5)]=1.2896015\ldots$; standard deviations $3.8688$ and $5.1584$ million to four decimals.)
A discrete loss model has outcomes 0, 10, 20, and 40 million with probabilities 0.50, 0.30, 0.15, and 0.05. A target of $24 million is unexpected-loss capital at $q=0.95$. If every outcome is scaled by one common factor, find the factor, calibrated mean, and calibrated $0.95$-quantile. (Answer: initial mean $=8$, initial quantile $=20$, initial capital $=12$; factor $=2$, calibrated mean $=$16$ million and quantile $=$40$ million.)
Annual independent normal losses have mean $6 million and standard deviation $2 million. For a two-year horizon at $q=0.95$, use $z_{0.95}=1.645$. State both unexpected-loss and gross-quantile capital, and explain which one uses the mean. (Answer: unexpected capital $=1.645\sqrt{2(2^2)}=$$4.6528 million; gross quantile capital $=2(6)+4.6528=$$16.6528 million; the gross quantile includes the two-year expected loss.)

What You'll Learn