Confidence Levels, Time Horizons, And Risk Calibration

Why Calibration Choices Matter

A solvency capital requirement is meaningful only in relation to the loss measure, confidence level, and time horizon used to calibrate it. A requirement calibrated to a higher confidence level represents a more extreme loss quantile under the same loss distribution. A requirement calibrated over a longer horizon may also be larger, but the size of that change depends on how losses evolve and aggregate over time.

These choices affect interpretation, not just arithmetic. A one-year requirement does not automatically describe protection over several years, and a percentage such as 99.5% does not mean that a firm is certain to remain solvent for a fixed number of years. The probability statement applies to a specified loss measure, distribution, confidence level, and horizon. This distinction is aligned with the CAS Exam 9 (Risk Management for Actuaries) syllabus.

Definition: For a loss $L$ measured over a specified time horizon, its value-at-risk at confidence level $q$, where $0<q<1$, is the $q$-quantile: $$ \operatorname{VaR}_q(L)=\inf\{x:F_L(x)\ge q\}. $$ The confidence level $q$ is the probability threshold used to select the quantile; the corresponding upper-tail probability is $1-q$. In a discrete distribution, the quantile is the smallest loss value at which the cumulative probability reaches or exceeds $q$.

A common capital convention is to measure capital above expected loss. Under that convention, the capital amount associated with a quantile is the loss quantile less the expected loss. This separates the average loss from the additional amount intended to cover adverse deviations. Regulatory frameworks can define their capital measures differently, so the convention must be stated rather than assumed.

Definition: For this tutorial, quantile-based capital at level $q$ for loss $L$ is $$ K_q(L)=\operatorname{VaR}_q(L)-E[L]. $$ This is a risk-capital measure above expected loss, not the total loss quantile itself. It is useful when expected losses are treated separately. If a rule defines its requirement as the full quantile or uses another adjustment, that rule’s definition takes precedence.

Confidence Level Changes the Tail Severity

For a fixed loss distribution and horizon, increasing $q$ cannot decrease the quantile: a higher confidence level requires looking farther into the upper tail. The change need not be proportional to the change in $q$. It depends on the shape of the loss distribution, particularly how rapidly losses increase in its tail.

For example, suppose a hypothetical one-year loss $L$, in thousands of dollars, is normally distributed with mean $\mu$ and standard deviation $\sigma$. Under this model, and only when the normal distribution is being used for the loss over that horizon,

L\sim N(\mu,\sigma^2), \qquad \operatorname{VaR}_q(L)=\mu+z_q\sigma, \qquad K_q(L)=z_q\sigma,

where $z_q$ is the $q$-quantile of the standard normal distribution. The second parameter in $N(\mu,\sigma^2)$ is the variance, so $\sigma$ is its standard deviation. The final equality follows because $E[L]=\mu$. This normal-model result does not establish that a normal distribution is appropriate for a particular insurer’s losses.

Worked Example: Comparing Two One-Year Confidence Levels

A hypothetical insurer models its one-year loss, in thousands of dollars, as $L\sim N(100,20^2)$. Compare the 99% and 99.5% loss quantiles and capital amounts above expected loss. Use $z_{0.99}=2.3263$ and $z_{0.995}=2.5758$. Report the final amounts to two decimal places in thousands of dollars.

Method and assumptions. The model specifies a normal one-year loss with mean $100$ and standard deviation $20$. Use the normal quantile formula; under the stated capital convention, subtract the mean loss from each quantile.

At the 99% confidence level:

\operatorname{VaR}_{0.99}(L)=100+2.3263(20) =100+46.526 =146.526.

The expected loss is $E[L]=100$, so capital above expected loss is

K_{0.99}(L)=146.526-100=46.526.

At the 99.5% confidence level:

\operatorname{VaR}_{0.995}(L)=100+2.5758(20) =100+51.516 =151.516,

and

K_{0.995}(L)=151.516-100=51.516.

Rounded only at the final step, the 99% quantile is $146.53 thousand and capital is $46.53 thousand; the 99.5% quantile is $151.52 thousand and capital is $51.52 thousand. The higher confidence level increases the quantile and the capital by $51.516-46.526=4.990$ thousand dollars.

Check and common mistake. Each capital amount is smaller than its corresponding total loss quantile by exactly the expected loss, $100$ thousand dollars. Treating the 99.5% quantile, $151.516$, as capital above expected loss would fail to subtract that mean; it would confuse the total loss quantile with the stated capital measure.

Time Horizon and the Conditions for Scaling

A time horizon specifies the period over which the loss variable is measured. To compare horizons, define the cumulative loss as the sum of losses over the periods in the horizon. A longer horizon can include more ordinary loss as well as more opportunity for adverse outcomes. The effect on a quantile depends on the loss process, including its distribution and dependence over time.

A useful special case is a sum of independent, identically distributed normal period losses. Suppose each period loss $L_t$ has mean $\mu$ and variance $\sigma^2$, and the period losses are independent. For an integer horizon $H$, define total loss as $L^{(H)}=\sum_{t=1}^{H}L_t$. Then the total has mean $H\mu$ and variance $H\sigma^2$, so its standard deviation is $\sqrt{H}\sigma$. Under those assumptions:

\operatorname{VaR}_q(L^{(H)})=H\mu+z_q\sqrt{H}\sigma, \qquad K_q(L^{(H)})=z_q\sqrt{H}\sigma.

The mean scales with $H$, while the capital above expected loss scales with $\sqrt{H}$. The latter scaling is not a general rule for solvency capital. It relies here on independent normal period losses with identical means and variances, a sum of losses, a fixed confidence level, and the stated capital convention.

Define the period loss.
Specify what counts as a loss in one period, its units, and the distributional assumptions.

Aggregate over the horizon.
For a cumulative-loss measure, sum period losses and account for their dependence. Do not assume independence without a basis.

Apply the chosen confidence level.
Find the quantile of the loss over the full horizon, not a quantile for a different horizon.

Translate the quantile into the stated capital measure.
If capital is defined above expected loss, subtract expected loss over the same horizon.

Worked Example: Two-Year Capital Under Independent Normal Losses

A hypothetical portfolio has annual losses $L_1$ and $L_2$, each normally distributed in millions of dollars with mean $12$ and variance $64$. Assume the two annual losses are independent. Find the two-year aggregate 99.5% loss quantile and capital above expected two-year loss. Use $z_{0.995}=2.5758$. Also compare the correct two-year capital with twice the one-year capital. Round final amounts to two decimal places in millions of dollars.

Method and assumptions. Because the losses are independent normal variables, their sum is normal. Its mean is the sum of the annual means, and its variance is the sum of the annual variances.

The two-year expected loss is

$$ E[L_1+L_2]=12+12=24. $$

The two-year variance and standard deviation are

\operatorname{Var}(L_1+L_2)=64+64=128, \qquad \sqrt{128}=11.3137\text{ (approximately)}.

Therefore the aggregate loss quantile is

\operatorname{VaR}_{0.995}(L_1+L_2) =24+2.5758(11.3137) \approx24+29.1418 =53.1418.

Capital above expected two-year loss is $53.1418-24=29.1418$, or $29.14 million after rounding. For one year, capital above expected loss is $2.5758(8)=20.6064$ million. Twice that amount is $41.2128$ million, or $41.21 million. Thus, under the specified assumptions, two-year capital is $\sqrt{2}$ times one-year capital, not twice one-year capital.

Check and common mistake. The two-year total quantile, $53.1418$, exceeds the two-year expected loss, $24$, by the calculated capital amount, $29.1418$. Doubling one-year capital incorrectly treats independent-period uncertainty as if standard deviations added directly; here variances add, producing a square-root-of-two increase in standard deviation.

Why Horizon Scaling Can Fail

The normal independent-loss calculation is a conditional result, not a default conversion rule. If annual losses are dependent, their aggregate variance includes covariance terms. If losses are skewed, heavy-tailed, discrete, or affected by changing conditions, the normal quantile formula may not describe the aggregate tail. Even when a normal approximation is used, its quality at a high confidence level must be assessed rather than presumed.

The quantile definition also matters for discrete losses. The quantile is the smallest value whose cumulative probability reaches the target. It can jump when the confidence level crosses a probability threshold. The following hypothetical example shows both a horizon effect and a confidence-level effect without invoking normality.

Worked Example: A Discrete Catastrophe Loss Over Two Years

In a hypothetical model, each year has a loss of $10 million with probability $0.02$, and zero otherwise. Assume annual losses are independent. Calculate the annual 99% quantile and capital above expected annual loss, then calculate the corresponding two-year amounts. Also determine the two-year 99.95% quantile and capital. Report final amounts in millions of dollars.

Method and assumptions. Enumerate the possible outcomes and their probabilities. For the two-year total, independence gives probability $0.98^2=0.9604$ of no event, $2(0.98)(0.02)=0.0392$ of exactly one event, and $0.02^2=0.0004$ of two events.

For one year, the expected loss is

$$ E[L]=0(0.98)+10(0.02)=0.2. $$

The annual cumulative probability at zero is $0.98$, below $0.99$, and at $10 million it is $1$. Therefore $\operatorname{VaR}_{0.99}(L)=10$ million and $K_{0.99}(L)=10-0.2=9.8$ million.

Over two years, expected loss is $0.2+0.2=0.4$ million. The cumulative probability at zero is $0.9604$, and at $10 million it is $0.9604+0.0392=0.9996$. Since $0.9604<0.99\le0.9996$, the two-year 99% quantile is $10 million. Its capital above expected loss is $10-0.4=9.6$ million.

At the 99.95% level, the cumulative probability at $10 million is $0.9996$, which is less than $0.9995$? No: $0.9996$ is greater than $0.9995$. Thus the 99.95% quantile remains $10 million, and capital is $10-0.4=9.6$ million. In contrast, at 99.99%, $0.9996<0.9999$, so the quantile is $20 million and capital is $20-0.4=19.6$ million.

Check and common mistake. The probabilities sum to $0.9604+0.0392+0.0004=1$, and the jump from a $10 million to a $20 million quantile occurs only when the target exceeds $0.9996$. A frequent error is to interpolate between the discrete outcomes or to assume that a longer horizon must multiply capital by a fixed factor; neither follows from this distribution.

Interpreting a Calibrated Requirement

A confidence level is not a complete description of risk. A 99.5% quantile is conditional on the specified loss definition, horizon, data, model, and assumptions. In the continuous case, it identifies a threshold with a modeled probability $1-q$ of loss exceeding that threshold; for discrete losses, the probability of exceeding the quantile can be strictly less than $1-q$ because probability mass can sit at the quantile.

When comparing two calibrations, hold the other choices in view. A difference in capital might reflect a change in confidence level, horizon, loss model, dependence assumptions, or the treatment of expected loss. If more than one choice changes, the effect cannot be attributed to confidence level alone. A useful comparison states the loss measure, horizon, confidence level, and whether the reported amount is a total quantile or capital above expected loss.

Higher confidence, same distribution and horizon: the quantile cannot decrease, though its increase depends on tail shape.
Longer horizon, same confidence: aggregate losses generally need a horizon-specific model; neither linear nor square-root scaling is universal.
Quantile versus capital: a loss quantile includes expected loss, while the convention $K_q(L)=\operatorname{VaR}_q(L)-E[L]$ removes it.
Model interpretation: calibration is a statement about a modeled distribution, not a guarantee that losses will stay below the requirement.

Key takeaway: State the loss variable, horizon, confidence level, and capital convention before interpreting a solvency calibration. Compare quantiles on a consistent basis, and use horizon scaling only when its distributional and dependence assumptions are justified.

Check Your Understanding

For each problem, identify the applicable horizon and capital convention before calculating.

A one-year hypothetical loss, in thousands of dollars, follows $L\sim N(50,10^2)$. Using $z_{0.975}=1.96$, find the 97.5% quantile and capital above expected loss. (Answer: quantile $69.60 thousand; capital $19.60 thousand.)
Annual losses are independent and normally distributed with mean $4 million and variance $9 million squared. At the 95% confidence level, use $z_{0.95}=1.6449$ to find three-year capital above expected loss. (Answer: $8.55 million.)
Annual loss is $5 million with probability $0.10$ and zero otherwise; annual losses are independent. Find the two-year 95% quantile and capital above expected loss. (Answer: quantile $5 million; capital $4 million.)
Independent annual losses are normal with mean $2 million and standard deviation $6 million. For a four-year aggregate at 99%, use $z_{0.99}=2.3263$. Find the quantile and capital above expected loss. (Answer: quantile $35.92 million; capital $27.92 million.)
A hypothetical one-year loss has a normal distribution with mean $20 thousand and standard deviation $5 thousand. Compare the 95% and 99% capital amounts, using $z_{0.95}=1.6449$ and $z_{0.99}=2.3263$. (Answer: $8.22 thousand at 95%; $11.63 thousand at 99%; the increase is $3.41 thousand.)

What You'll Learn