Confidence Levels in Capital Estimates

What a Capital Confidence Level Says

A capital estimate is easier to interpret when its confidence level is stated as a probability about losses. For example, a 99% level is not a statement that a company is “99% safe” in every sense. It describes how often a specified loss threshold is expected to cover losses under a particular model, on a defined loss basis and over a defined horizon.

The loss basis matters: it must be clear which losses are included and whether the threshold is for total losses or for losses above expected loss. The horizon matters too. In this tutorial, examples use a fixed one-year horizon; a probability for that horizon should not be transferred to a different period without a suitable model.

Definition: Let $L$ be the random variable for losses over a specified horizon, measured on a stated basis, and let $C$ be a loss threshold. The model-based coverage probability of $C$ is $P(L \leq C)$. A target confidence level $p$ means the threshold is selected so that $P(L \leq C) \geq p$. The complementary probability $P(L > C)$ is the probability that losses exceed the threshold under the same model and assumptions.

A confidence level belongs to the combination of a loss model, a loss definition, a horizon, and a threshold. It is not an intrinsic property of the company or of the dollar amount alone. For example, the same $30 million threshold can have different coverage probabilities under two different loss models.

From a Confidence Level to a Probability Statement

For a continuous loss distribution, define the threshold associated with level $p$ as the $p$-quantile:

q_p(L)=\inf\{c:F_L(c)\geq p\}, \qquad F_L(c)=P(L\leq c).

The definition chooses the smallest threshold whose cumulative probability reaches or exceeds $p$. If the distribution is continuous and its cumulative distribution function is strictly increasing at that threshold, then $P(L\leq q_p(L))=p$. The probability of exceeding the threshold is therefore $1-p$.

This equality need not hold for a discrete loss distribution. Its cumulative probability can jump from below $p$ to above $p$ at a possible loss amount. The selected threshold still reaches the target, but its actual coverage can be greater than the target.

Important distinction: If a capital amount $K$ is defined as a buffer for unexpected loss, the relevant total-loss threshold may be $E[L]+K$. The confidence statement is then $P(L\leq E[L]+K)\geq p$. If the amount is instead already a threshold for total losses, the statement is $P(L\leq C)\geq p$. Always identify which convention is being used before interpreting a confidence level.

Worked Examples

Worked Example: Interpreting Coverage Under a Normal Loss Model

A hypothetical insurer models its one-year aggregate loss $L$, in millions of dollars, as $L\sim N(20,25)$. The first parameter is the mean and the second is the variance, so the standard deviation is $5$ million dollars. The model is assumed to apply to the stated one-year loss basis, with parameters treated as known. What coverage probability is provided by a $30 million total-loss threshold, and what does that imply at a 95% target level?

Method. Standardize the threshold and use the standard normal cumulative distribution function $\Phi$, where $\Phi(z)=P(Z\leq z)$ for $Z\sim N(0,1)$.

The standard deviation is $\sqrt{25}=5$. The standardized threshold is

z=\frac{30-20}{5}=2.

Thus the coverage probability is $P(L\leq 30)=\Phi(2)=0.9772$, to four decimals. The probability of exceeding the threshold is $1-0.9772=0.0228$. This threshold provides 97.72% model-based coverage, which is greater than the 95% target.

Check. The two probabilities sum to $0.9772+0.0228=1.0000$, as required for the events $L\leq 30$ and $L>30$.

Common mistake avoided. Calling the threshold “95% capital” just because it exceeds a 95% target would obscure the model’s actual result. Its calculated coverage is 97.72%, not 95%.

Worked Example: A Discrete Model Can Exceed Its Target

A hypothetical specialty insurer models its one-year aggregate loss, in millions of dollars, using the following distribution. Losses take only the listed values, and the probabilities are assumed to be complete and mutually exclusive. Find the smallest threshold that meets a 95% confidence target, then state its actual coverage.

One-year loss	Probability	Cumulative probability
$0.5 million	0.50	0.50
$1.5 million	0.30	0.80
$3.0 million	0.17	0.97
$6.0 million	0.03	1.00

Method. Add the probabilities of losses at or below each candidate threshold. At $1.5 million, coverage is $0.50+0.30=0.80$, which is below 0.95. At $3.0 million, coverage is $0.50+0.30+0.17=0.97$, which reaches the target. Therefore the smallest qualifying threshold is $3.0 million.

The actual coverage is 97%, not exactly 95%. The probability of exceeding $3.0 million is the probability of the $6.0 million outcome, or 0.03. The jump from 80% to 97% occurs because the loss distribution assigns a probability of 17% to exactly $3.0 million.

Check. $P(L\leq 3.0)+P(L>3.0)=0.97+0.03=1.00$. The next lower possible loss threshold, $1.5 million, provides only 80% coverage, confirming that $3.0 million is the smallest threshold meeting the target.

Common mistake avoided. Reporting “exactly 95% coverage” at $3.0 million ignores the probability mass at that loss amount. The threshold meets or exceeds the target; the model’s actual coverage is 97%.

Worked Example: Comparing Coverage at Different Thresholds

A hypothetical insurer uses the same one-year loss model $L\sim N(10,4)$, with losses and parameters in millions of dollars. The mean is $10 million and the standard deviation is $\sqrt{4}=$2$ million. Compare the model-based coverage of total-loss thresholds of $12 million, $14 million, and $16 million.

Method. Standardize each threshold using $z=(C-10)/2$, then read the corresponding standard normal cumulative probability.

For $C=12$, $z=(12-10)/2=1$, so $P(L\leq 12)=\Phi(1)=0.8413$. For $C=14$, $z=(14-10)/2=2$, so $P(L\leq 14)=\Phi(2)=0.9772$. For $C=16$, $z=(16-10)/2=3$, so $P(L\leq 16)=\Phi(3)=0.9987$. Probabilities are rounded to four decimals.

The respective coverage levels are 84.13%, 97.72%, and 99.87%. Their corresponding exceedance probabilities are $1-0.8413=0.1587$, $1-0.9772=0.0228$, and $1-0.9987=0.0013$. Each higher threshold covers a larger portion of the modeled loss distribution.

Check. The thresholds increase in order, and so do their calculated coverage probabilities. Each coverage probability plus its exceedance probability equals 1.0000 after rounding.

Common mistake avoided. A confidence level is not the percentage of expected loss held as capital. For instance, $12 million is 120% of the $10 million mean, but its model-based coverage is 84.13%, not 120%.

Why the Quantile Statement Holds

The probability interpretation follows directly from the cumulative distribution function. By definition, $F_L(c)$ is the probability that loss is at or below $c$. The threshold $q_p(L)$ is the smallest $c$ for which this cumulative probability reaches $p$. Therefore:

F_L(q_p(L))=P(L\leq q_p(L))\geq p.

For a continuous distribution with no probability mass at the threshold and a cumulative distribution function that reaches $p$ exactly, the inequality becomes equality. Since $L\leq q_p(L)$ and $L>q_p(L)$ are complementary events, their probabilities add to 1:

P(L>q_p(L))=1-P(L\leq q_p(L))=1-p.

For a discrete distribution, the first inequality may be strict because the cumulative probability can jump over $p$. This is why “at least the target confidence level” is the generally correct wording.

What the Confidence Level Does—and Does Not—Mean

A model-based confidence level is conditional on the chosen assumptions. It describes the probability assigned by the loss model to the event that losses do not exceed the specified threshold. It does not mean that losses are guaranteed to remain below that amount, nor does it mean the company has that percentage chance of surviving in every possible sense.

The probability statement also does not automatically describe the accuracy of the model itself. A stated 99% level does not account for model error, changes in the exposure mix, or uncertainty in estimated parameters unless those sources of uncertainty have been included in the model. The result should therefore be read as a calibrated statement under stated assumptions, not as certainty about future outcomes.

Finally, the loss basis must match the capital definition. If the threshold concerns total losses, interpret $P(L\leq C)$. If capital is an unexpected-loss buffer $K$, identify the corresponding total-loss threshold before stating coverage. Mixing these conventions can shift the threshold by expected loss and materially change the probability statement.

Key takeaway: A confidence level $p$ means that, under a specified loss model and over a specified horizon, the chosen threshold covers at least a proportion $p$ of modeled outcomes. For continuous distributions coverage may equal $p$; for discrete distributions it can be higher. Always state the loss basis, horizon, and model assumptions.

Check Your Understanding

For each problem, interpret the result on the stated loss basis and horizon. Probabilities in the answers are rounded to four decimals where applicable.

A hypothetical insurer models one-year losses in thousands of dollars as $L\sim N(50,100)$. The mean is 50 and the variance is 100. For a total-loss threshold of $70 thousand, calculate the coverage and exceedance probabilities. (Answer: Coverage $0.9772$; exceedance $0.0228$, or 97.72% and 2.28%.)
A hypothetical portfolio has one-year losses $L\sim N(8,4)$, in millions of dollars, with known parameters and no additional uncertainty included. Compare the coverage at thresholds of $10 million and $12 million. (Answer: At $10 million, $z=1$ and coverage is $0.8413$; at $12 million, $z=2$ and coverage is $0.9772$.)
A discrete one-year loss model assigns probabilities 0.60, 0.25, and 0.15 to losses of $1 million, $4 million, and $9 million, respectively. Find the smallest threshold meeting an 80% target and its actual coverage. (Answer: $4 million; actual coverage $0.60+0.25=0.85$, or 85%.)
A model gives a 99% coverage level for a threshold of $25 million on a specified one-year total-loss basis. State the model-based exceedance probability and one conclusion that cannot be drawn from this statement alone. (Answer: Exceedance probability $1-0.99=0.01$, or 1%; it does not guarantee that actual losses will stay below $25 million or establish solvency under all circumstances.)
A hypothetical insurer reports unexpected-loss capital $K=$6 million$, with expected one-year loss $E[L]=$14 million. Under its stated model, \(P(L\leq E[L]+K)=0.96$. Identify the total-loss threshold and state the model-based coverage and exceedance probabilities. (Answer: The total-loss threshold is $20 million; coverage is 0.96 and exceedance is 0.04, or 96% and 4%.)

What You'll Learn