Indicator Variables For Binary Risk Events

Why Represent Risk Events With Indicators?

Many risk questions are about whether a loss crosses a threshold, rather than the precise size of every loss. An insurer may ask whether two lines each exceed a specified loss level during the same year, or how variable the number of risk limits breached will be. An indicator variable turns each yes-or-no event into a random variable that can be analyzed with expectations, variances, and covariances.

This representation is especially useful for dependence. Marginal exceedance probabilities describe each risk event separately; a joint exceedance probability describes how often both occur. Their difference from what independence would imply gives the covariance of the indicators.

Definition: For an event $A$, its indicator random variable is $I_A$, defined by $I_A=1$ if $A$ occurs and $I_A=0$ otherwise. For a loss $X$ and threshold $x$, the threshold-exceedance indicator is $I_{\{X>x\}}$. Its event probability is $p=P(A)$.

An indicator is a Bernoulli random variable: it takes only the values 0 and 1. Its mean is the probability of the event, and its variance is determined by that same probability. For an event with probability $p$, the probability of the indicator being 1 is $p$, and the probability of it being 0 is $1-p$.

E[I_A]=p,\qquad \operatorname{Var}(I_A)=p(1-p).

For a loss threshold, $p=P(X>x)=S_X(x)$ when $S_X(x)$ denotes the probability of exceeding $x$. When a loss distribution has point masses, use the event probability specified by the strict inequality $X>x$; do not silently replace it with $X\ge x$. The distinction matters if the loss can equal the threshold with positive probability.

Two Binary Events: Joint Probability and Covariance

Let $A$ and $B$ be two risk events, with indicators $I_A$ and $I_B$. Write $p_A=P(A)$, $p_B=P(B)$, and $q=P(A\cap B)$. The joint probability $q$ records how often both events occur. The covariance compares that joint probability with the product of the marginal probabilities.

Theorem: For two event indicators, $\operatorname{Cov}(I_A,I_B)=P(A\cap B)-P(A)P(B)$. If both event probabilities are strictly between 0 and 1, their correlation is \[ \operatorname{Corr}(I_A,I_B)= \frac{P(A\cap B)-P(A)P(B)} {\sqrt{P(A)(1-P(A))P(B)(1-P(B))}}. \] If either event has probability 0 or 1, its indicator has zero variance and the correlation is undefined.

The covariance formula follows directly from the definition of covariance. The product $I_AI_B$ equals 1 exactly when both events occur, and equals 0 otherwise. Thus its expected value is the probability of their intersection. Subtracting the product of the two indicator means gives the result:

\begin{aligned} \operatorname{Cov}(I_A,I_B) &=E[I_AI_B]-E[I_A]E[I_B]\\ &=P(A\cap B)-P(A)P(B). \end{aligned}

If $A$ and $B$ are independent, then $P(A\cap B)=P(A)P(B)$, so their indicator covariance is zero. For two binary events, the converse also holds: zero indicator covariance means their joint probability equals the product of the marginals, which is the independence condition for the pair. A positive covariance means joint occurrence is more likely than under independence; a negative covariance means it is less likely.

The joint probability must also be feasible given the marginals. In particular, the probability of both events cannot exceed the probability of either event, and it cannot be less than zero. Also, the probability of their union cannot exceed 1. These requirements give the bounds below.

\max(0,p_A+p_B-1)\leq P(A\cap B)\leq \min(p_A,p_B).

These bounds are useful checks, not a dependence model. Marginal probabilities alone generally do not determine the joint probability. When $p_B>0$, the conditional probability $P(A\mid B)$ can be found from the joint probability by dividing by $p_B$. It is not the same quantity as $P(A\cap B)$.

Worked Example: Two Annual Loss Events

A hypothetical insurer studies two annual events: $A$, a property loss exceeds $4 million, and $B$, a cyber loss exceeds $2 million. The specified joint distribution is: both events, 0.08; $A$ only, 0.12; $B$ only, 0.22; neither, 0.58. The four probabilities sum to 1.00. The horizon is one year; no losses are discounted, so interest is not used, and no mortality assumption applies. Use the stated joint distribution; no further loss-severity distribution is assumed. Find the indicator covariance, correlation, and $P(A\mid B)$.

Method and marginals. The indicators are $I_A$ and $I_B$. Add the relevant cells to obtain the marginal event probabilities:

p_A=0.08+0.12=0.20,\qquad p_B=0.08+0.22=0.30,\qquad P(A\cap B)=0.08.

Covariance. Subtract the product of the marginals from the joint probability:

\operatorname{Cov}(I_A,I_B)=0.08-(0.20)(0.30)=0.08-0.06=0.02.

Correlation. First calculate the indicator variances: $0.20(0.80)=0.16$ and $0.30(0.70)=0.21$. Therefore the correlation is

\frac{0.02}{\sqrt{(0.16)(0.21)}}= \frac{0.02}{\sqrt{0.0336}}= \frac{0.02}{0.183303\ldots}=0.1091.

Conditional probability. Since $P(B)=0.30>0$, $P(A\mid B)=0.08/0.30=0.2667$. The joint probability is 0.0800, while the conditional probability is 0.2667.

Check and common mistake. The joint probability 0.08 is between 0 and $\min(0.20,0.30)=0.20$, as required. The positive covariance is consistent with the joint probability exceeding the independence value $0.20(0.30)=0.06$. A common error is to report 0.08 as $P(A\mid B)$; the correction is to divide the intersection probability by $P(B)$.

Threshold Exceedances and Event Counts

For a pair of losses $X$ and $Y$, choose thresholds $x$ and $y$, and define $A=\{X>x\}$ and $B=\{Y>y\}$. Then $I_A$ and $I_B$ measure whether each threshold is crossed. Their covariance captures dependence specifically at those thresholds. It need not describe all aspects of dependence between the underlying loss amounts.

A related use is counting how many events occur. For $n$ event indicators, let $N=\sum_{j=1}^n I_{A_j}$. Linearity of expectation gives the expected count, whether or not the events are independent. The variance, however, includes the covariances between pairs.

E[N]=\sum_{j=1}^n P(A_j),\qquad \operatorname{Var}(N)= \sum_{j=1}^n \operatorname{Var}(I_{A_j}) +2\sum_{j<k}\operatorname{Cov}(I_{A_j},I_{A_k}).

To justify the variance expression, expand the variance of the sum. Each individual variance appears once, and each distinct pair contributes twice because $\operatorname{Var}(U+V)=\operatorname{Var}(U)+\operatorname{Var}(V)+2\operatorname{Cov}(U,V)$. The expectation requires no independence assumption; omitting the covariance terms from the variance does require zero covariance for every pair.

Worked Example: Variability in the Number of Breaches

A hypothetical risk team tracks three annual indicators: $I_1$ is a reserve shortfall, $I_2$ is a vendor outage, and $I_3$ is a claims-processing delay. Their probabilities are $0.10$, $0.20$, and $0.30$. The pairwise joint probabilities are $P(A_1\cap A_2)=0.04$, $P(A_1\cap A_3)=0.05$, and $P(A_2\cap A_3)=0.09$. The horizon is one year; no discounting or mortality assumption is involved. Assume a joint distribution with no three-way occurrence. Find the expected number of events and its variance.

Method and compatibility check. With no three-way occurrence, the probabilities of exactly each pair are 0.04, 0.05, and 0.09. The probabilities of exactly one event are $0.10-0.04-0.05=0.01$, $0.20-0.04-0.09=0.07$, and $0.30-0.05-0.09=0.16$. The probability of no events is $1-(0.04+0.05+0.09+0.01+0.07+0.16)=0.58$. These nonnegative cells sum to 1, so the stated joint distribution is feasible.

Expected count. Let $N=I_1+I_2+I_3$. By linearity of expectation,

$$ E[N]=0.10+0.20+0.30=0.60. $$

Individual variances. The three indicator variances are $0.10(0.90)=0.09$, $0.20(0.80)=0.16$, and $0.30(0.70)=0.21$. Their sum is $0.46$.

Pairwise covariances. Subtract each product of marginals from its joint probability:

\begin{aligned} \operatorname{Cov}(I_1,I_2)&=0.04-(0.10)(0.20)=0.02,\\ \operatorname{Cov}(I_1,I_3)&=0.05-(0.10)(0.30)=0.02,\\ \operatorname{Cov}(I_2,I_3)&=0.09-(0.20)(0.30)=0.03. \end{aligned}

Variance of the count. Add twice the sum of the three covariances to the sum of the individual variances:

\operatorname{Var}(N)=0.46+2(0.02+0.02+0.03)=0.46+0.14=0.60.

Check and common mistake. From the compatible joint distribution, $P(N=1)=0.24$, $P(N=2)=0.18$, and $P(N=0)=0.58$. Thus $E[N^2]=0.24+4(0.18)=0.96$, and $E[N^2]-(E[N])^2=0.96-0.36=0.60$, agreeing with the covariance calculation. Treating the events as independent would incorrectly omit the positive covariance terms and give variance $0.46$.

Interpreting the Measure and Avoiding Traps

Indicator covariance is expressed in probability units, while indicator correlation is unitless and lies between $-1$ and 1 when defined. The covariance depends on the event probabilities as well as their dependence; the same qualitative association can yield different covariance values for events with different marginal probabilities. For that reason, state the thresholds and the marginal probabilities when interpreting a binary-risk dependence calculation.

For event indicators, the joint-probability formula gives a direct interpretation: compare the observed or specified $P(A\cap B)$ with $P(A)P(B)$. Do not infer independence merely because each event has a known marginal probability. Conversely, a correlation calculated from indicators is only a summary for those particular binary events; it is not automatically the correlation of the underlying loss amounts.

Key takeaway: Define the event before calculating. For two indicators, covariance is joint probability minus the product of marginal probabilities. For an event count, include twice every pairwise covariance in its variance; independence must not be assumed without support.

Check Your Understanding

For each problem, define the event indicators first and keep joint probabilities distinct from conditional probabilities.

A hypothetical annual event has probability $0.07$. Find the expectation and variance of its indicator. (Answer: $E[I]=0.0700$, $\operatorname{Var}(I)=0.0651$.)
Two binary risk events have probabilities $0.18$ and $0.25$, and their joint probability is $0.06$. Find their covariance and $P(A\mid B)$. (Answer: covariance $=0.0150$; $P(A\mid B)=0.2400$.)
Two events have probabilities $0.35$ and $0.50$. Give the feasible range for their joint probability and the corresponding feasible range for their indicator covariance. (Answer: joint probability from $0$ to $0.35$; covariance from $-0.175$ to $0.175$.)
Three event indicators have marginal probabilities $0.20,0.25,0.40$, and pairwise joint probabilities $0.08,0.12,0.15$, respectively. Find the variance of their sum. (Answer: $0.8275$.)
Two loss thresholds are exceeded with probabilities $0.12$ and $0.08$. If $P(A\mid B)=0.50$, find the joint probability and indicator covariance. (Answer: joint probability $=0.0400$; covariance $=0.0304$.)

What You'll Learn