Why Simultaneous Exceedances Matter
A loss can be material on its own, but risk managers are often concerned with whether two losses are large at the same time. For example, a hypothetical insurer might ask whether claims from two lines can both exceed their respective planning thresholds in the same year. The probability of that joint event depends on the losses’ dependence, not just on their separate probabilities of exceeding those thresholds.
This calculation is a focused way to examine dependence at chosen thresholds. It does not, by itself, describe the entire joint distribution or establish that the losses are independent. A useful reference is the probability that would apply if the two losses were independent.
The thresholds must be interpreted as stated. “Exceeds 100” means strictly greater than 100, not greater than or equal to 100. For discrete losses, the distinction can change the answer if there is probability mass at a threshold. In this tutorial, all exceedance events use strict inequalities.
Three Routes to the Joint Probability
The calculation method depends on how the joint information is supplied. For a discrete joint law, add the probabilities of the outcome pairs satisfying both inequalities. For a continuous joint density, integrate over the region satisfying both inequalities. If the relevant conditional probability is given, use the definition of conditional probability.
A general identity gives another route. Let \(F_X(x)=P(X\leq x)\), \(F_Y(y)=P(Y\leq y)\), and \(F_{X,Y}(x,y)=P(X\leq x,Y\leq y)\). Then
This identity holds for discrete and continuous losses. Its use requires the distribution functions and joint distribution function to be evaluated at the stated thresholds; it does not require independence.
Why the Complement Identity Holds
Let \(A=\{X\leq x\}\) and \(B=\{Y\leq y\}\). The event that both losses exceed their thresholds is the complement of \(A\cup B\). The addition rule for probabilities gives \(P(A\cup B)=P(A)+P(B)-P(A\cap B)\), where \(P(A\cap B)=F_{X,Y}(x,y)\). Taking the complement therefore gives the identity.
The subtraction of \(P(A\cap B)\) is essential: without it, the outcomes where both losses are at or below their thresholds would be subtracted twice. This inclusion-exclusion reasoning also explains why the joint distribution function is needed; the marginal distributions alone do not generally determine the joint exceedance probability.
Worked Example: Adding the Qualifying Cells in a Discrete Law
A hypothetical insurer models annual losses \(X\) and \(Y\), in thousands of dollars, using the following joint probability table. Calculate \(P(X>0,\;Y>0)\), then compare it with the independence benchmark.
| \(X\) | \(Y=0\) | \(Y=3\) | \(Y=6\) | Row total |
|---|---|---|---|---|
| \(X=0\) | 0.20 | 0.15 | 0.10 | 0.45 |
| \(X=2\) | 0.10 | 0.20 | 0.05 | 0.35 |
| \(X=5\) | 0.05 | 0.05 | 0.10 | 0.20 |
| Column total | 0.35 | 0.40 | 0.25 | 1.00 |
Method. First identify the cells for which both losses are strictly positive. These are the four cells with \(X=2\) or \(5\) and \(Y=3\) or \(6\). Add their probabilities:
The row totals give \(P(X>0)=0.35+0.20=0.55\). The column totals give \(P(Y>0)=0.40+0.25=0.65\). If the losses were independent, the joint probability would be
Thus, the actual probability is \(0.4000\), compared with an independence benchmark of \(0.3575\), a difference of \(0.0425\). The table sums to 1.00, and the four selected cells account for exactly the outcomes meeting both strict threshold conditions. The common error avoided here is to include a cell where one loss equals its threshold; that would be wrong if the threshold condition were strict.
Calculating From a Joint Density
For continuous losses with joint density \(f_{X,Y}\), the probability of both exceedances is the integral of the density over the region \(X>x\) and \(Y>y\). The independence benchmark still comes from the marginal exceedance probabilities. A product density is not assumed: the joint density must be integrated as given.
Worked Example: Integrating Over a Joint Exceedance Region
Suppose dimensionless losses \(U\) and \(V\) have joint density \(f_{U,V}(u,v)=1+0.6(1-2u)(1-2v)\) on the unit square \(0\leq u\leq1,\;0\leq v\leq1\), and density zero elsewhere. Calculate \(P(U>0.25,V>0.75)\) and compare it with independence.
Check the model first. Since each factor \(1-2u\) and \(1-2v\) is between \(-1\) and \(1\), the density is at least \(1-0.6=0.4\), so it is nonnegative. Each factor has integral zero over \([0,1]\), so the density integrates to 1 over the square. Integrating over either coordinate also gives a marginal density of 1 on \([0,1]\); hence both marginals are uniform.
Integrate the joint density over \(0.25<u\leq1\) and \(0.75<v\leq1\). The constant part contributes the area of that rectangle:
For the interaction term, the two one-dimensional integrals are
The product is \((-0.1875)(-0.1875)=0.03515625\). Multiplying this by the density coefficient \(0.6\) gives \(0.02109375\). Therefore,
Because each marginal is uniform, \(P(U>0.25)=0.75\) and \(P(V>0.75)=0.25\). The independence benchmark is \(0.75(0.25)=0.1875\). Rounded to four decimal places, the actual probability is \(0.2086\), the benchmark is \(0.1875\), and the difference is \(0.0211\). The result is consistent with the integration: the interaction contribution is positive because both one-dimensional integrals are negative, so their product is positive.
Using Conditional Exceedance Information
If \(P(Y>y\mid X>x)\) is supplied and \(P(X>x)>0\), the definition of conditional probability gives a direct calculation:
This route is particularly convenient when the problem describes how often one loss exceeds its threshold among cases where the other loss already exceeds its threshold. It is important to keep the conditioning event and the target joint event in the correct order.
Worked Example: Converting a Conditional Probability
For a hypothetical portfolio, let \(A\) be the event that a technology-services loss exceeds \$2 million and \(B\) the event that a data-recovery loss exceeds \$1 million in the same year. Suppose \(P(A)=0.12\) and \(P(B\mid A)=0.30\). Calculate \(P(A\cap B)\), compare it with independence, and check that the result is feasible.
Method. Since \(P(A)>0\), multiply the conditional probability by the probability of its conditioning event:
The independence benchmark requires \(P(B)\), which is not yet stated. Suppose the portfolio model also gives \(P(B)=0.08\). Then the benchmark is \(P(A)P(B)=0.12(0.08)=0.0096\). The actual probability exceeds the benchmark by \(0.0360-0.0096=0.0264\).
A feasibility check is \(P(A\cap B)\leq \min(P(A),P(B))\). Here \(0.0360\leq\min(0.12,0.08)=0.08\), so the joint probability is within the required bound. The calculation avoids confusing \(P(B\mid A)\) with \(P(A\cap B)\): the conditional probability \(0.30\) is not itself the probability that both events occur.
Interpreting the Independence Comparison
The difference between the actual joint probability and the independence benchmark shows the direction and size of the departure for the particular pair of thresholds. A positive difference means that the two exceedance events occur together more often than they would under independence at those thresholds; a negative difference means less often. This is a statement about these events, not a complete description of all dependence between \(X\) and \(Y\).
The comparison is useful in risk assessment because simultaneous threshold breaches can have consequences that separate loss probabilities do not reveal. However, do not infer that losses are independent merely because one chosen joint exceedance probability equals the product of the marginals. Independence requires the appropriate factorization across all relevant events, not just equality at one pair of thresholds.
Common Exam Trap: Multiplying Marginals Too Soon
A frequent error is to calculate \(P(X>x)P(Y>y)\) and report it as the actual joint probability without establishing independence. That product is the correct benchmark under independence, but dependence can make the actual probability larger or smaller. Another error is to replace a strict exceedance with an inclusive one for a discrete loss; probabilities at the threshold then enter the calculation incorrectly.
Check Your Understanding
For each problem, calculate the simultaneous exceedance probability and compare it with independence when the required information is provided.
- A joint law assigns probability \(0.07\) to both threshold exceedances, while the separate exceedance probabilities are \(0.18\) and \(0.25\). Find the independence benchmark and the difference between the actual probability and that benchmark. (Answer: benchmark \(0.0450\); difference \(0.0250\); actual probability is higher.)
- On the unit square, \(f(u,v)=1+0.4(1-2u)(1-2v)\), with density zero outside. Find \(P(U>0.2,V>0.6)\) and its independence benchmark. (Answer: actual \(0.3354\); benchmark \(0.3200\), rounded to four decimals.)
- Suppose \(P(A)=0.15\), \(P(B)=0.20\), and \(P(B\mid A)=0.40\). Calculate the actual joint probability and the independence benchmark. (Answer: actual \(0.0600\); benchmark \(0.0300\); actual is higher.)
- At thresholds \(x=3\) and \(y=4\), suppose \(F_X(3)=0.65\), \(F_Y(4)=0.70\), and \(F_{X,Y}(3,4)=0.48\). Calculate the joint exceedance probability and the independence benchmark. (Answer: actual \(0.1300\); benchmark \(0.1050\); actual is higher.)
- A hypothetical model gives the following probabilities for whether losses \(X\) and \(Y\) exceed their respective thresholds: both exceed, \(0.09\); only \(X\) exceeds, \(0.11\); only \(Y\) exceeds, \(0.06\); neither exceeds, \(0.74\). Find the actual joint exceedance probability and independence benchmark. Does this one comparison establish that the loss variables are independent? (Answer: actual \(0.0900\); benchmark \(0.20(0.15)=0.0300\); no, it establishes only that the specified exceedance events do not factor as under independence.)