Why Joint Survival Probabilities Matter
For a risk manager, knowing how often each loss is large on its own is not enough. Two business lines might each have a modest probability of exceeding a threshold, yet the chance that both losses exceed their thresholds at the same time may be much higher or lower than an independence calculation suggests. That simultaneous tail probability is important when assessing accumulation risk and the potential for several losses to strain capital together.
This tutorial relates the joint survival probability of paired losses to their marginal exceedance probabilities. The distinction between a strict exceedance, such as \(X>x\), and a threshold-inclusive event, such as \(X\ge x\), is especially important for discrete losses or any model with positive probability at a threshold.
Relating the Joint Tail to the Marginal Tails
The marginal exceedance probabilities describe the sizes of the two separate tails. The joint survival probability describes their overlap: the probability that both losses are in their respective tails at once. The exact relationship uses the joint CDF as well as both marginal CDFs.
Since \(S_X(x)=1-F_X(x)\) and \(S_Y(y)=1-F_Y(y)\), the same relationship can be written as
The first formula follows directly from inclusion-exclusion. The complement of \(\{X>x,Y>y\}\) is the union \(\{X\le x\}\cup\{Y\le y\}\). The probability of that union is \(F_X(x)+F_Y(y)-F_{X,Y}(x,y)\), because the event \(\{X\le x,Y\le y\}\) was counted twice. Subtracting the union probability from 1 gives the displayed joint-tail formula.
The two marginal tails alone do not generally determine the joint tail. They say how often each threshold is crossed, but not how the crossings overlap. If \(X\) and \(Y\) are independent, the overlap is the product \(S_X(x)S_Y(y)\). Without independence, that product is only a comparison benchmark, not a general formula.
These bounds are useful checks, but they do not usually identify the joint probability exactly. A joint distribution or other information about dependence is needed for an exact value. When comparing an observed joint tail with \(pq\), describe it as higher or lower than the independence benchmark; do not infer a complete dependence structure from that one comparison.
Worked Paired-Loss Calculations
Worked Example: Calculating a Joint Tail From a Discrete Law
A hypothetical insurer models two annual losses, \(X\) and \(Y\), in thousands of dollars. Their joint probability distribution is shown below. Calculate \(P(X>4,Y>5)\), the two marginal exceedance probabilities, and the independence benchmark.
| \(X\backslash Y\) | \(Y=0\) | \(Y=5\) | \(Y=12\) | Row total |
|---|---|---|---|---|
| \(X=0\) | 0.20 | 0.10 | 0.05 | 0.35 |
| \(X=4\) | 0.10 | 0.15 | 0.10 | 0.35 |
| \(X=10\) | 0.05 | 0.10 | 0.15 | 0.30 |
| Column total | 0.35 | 0.35 | 0.30 | 1.00 |
Method. Sum the cells satisfying each strict threshold, keeping the strict inequalities in view. The row totals and column totals check that the listed probabilities sum to 1.00.
Marginal tail for \(X\). The only listed \(X\) value greater than 4 is 10, so \(P(X>4)=0.05+0.10+0.15=0.30\).
Marginal tail for \(Y\). The only listed \(Y\) value greater than 5 is 12, so \(P(Y>5)=0.05+0.10+0.15=0.30\).
Joint tail. Both strict inequalities hold only in the cell \(X=10,Y=12\). Therefore \(P(X>4,Y>5)=0.15\).
Independence benchmark and check. If the losses were independent, the joint probability would be \(0.30(0.30)=0.09\). The actual value, 0.15, is greater than that benchmark. It also satisfies the general bounds \(0\le0.15\le\min(0.30,0.30)=0.30\). A common mistake is to include the \(Y=5\) column because it meets the threshold numerically; the event is \(Y>5\), so that column is excluded.
Worked Example: Interpreting Simultaneous Large Losses
In a hypothetical portfolio, let \(X\) be a wind-related loss and \(Y\) a water-damage loss, both measured in thousands of dollars. Each loss is either zero or a positive amount. The model assigns the following probabilities: both losses exceed their respective thresholds of 100 and 80, 0.12; only \(X\) exceeds its threshold, 0.18; only \(Y\) exceeds its threshold, 0.08; neither exceeds, 0.62. Calculate the marginal tail probabilities, the joint tail, the independence benchmark, and the conditional probability that \(Y>80\) given \(X>100\).
Method. First add the scenario probabilities that satisfy each marginal event, then identify the scenario satisfying both. The four scenario probabilities sum to \(0.12+0.18+0.08+0.62=1.00\).
Marginal tails. The event \(X>100\) occurs in the “both” and “only \(X\)” scenarios: \(P(X>100)=0.12+0.18=0.30\). The event \(Y>80\) occurs in the “both” and “only \(Y\)” scenarios: \(P(Y>80)=0.12+0.08=0.20\).
Joint tail and benchmark. Both events occur in the “both” scenario, so \(P(X>100,Y>80)=0.12\). Under independence, the benchmark is \(0.30(0.20)=0.06\). The modeled joint tail is twice that benchmark.
Conditional interpretation. The conditional probability is \(P(Y>80\mid X>100)=0.12/0.30=0.40\). Thus, in this hypothetical model, 40% of cases in which \(X\) exceeds 100 also have \(Y\) exceeding 80. This is higher than the unconditional \(Y\)-tail probability of 20%. The bounds check gives \(0\le0.12\le\min(0.30,0.20)=0.20\).
The common mistake is to treat the two marginal exceedance probabilities as enough to calculate the joint tail and report 0.06 as the answer. That is the answer only under independence; here the scenario probabilities specify a joint tail of 0.12.
Threshold Atoms and Strict Versus Inclusive Events
For a continuous loss distribution with no probability mass at a threshold, \(P(X>x)=P(X\ge x)\). For a discrete loss, those probabilities can differ because \(P(X=x)\) may be positive. A CDF value \(F_X(x)=P(X\le x)\) includes the probability at \(x\); therefore \(1-F_X(x)\) is \(P(X>x)\), not \(P(X\ge x)\).
The joint CDF formula above is specifically for the strict event \(X>x,Y>y\). If a problem changes one inequality to inclusive, the event changes and threshold mass may matter. For example, with \(F_X(x^-)=P(X<x)\), the corresponding formula for one inclusive and one strict threshold is
This follows by taking the complement of \(\{X\ge x,Y>y\}\), which is \(\{X<x\}\cup\{Y\le y\}\), and applying inclusion-exclusion. The left-limit terms and mixed joint probability are not generally recoverable from \(F_X(x)\), \(F_Y(y)\), and \(F_{X,Y}(x,y)\) alone. Thus, an inclusive event may be only partly identified by CDF values at the thresholds.
Worked Example: Rejecting an Impossible Inclusive-Tail Report
A hypothetical loss pair has \(F_X(10)=0.60\), \(F_Y(25)=0.50\), and \(F_{X,Y}(10,25)=0.30\). An analyst reports \(P(X\ge10,Y>25)=0.18\). Determine the strict joint survival probability, assess the report, and state what can be concluded about the inclusive event.
Method. First use the joint CDF identity for the strict event, then compare the inclusive event with it by set containment.
Strict joint survival. \(P(X>10,Y>25)=1-0.60-0.50+0.30=0.20\).
Check the reported value. Every outcome satisfying \(X>10,Y>25\) also satisfies \(X\ge10,Y>25\). Hence \(\{X>10,Y>25\}\subseteq\{X\ge10,Y>25\}\), and so \(P(X\ge10,Y>25)\ge0.20\). The reported value 0.18 is impossible and is inconsistent with the supplied CDF values.
What is and is not known. The exact probability of the inclusive event is not determined by these three CDF values alone: probability at \(X=10\) while \(Y>25\) could add to the strict probability. The marginal probability \(P(Y>25)=1-0.50=0.50\) supplies an upper bound. Therefore the given information implies \(0.20\le P(X\ge10,Y>25)\le0.50\), not an exact value. The trap is to say the inclusive probability is simply unknown and stop there: although its exact value is unknown, the strict event provides a lower bound that rules out 0.18.
Practical Checks and Risk Interpretation
A disciplined joint-tail calculation separates three questions: which inequalities define the event, what information determines the probability, and whether the resulting number is feasible. In particular, calculate marginal tails using the stated strict or inclusive comparisons before comparing them with the joint probability.
- Check event containment. A joint exceedance cannot be more likely than either marginal exceedance: \(P(X>x,Y>y)\le\min(S_X(x),S_Y(y))\).
- Check the lower bound. If the two marginal tails sum to more than 1, their overlap must be at least \(S_X(x)+S_Y(y)-1\).
- Use independence only when justified. The product of marginal tails is an independence result, not a general aggregation rule.
- Inspect equality at the threshold. For discrete or mixed models, changing \(>\) to \(\ge\) can change the event and its probability.
In capital and risk aggregation, the marginal tails answer how often each component is large, while the joint tail helps answer how often large losses coincide. The same marginal distributions can be paired in different ways and produce different joint tails. Consequently, marginal severity information alone is not a complete description of simultaneous large-loss risk.
Check Your Understanding
For each problem, preserve the stated threshold inequalities and check that the result is consistent with the marginal tail probabilities.
- For two losses, \(P(X>x)=0.18\), \(P(Y>y)=0.12\), and the losses are independent. Calculate the joint exceedance probability and the general lower and upper bounds that would apply even without independence. (Answer: independence benchmark \(0.0216\); general bounds \(0\le P(X>x,Y>y)\le0.12\).)
- A hypothetical discrete model gives the joint probabilities below, where \(X\) takes values 0, 50, 100 and \(Y\) takes values 0, 40, 90. The rows of the table are \(X=0:(0.25,0.10,0.05)\), \(X=50:(0.10,0.20,0.05)\), and \(X=100:(0.05,0.10,0.10)\), with entries ordered by increasing \(Y\). Find \(P(X>50)\), \(P(Y>40)\), \(P(X>50,Y>40)\), and the independence benchmark. (Answer: \(0.25,\ 0.20,\ 0.10,\ 0.05\), respectively.)
- Suppose \(F_X(8)=0.72\), \(F_Y(15)=0.65\), and \(F_{X,Y}(8,15)=0.50\). Calculate the strict joint survival probability. (Answer: \(P(X>8,Y>15)=1-0.72-0.65+0.50=0.1300\).)
- For some thresholds \(a,b\), the strict joint probability \(P(X>a,Y>b)\) is 0.11, but a report gives \(P(X\ge a,Y>b)=0.09\). Can the report be correct? Is the exact inclusive probability determined by the strict probability alone? (Answer: No; the inclusive event contains the strict event, so its probability must be at least 0.11. The exact inclusive probability is not determined by the strict probability alone.)
- Let \(P(X>x)=0.35\), \(P(Y>y)=0.40\), and \(P(X>x,Y>y)=0.28\). Find \(P(Y>y\mid X>x)\), compare the joint probability with the independence benchmark, and verify the bounds. (Answer: conditional probability \(0.28/0.35=0.8000\); independence benchmark \(0.35(0.40)=0.14\); bounds \(0\le0.28\le0.35\).)