Subexponential Distributions and the Single Big Jump

If a portfolio's monthly loss is unusually large, is that because of one single catastrophic position blowing up, or because many positions each lost a moderate amount at the same time? The answer changes what you should actually worry about and defend against — one bad position or systemic correlation across many. Subexponential distributions give this everyday intuition a precise mathematical form: for a well-defined class of heavy-tailed distributions, an unusually large sum of independent losses is, overwhelmingly, caused by exactly one of those losses being unusually large on its own, while the rest are perfectly ordinary. This is called the single big jump principle.

An analogy: a marathon versus a plane crash

If a group of marathon runners' total combined finishing time was unusually slow, the likely explanation is that everyone ran a bit slower than expected — bad weather, a tough course — not that one runner alone took ten times as long (running times don't vary that wildly). But if a group of insurance claims in a portfolio was unusually large in total, the likely explanation often is that one single claim was catastrophic — a total loss, a lawsuit — while every other claim was perfectly typical. Loss distributions with this second character, where extremes are dominated by one outlier rather than distributed effort, are subexponential.

The idea, one symbol at a time

Let $X_1, \ldots, X_n$ be independent, identically distributed random variables (e.g., $n$ independent claim or loss sizes) with common distribution $F$ and survival function $\overline{F}(x) = P(X > x)$. Let $S_n = X_1 + \cdots + X_n$ be their sum. $F$ is called subexponential if, as $x \to \infty$:

$$P(S_n > x) \sim n \, \overline{F}(x) .$$

In plain English: the probability that the total of $n$ losses exceeds a large threshold $x$ is, for large $x$, essentially the same as the probability that just one of the $n$ losses alone exceeds $x$, multiplied by $n$ simply because there are $n$ chances for that outlier to occur. The other $n-1$ terms contribute negligibly; they're assumed ordinary. Equivalently, the single big jump principle states:

$$P(S_n > x \mid \max(X_1,\ldots,X_n) > x) \to 1 \quad \text{as } x \to \infty .$$

In plain English: conditional on the sum being extremely large, the probability this happened because one individual term was itself that large approaches 1 — sum-being-big and one-term-being-big become, in the extreme tail, the same event. This is the opposite of light-tailed distributions (like the normal), where a large sum is instead explained by all the terms being simultaneously a bit above average.

Worked example 1: Pareto losses (subexponential)

Take losses with a Pareto tail $\overline{F}(x) = x^{-2}$ for $x \ge 1$ (subexponential). For $n = 2$ independent losses, the theory predicts $P(S_2 > x) \sim 2\overline{F}(x) = 2x^{-2}$ for large $x$. Checking at $x = 100$: $2 \times 100^{-2} = 2 \times 0.0001 = 0.0002$. This says a combined loss of over $100m from two independent positions is (in the extreme) about twice as likely as a single position alone losing over $100m — consistent with "it took one big loser, and there were two chances for one to appear," not with "both lost $50m each."

Worked example 2: normal losses (not subexponential) for contrast

Take two independent losses each $N(0, 10^2)$ (in $m, light-tailed). Their sum is $N(0, (10\sqrt{2})^2) = N(0, 14.14^2)$. The probability the sum exceeds $40m: using the normal tail, $P(S_2 > 40) = P(Z > 40/14.14) = P(Z > 2.83) \approx 0.0023$. Compare to the probability a single loss alone exceeds $40m: $P(X_1 > 40) = P(Z > 4) \approx 0.00003$ — vastly smaller than the sum probability. Here the sum being large is overwhelmingly explained by both variables being moderately elevated together, not by one alone spiking — exactly the opposite pattern from the subexponential case.

What this means in practice

The single big jump principle tells a risk manager where to focus: for subexponential loss distributions (common for credit losses, insurance claims, and single-name equity blowups), extreme portfolio losses are best defended by capping the size of any single position or exposure, not by diversifying across many moderate ones — diversification barely helps against a single-big-jump tail, since the danger was never "many things going wrong together." This is the mathematical justification behind position limits and concentration limits as the primary tail-risk control for such exposures, in contrast to correlation-based diversification, which is the right tool for light-tailed risk instead.