The Poisson Clumping Heuristic

Many real systems fail in ways that don't look Poisson at all up close — a market crash builds up from correlated stress across dozens of positions, a server outage cascades through dependent services, three fraud checks each individually depend on messy, non-random conditions. And yet the timing of rare failures in systems like these often still ends up looking approximately Poisson — random, memoryless arrival times. The Poisson clumping heuristic, developed by David Aldous, explains why, and gives a way to estimate rare-event probabilities without fully modelling the underlying mechanism.

The idea: group the space of possible states into overlapping "clumps" around each way the rare bad event could occur, and note that when the event is genuinely rare, these clumps are almost always far apart in time and hardly ever overlap. Under fairly general conditions, a rare event's occurrences behave like a Poisson process for exactly this reason — not because the individual triggering mechanism is memoryless, but because rare events are, almost by definition, so infrequent that whatever local correlation exists within one "clump" of near-misses barely matters, and the clumps themselves arrive independently of each other. This lets one estimate the rate of the rare event by computing the expected number of clumps per unit time, without needing an exact model of the underlying process.

A common quant use: estimating how often a portfolio breaches a severe drawdown threshold. The exact path to breaching it is complex and correlated across positions, but if breaches are rare enough, the heuristic justifies modelling breach times as approximately Poisson, which then unlocks simple formulas (expected time between breaches is one divided by the rate) that would otherwise require simulating the full correlated system.