Poisson Approximation and the Chen-Stein Method

Counting rare events — the number of large drawdowns in a year, the number of simultaneous defaults in a credit portfolio, the number of extreme price jumps in a session — often comes down to summing up many individual indicator variables, each one rarely equal to 1. When those indicators are independent, the classical result is clean: the sum converges to a Poisson distribution as the number of trials grows and their individual probability shrinks. Real events are rarely perfectly independent, though — a market jump can trigger a cluster of related jumps. The Chen-Stein method answers exactly how much that dependence matters, with an explicit numerical error bound rather than a vague "approximately."

An analogy: counting rare coincidences at a party

Imagine a large party where, for every pair of guests, there's a tiny chance they happen to share a birthday. With many guests, the number of such coincidental pairs behaves like a Poisson-distributed count — a small number of pairwise "hits" out of an enormous number of pairwise chances, exactly the regime Poisson approximation was built for. Now suppose some pairs aren't independent — say, twins in the room automatically share a birthday, creating a cluster of dependent "hits." The Chen-Stein method is the tool for saying precisely how much this small pocket of dependence throws off the otherwise-clean Poisson prediction, rather than just hand-waving that the approximation "should still be roughly fine."

The idea, one symbol at a time

Let $X = \sum_{i=1}^n I_i$ be a sum of $n$ indicator variables, each $I_i = 1$ with small probability $p_i$ and $0$ otherwise, and let $\lambda = \mathbb{E}[X] = \sum_i p_i$. If the $I_i$ were independent, the classical Poisson limit theorem says $X$ converges to a Poisson distribution with mean $\lambda$ as $n$ grows and each $p_i$ shrinks. The Chen-Stein method extends this to weakly dependent indicators and produces an explicit bound of the form

$$d_{TV}(X, \text{Poisson}(\lambda)) \le b_1 + b_2,$$

where $d_{TV}$ measures the total distance between the true distribution of $X$ and a Poisson$(\lambda)$ distribution, and $b_1, b_2$ are computable quantities built from how strongly each $I_i$ depends on the others near it. In plain English: the method doesn't just assert the approximation is good, it produces a specific number bounding exactly how wrong the Poisson approximation could be, built entirely from the dependence structure of the underlying events — small $b_1, b_2$ (weak, local dependence) means a trustworthy Poisson approximation; large values are an explicit warning that dependence has broken the approximation.

Worked example 1: independent rare events

Suppose 1,000 independent stocks each have a 0.2% daily chance of a large idiosyncratic jump. The expected number of jumping stocks on a given day is $\lambda = 1000 \times 0.002 = 2$. Under the classical Poisson limit (a special case of Chen-Stein with the dependence bound equal to zero), the probability of seeing exactly 3 jumps on a given day is

$$\Pr(X = 3) = \frac{\lambda^3 e^{-\lambda}}{3!} = \frac{8 \times e^{-2}}{6} \approx \frac{8 \times 0.1353}{6} \approx 0.180.$$

With true independence, this Poisson formula is essentially exact for this scale of problem, and no correction term is needed.

Worked example 2: adding dependence and bounding the error

Now suppose 20 of the 1,000 stocks belong to one sector and are linked so that if any one of them jumps, the conditional probability the others in that sector jump within the same day roughly doubles. This local clustering doesn't change $\lambda = 2$ (the mean count is unaffected by correlation), but it does inflate the variance of $X$ above what a true Poisson$(2)$ distribution has (whose variance equals its mean, 2). The Chen-Stein bound formalizes this: $b_2$, the term capturing dependence among nearby indicators, grows roughly with the size and strength of the correlated cluster — here, a 20-stock cluster with roughly doubled conditional jump probability contributes a small but non-zero $b_2$, giving an explicit, computable upper limit (rather than a guess) on how far $\Pr(X=3)$ can be from the Poisson value of 0.180 computed above.

What this means in practice

Poisson approximation and Chen-Stein bounds underpin models of rare, clustered financial events: joint default counts in credit portfolios, the frequency of simultaneous risk-limit breaches, and extreme-move clustering across correlated assets. The practical value of Chen-Stein specifically is that it turns "dependence probably doesn't matter much here" from an assumption into a checkable, quantitative claim — useful precisely in the situations, like sector-clustered jump risk, where naively assuming independence would understate tail risk.