Poisson Processes: Why Rare Events Cluster When You Expect Them to Spread Out

You've been waiting 20 minutes for a bus. Then three show up at once. Your first instinct is to blame the transit authority, bad scheduling, or the universe personally targeting you. The real culprit is simpler: you misunderstand how rare events actually distribute themselves through time.

Photo by DS stories on Pexels.

This is the Poisson process, and it's one of the most counterintuitive workhorses in all of probability.

What the Poisson Model Actually Says

A Poisson process models events that occur randomly over time (or space), at some average rate, and independently of each other. "Independently" is the key word. Each event has no memory of the last one. No waiting period. No recovery time. The process doesn't know it just fired.

The formula for the probability of observing exactly k events in a time interval, given an average rate of λ events per interval:

P(X = k) = (λ^k * e^(-λ)) / k!

Seems clean. Almost too clean. The trouble is what this math implies about clustering.

When λ is small (rare events), the distribution skews heavily toward zero. Most intervals will contain nothing at all. But the variance of a Poisson distribution equals λ exactly. That sounds stable until you realize that when events do arrive, they can pile up fast enough to compensate for all those empty stretches.

In plain terms: rare events don't spread out evenly. They clump.

Your Brain Expects a Grid. Reality Draws Dots at Random.

Imagine dropping 100 points randomly onto a line segment. Your intuition says they'll spread out roughly evenly, maybe 10 per section if you divide into 10 sections. Run the actual simulation and you'll get sections with 0, 1, 17, 3, 0, 22. Horrifyingly clumped. Completely correct behavior.

This is why:

• Server outages seem to happen in clusters
• Shooting stars come in bursts during a meteor shower, then go quiet
• Goals in soccer pile up in certain 10-minute windows
• Shark attacks appear to "epidemic" in a single summer

None of these require a causal story about why clustering happened. The clustering is the baseline expectation. Uniform spacing would actually be suspicious; it would suggest some repelling force keeping events apart.

graph TD
    A[Events occur at rate λ] --> B{Independent arrivals}
    B --> C[Most intervals: zero or few events]
    B --> D[Some intervals: multiple events]
    C --> E[Apparent quiet stretches]
    D --> F[Apparent clusters]
    E --> G((Perceived pattern where none exists))
    F --> G

Where Analysts Go Wrong

The failure mode shows up constantly in operational data. A data team sees three database failures in one week after months of stability. They write a post-mortem. They hunt for a root cause. They ship a fix. Then everything goes quiet for two months, and they credit the fix.

Maybe the fix helped. Maybe those three failures were just a Poisson cluster, and two months of quiet was the statistical echo of having already used up your "bad luck quota." You can't tell from a sample of one cluster.

The correct response to a Poisson cluster is: estimate your actual rate first. If your historical failure rate is λ = 0.5 per week, then three failures in one week has probability roughly 1.4%. Low, yes. But across 52 weeks and multiple systems, a 1.4%-chance event becomes nearly certain to appear somewhere. The cluster was always coming. You just didn't know which week it would land on.

This matters for staffing decisions, infrastructure capacity planning, and any domain where you respond to runs of bad events by overcorrecting.

The Exponential Connection

Here's the part that ties the model together: the waiting times between events in a Poisson process follow an exponential distribution. And the exponential distribution is memoryless.

If you've already waited 10 minutes for the bus, your expected additional wait is the same as when you first arrived. The past means nothing. Which is why "it's been so long, it must be coming soon" is exactly backwards. And why "we just had a crash, we're safe for a while" is equally wrong.

Both intuitions apply human social logic (debts get paid, streaks end, the universe balances) to a process that has no memory and owes you nothing.

Practical Takeaway

Before you explain a cluster, check whether a cluster was probable in the first place. Estimate your baseline rate. Calculate the probability of the observed count under a Poisson model. If it lands inside the top 10% of expected outcomes, you're describing noise as signal.

Rare events bunch. Empty stretches follow. Neither phenomenon needs a story. The mean rate is real; the smooth, evenly-spaced picture it implies is not.