How Rare Events Can Occur in Clusters
- Recent failures in computer, aviation and banking systems can lead you to think these events are related and share a common root cause.
- But random, unrelated events can also land in a short period of time out of sheer coincidence, and look like a cluster of related events.
- This model shows rare failure events over a century across 10 to 30 entities: NASDAQ-listed companies and federal agencies, one per timeline.
- Change the characteristics of the entities with the sliders, such as how often they fail and how many failures get noticed, and watch clusters appear out of pure randomness.
- The companies and agencies are assigned entirely at random. The events are simulated and don't reflect anything that really happened.
Parameters
TimelinesEach row is one entity, a company or federal agency, whose failures happen at random and independently of every other row. Adding timelines adds more places for failures to happen, so coincidences become more likely even though nothing connects the rows.
Events/timeline/centuryHow often each entity fails, on average. At 10, an entity fails about once a decade. Failures follow a random (Poisson) process: the timing of one failure says nothing about when the next will come.
Cluster windowHow close together failures must fall to look related. A short window, like 3 months, takes a tight coincidence; a long window, like 5 years, makes clusters easy to find. News cycles tend to link events that happen within weeks or months of each other.
Cluster cutoffHow many counted failures must land inside one window before they look like a cluster. 5+ is a striking coincidence; 2+ is easy to hit. Only failures that are seen and of medium or high impact count.
VisibilityNot every failure makes the news. This sets the share of failures the public sees. Unseen failures (cross-hatched) still happen, but nobody notices them, so they can't be part of a perceived cluster. More visibility means more apparent clusters.
Impact skewEach failure has a low (L), medium (M) or high (H) impact, shown as light, medium or dark blue. Only M and H failures are serious enough to count toward a cluster. At the left the three levels are equally likely; moving right turns low-impact failures into high-impact ones, so more failures count.
1030
120
3 months5 years
5+ events2+ events
20% seen80% seen
1/3L 1/3M 1/3H0/3L 1/3M 2/3H
Events Over Time
Seen event, impact L/M/H (only M/H count)
Unseen event (not counted)
Cluster
Clusters vs Number of Timelines
2+ in a window
3+
4+
5+
Current setting