Rare events and tail risk: probability, impact and repeated opportunities

Specify what makes an event rare

A probability needs a unit of observation and a time or opportunity frame. A hypothetical flag probability of 0.001 per independent inspection is not the same statement as 0.001 per year or per machine lifetime. The Stanford probability reference defines events and probability distributions; applying those ideas requires a precise event boundary. Decide whether repeated flags on one device count as separate events, whether the process changes over time and whether the published denominator includes all inspections. Without those details, a small percentage can be deceptively reassuring.

Calculate at least one event using the complement

Assume, solely for this example, that each of n inspections is independent and has the same flag probability p=0.001. The probability of no flag is (1−p)^n, so the probability of at least one is 1−(1−p)^n. This formula is derived from the stated assumptions. It does not describe a real inspection programme. Multiplying n by p gives the expected count, which is a different quantity and only approximates the at-least-one probability when the relevant conditions make multiple events sufficiently uncommon.

Compare repeated opportunities

At one thousand inspections the expected count is one, but that does not mean exactly one flag is certain. The table separates a count expectation from an event probability. If inspections share a common defect or environment, independence may fail and this particular calculation is no longer justified without a revised joint model.

Separate probability from consequence

Imagine two hypothetical failure categories with the same probability but very different repair times. A probability-only ranking treats them alike even though their practical effects differ. An expected consequence multiplies possible consequences by their probabilities and adds them, but that average still hides the shape of the distribution. A rare large effect can dominate an expectation while most observations remain small. Describe the consequence scale and the affected population rather than using the word rare as a synonym for unimportant. This is a principle of clear reporting, not a specific operational risk recommendation.

Use bounds with their assumptions visible

The Carnegie Mellon material on deviation inequalities distinguishes bounds based on moments from exact distributional probabilities. Markov’s inequality requires a nonnegative variable with finite expectation; Chebyshev’s uses finite variance. A bound can remain valid across many distributions while being loose for a particular one. Reporting an upper bound as an estimated frequency overstates what was calculated. Conversely, a model-specific tail estimate may look precise while depending strongly on an assumed distribution that has little support in the observed extremes.

Recognise model-dependent waiting times

The NIST exponential-distribution reference describes a model with a constant hazard and a survival function of exponential form. That property can be useful for some waiting-time models, but it is not a universal law of physical devices. Wear, changing operating conditions or shared external shocks can produce different behaviour. If a tail claim relies on an exponential or normal assumption, name that assumption. Evidence near the centre of a dataset may not strongly establish its extreme tail, especially when only a few rare observations are available.

Inspect selection and reporting opportunities

A dramatic event reported across many organisations may have had far more opportunities to occur than the story suggests. Counting only the one surprising case omits the many uneventful observations. Conversely, a system that records only completed inspections may omit failures that interrupted recording itself. Define the reporting process as well as the physical process. A useful analysis distinguishes an event’s probability under the model from the probability that someone notices, records or publicises it. Those filters can change the apparent frequency in an informal collection of examples.

Rare-event communication checklist

• Name the event, denominator and time frame.
• Separate a single opportunity from many opportunities.
• State dependence assumptions.
• Distinguish expected counts from at-least-one probabilities.
• Describe consequences as well as likelihood.
• Label bounds and model-based estimates accurately.
• Document missing events and reporting filters.

Use rounded figures consistent with the uncertainty in the inputs. A calculation based on an assumed probability should not gain a false aura of measurement merely because software prints many decimal places. Keep the original hypothetical parameters next to the result so readers can see which part is arithmetic and which part is a substantive assumption about the process.

Sources and verification

• Probability and statistics refresher — Stanford University teaching materials — 2026-10-23
• Deviation bounds and the law of large numbers — Carnegie Mellon University — 2026-10-23
• Exponential distribution — NIST/SEMATECH e-Handbook — 2026-10-23