Glossary

The core vocabulary shared across every field surveyed on this site, in plain language.

Half-life

The time it takes for a quantity to fall to half its starting value — or, in engagement contexts, the time it takes a piece of content to accumulate half of the total engagement it will ever receive. Used across radioactive decay, pharmacology, and, more recently, social-media engagement studies. See the content decay page for published half-life figures.

Hazard rate

The instantaneous probability that an item fails (or an event occurs) right now, given that it has survived until now. Also called the failure rate or force of mortality depending on the field. A hazard rate that rises over time describes "wear-out"; one that falls describes "infant mortality"; a constant hazard rate describes purely random, age-independent failure. See the machines page for the bathtub curve, which traces all three phases.

Survival curve

A function that shows the probability an item is still alive (or a subject is still event-free) past a given time. Starts at 1.0 (100% survival at time zero) and decreases toward zero. The Kaplan-Meier estimator, described on the medicine page, is the standard way to estimate a survival curve from real, partially-observed data.

TTL / RUL

Time-To-Live and Remaining Useful Life are the same underlying concept named differently by different fields: how much time is left before an item reaches its end-of-life or failure threshold. RUL is the term of art in predictive maintenance (see the maintenance page); TTL is the more general/networking-derived term used elsewhere.

Censoring

What happens when you don't get to observe an item's actual failure or event time — a patient is still alive when a clinical study ends, a subscriber hasn't churned yet when the data was pulled, a machine is still running when the observation window closes. Survival analysis is built specifically to use this partial information correctly rather than discarding incomplete cases. See the medicine page.

Calibration

Whether a model's predicted probabilities or times match reality on average — as opposed to whether it merely ranks items in the right order (see Discrimination, below). A well-calibrated model that says "70% survival probability at one year" should be right about 70% of the time across all the cases where it made that prediction. Calibration is checked separately from ranking ability in every mature field surveyed here; see the methodology page for why the two are not interchangeable.

Discrimination

A model's ability to correctly rank items relative to each other — does the item that actually failed first get ranked as higher-risk than the item that failed later? Measured by metrics like the C-index (see medicine) or AUC. A model can discriminate well while still being poorly calibrated, which is why mature fields always report both.

Confidence interval

A range around a point estimate that is expected to contain the true value at some stated probability (commonly 90% or 95%). Reliability engineering, in particular, treats the confidence interval — not the point estimate alone — as the actual deliverable of a lifetime prediction. See the machines page for the chi-square-based MTBF bound formula.

Convergence

How quickly a prediction's accuracy or precision improves as more data arrives — does the algorithm "lock in" on the right answer as the true failure point approaches, or does it stay noisy until the very end? Formally defined in the predictive-maintenance literature as the geometric distance from the first prediction to the centroid of the accuracy-vs-time curve. See the maintenance page.