Prediction interval
A prediction interval bounds a specified future observation or statistic with stated coverage under a model and sampling assumptions.
The target is a future random outcome, not a fixed parameter. Prediction uncertainty includes variation in that outcome and, when model parameters are estimated, estimation uncertainty. A confidence interval for a fleet failure rate therefore does not directly bound its next-period failure count. Specify the population, future exposure or horizon, prediction target and coverage. [1]
An interval for one future observation differs from an interval for a future sample mean or simultaneous bounds for several observations. Normal-theory limits for a future mean assume independent observations from the same normally distributed population; they are not a generic formula for discrete failure counts or censored lifetimes. [2]
Relations
Limitations
- Predictive coverage remains model-dependent even when a parameter is known exactly. A Poisson future count with known mean 1 equals 1 with probability exp(-1), not certainty; a fitted rate also introduces estimation uncertainty. [3]