A prediction interval is a range of values intended to contain a future or otherwise unobserved outcome with a specified probability of coverage. In statistics, it expresses uncertainty about an observation rather than only uncertainty about a population parameter or an average response. Its construction depends on the available data, the quantity being predicted, and assumptions about how observations arise. Prediction intervals are used in regression, forecasting, experimental analysis, and other settings where a point prediction alone does not adequately describe uncertainty.
Definition and interpretation
Let denote the observed data and a future random variable. A prediction interval is a data-dependent interval
A frequentist prediction procedure with nominal coverage is typically designed to satisfy
for every parameter value allowed by the assumed model. Exact procedures may attain equality; approximate procedures may attain the target only approximately or as the sample becomes large.
The probability in this expression generally averages over both the data used to construct the interval and the future outcome. Thus, a 95% prediction procedure is designed so that, over repeated applications under the assumed data-generating process, approximately 95% of the resulting intervals contain their corresponding future observations.
This does not automatically imply that the conditional coverage is 95% for every realized dataset, every subgroup, or every predictor value. These are stronger and distinct requirements.
Prediction intervals usually have two endpoints, but one-sided prediction bounds are also useful. An upper prediction bound , for example, may be constructed so that
Distinction from related intervals
Confidence intervals
A confidence interval usually concerns a fixed but unknown parameter, such as a population mean or regression coefficient. A prediction interval concerns an outcome that remains random even if the parameters were known.
In regression, a confidence interval for the mean response at estimates
where the right-hand side is a conditional expectation. A prediction interval instead concerns a new response at that predictor value.
Under the same conventional regression model and confidence level, the prediction interval for a single new observation is wider than the confidence interval for its mean response. It must account for both uncertainty in estimating the mean and variation of individual observations around that mean.
Tolerance intervals
A tolerance interval is constructed to contain a specified proportion of an entire population distribution with a specified confidence level. For example, a tolerance interval might be designed to contain at least 99% of the population with 95% confidence.
A 95% prediction interval for one future observation does not provide that guarantee. Prediction coverage for an individual future outcome and confidence about population content are different probability statements.
Credible intervals
A credible interval describes posterior uncertainty about a parameter in Bayesian analysis. A Bayesian prediction interval instead describes uncertainty about a future observable outcome, usually through the posterior predictive distribution.
Sources of predictive uncertainty
A useful model is
where is the underlying mean response and is random variation around it.
Predictive uncertainty can include:
- Outcome variability: variation among observations even when the model and its parameters are known.
- Estimation uncertainty: uncertainty caused by estimating model parameters from finite data.
- Model uncertainty: uncertainty about the model's functional form, included predictors, or distributional assumptions.
- Future-input uncertainty: uncertainty about predictors that are themselves unknown at prediction time.
Not every method includes all these components. A regression interval calculated at specified future predictor values generally treats those values as known. An interval based on a single selected model often does not fully account for model selection.
Where future outcomes retain positive conditional variance, collecting more training data can reduce estimation uncertainty without eliminating outcome variability. Consequently, a prediction interval need not shrink toward zero width as the dataset grows.
Prediction for a normal population
Suppose are independent observations from a normal distribution with unknown mean and unknown variance . Let be the sample mean and let
be the sample variance.
If is an independent observation from the same population, then
has a Student's -distribution with degrees of freedom. An exact two-sided prediction interval is therefore
where is the -quantile of a -distribution with degrees of freedom.
The factor separates the variability of the new observation from uncertainty in the estimated mean. By comparison, the confidence interval for uses .
For example, if , , and , a 95% prediction interval is approximately
or approximately . The corresponding confidence interval for the population mean is approximately .
The prediction target matters. For the mean of independent future observations from the same population, the appropriate interval is
A future average and a single future observation therefore require different intervals.
Prediction intervals in linear regression
Consider the standard linear regression model
where has rows, linearly independent columns, and includes an intercept if required.
The ordinary least squares estimate is
For a new predictor vector , define
If the new error is independent of the observed errors and has the same normal distribution, an exact prediction interval for one new response is
The term
measures uncertainty in the estimated mean at . The additional accounts for the new observation's own error variance.
For simple linear regression, this becomes
The interval becomes wider as moves away from the center of the observed predictor values. Nevertheless, this widening does not protect against an incorrect linear relationship outside the observed range. Extrapolation can produce poor coverage even when the formula is applied correctly.
General methods of construction
Predictive quantiles
If a predictive probability distribution is available, an equal-tailed interval can be formed from its quantiles:
This construction need not be symmetric about the mean or median. For a skewed distribution, the upper and lower distances from a point prediction may differ substantially.
Equal-tailed intervals are not necessarily the shortest intervals with the required probability. For multimodal distributions, a prediction set consisting of separate regions may be more informative than a single interval. For discrete outcomes, attainable coverage levels are constrained by the probability masses, so coverage may exceed the nominal level.
Bayesian posterior prediction
In Bayesian inference, prediction is based on the posterior predictive distribution:
Here is the posterior distribution of the model parameters. Integrating over it propagates parameter uncertainty into predictions.
A Bayesian predictive interval contains a specified amount of posterior predictive probability conditional on the observed data and the chosen model and prior. This interpretation differs from repeated-sampling coverage. Bayesian predictive intervals can have good frequentist coverage, but that property does not follow automatically from their posterior probability.
Bootstrap methods
Bootstrap methods approximate predictive uncertainty by resampling data, simulating from a fitted model, or resampling estimated errors.
A prediction bootstrap must represent both uncertainty in the fitted prediction and randomness of the future outcome. Merely refitting a model on bootstrap samples and taking quantiles of the fitted mean predictions generally produces an interval for the mean response, not a complete prediction interval for an individual observation.
The resampling scheme must also respect relevant features of the data. For example, ordinary independent resampling is generally unsuitable for observations with substantial temporal dependence.
Quantile regression and conformal prediction
Quantile regression estimates conditional quantiles directly. Models for the and conditional quantiles can therefore provide the endpoints of a prediction interval. Their coverage depends on how accurately those quantiles are estimated.
Conformal prediction constructs prediction sets using conformity or nonconformity scores. In a common split-conformal procedure, a model is fitted on one subset of the data, while absolute prediction errors on a separate calibration subset determine an interval radius.
With appropriate finite-sample quantile adjustment and exchangeable calibration and future observations, split conformal prediction provides marginal coverage of at least . This guarantee does not require a correct parametric model for the outcome distribution. It does, however, depend on the exchangeability assumption and generally does not guarantee coverage at every individual predictor value.
Time-series forecasting
In time-series analysis, prediction intervals are often called forecast intervals. Their construction must account for temporal dependence and the forecast horizon.
For a forecast , an approximately normal forecast error with estimated standard error leads to the interval
where is a standard normal quantile.
The value of depends on the forecasting model and horizon . Uncertainty often increases with the horizon, but its behavior is model-dependent: it can approach a finite limit in some stationary processes and grow without bound in some nonstationary processes.
Forecasting software may condition on estimated model parameters as though they were known. Such intervals can omit parameter uncertainty and be too narrow, particularly in small samples.
Coverage, width, and limitations
A useful prediction interval balances coverage and informativeness. Wide intervals can achieve high coverage while contributing little precision. Narrow intervals are useful only when their stated coverage is credible.
Evaluation on held-out observations commonly examines:
- Empirical coverage: the proportion of outcomes inside their intervals.
- Average width: the mean difference between upper and lower endpoints.
- Conditional performance: coverage and width across predictor ranges, groups, or forecast horizons.
- Interval scores: criteria that penalize both excessive width and missed observations.
For a central interval and observed outcome , a commonly used interval score is
Lower scores favor narrower intervals while penalizing failures to cover the outcome.
Nominal coverage can fail when errors are non-normal, variance changes with predictors, dependence is ignored, or future observations differ systematically from the training population. Unaccounted model selection can also distort coverage.
Coverage must be matched to the intended claim. Separate 95% intervals for many future observations do not generally give a 95% probability that all observations will lie within their intervals simultaneously. Simultaneous prediction requires a joint construction or an adjustment such as the Bonferroni correction. Similarly, good overall coverage can coexist with substantial undercoverage in particular subgroups.