A point forecast answers one question: what value is most useful as the centre? A prediction interval answers the question the receiver usually cares about more: how far could the next real outcome reasonably land from that centre?
A model predicts tomorrow’s journey time as 31 minutes. That number is easy to read and easy to misuse. The next journey could take 27 minutes, 35 minutes or 44 minutes depending on ordinary variability, model uncertainty and operating conditions.
A prediction interval is a range constructed to contain a future observation with a stated level of coverage under a model and its assumptions. It is not merely a confidence interval around an estimated mean.
This is a specialist branch beneath How Forecasting Works, How Uncertainty Works and How Uncertainty Propagation Works. The narrow question is: how should uncertainty about one future outcome be shown?
Mean Uncertainty and Future-Outcome Uncertainty Are Different
Suppose we estimate the average height of a population. With enough data, the average can be estimated very precisely.
That does not mean the height of the next person is nearly known. Individual heights remain variable around the average.
A confidence interval for the mean reflects uncertainty in the estimated centre. A prediction interval for a future observation must also include the natural or modelled variation of individual outcomes.
That is why prediction intervals are usually wider.
Coverage Is a Long-Run Property Under the Assumptions
A 95% prediction interval does not mean there is a universal metaphysical guarantee that one particular future observation has a 95% probability of landing inside the interval under every statistical framework.
In frequentist use, the interval procedure is designed so that across repeated applications under the model assumptions, roughly the stated proportion of future observations would be captured.
Bayesian predictive intervals are interpreted through the posterior predictive distribution. The language differs; both approaches insist that future-outcome uncertainty should be explicit rather than hidden behind a point.
Prediction Intervals Depend on the Operating State
Uncertainty is often not constant across the input range.
Journey-time variation may be small late at night and large at peak hour. demand forecasts may be tighter on ordinary weekdays than during holidays. machine measurements may become less stable near operating limits.
A useful interval should widen and narrow with the modelled uncertainty rather than applying one generic band everywhere.
Intervals Become Fragile Under Misspecification
An interval can have mathematically correct coverage under the wrong model and poor real-world coverage.
If the model ignores heavy tails, changing variance, regime shifts or structural breaks, the stated 95% interval may contain far fewer than 95% of future observations.
This is why interval quality must be validated empirically.
See How Model Misspecification Works.
Calibration Can Be Checked
If a forecasting system issues thousands of nominal 90% prediction intervals, we can later ask what fraction of realised outcomes actually landed inside them.
If only 65% did, the intervals were too narrow or the model was failing in some regimes. If 99.9% did, they may be so wide that they provide little operational value.
Good intervals balance honest coverage with useful sharpness.
Intervals Should Be Evaluated by Regime, Not Only Overall
A model can achieve respectable average coverage while failing badly in the states that matter most.
A flood forecast might be well calibrated during ordinary rainfall and dangerously overconfident during extreme storms. A financial model may cover typical returns while missing crisis tails.
The receiver cares about conditional reliability, especially near consequential thresholds.
Worked Example: MRT Journey Time
A model predicts a 31-minute journey with a 90% prediction interval of 28 to 37 minutes under ordinary peak conditions.
The point forecast supports planning. The interval tells the passenger or operator how much routine uncertainty remains.
If the same model enters a disruption state it has barely seen before, the interval should not be trusted automatically. The operating regime has moved beyond the evidence supporting its calibration.
The railway owner remains How MRT Works | It’s Mathematics.
Worked Example: Finance
A bank forecasts next-quarter credit losses.
The expected loss is one number. A prediction range around realised future loss must reflect borrower-level variability, macroeconomic uncertainty, parameter uncertainty and model assumptions.
If the range ignores uncertainty about the economic regime, it can look sophisticated while remaining too narrow for stress.
A Careful Analogy: Education
A teacher predicts that a learner is likely to score around 75 on a future test.
A useful planning statement may be “given current evidence, ordinary performance could plausibly vary across a broader band.” The precise formal interval depends on the assessment model, but the principle transfers: future performance is not one predetermined point.
Prediction Intervals Need a Receiver
A range is useful only if it connects to a decision.
If even the upper bound remains safe, action may not change. If the interval crosses a safety, capacity or financial threshold, uncertainty itself becomes operational information.
This is why the correct interval is not simply the narrowest one. It is the narrowest interval whose calibration is still defensible for the receiver’s job.
A Prediction-Interval Checklist
- Separate uncertainty in the estimated mean from variability in future observations.
- Choose a coverage level suited to the decision.
- Allow interval width to vary across operating states where appropriate.
- Include relevant parameter and model uncertainty.
- Backtest empirical coverage on future or held-out data.
- Check calibration by regime and near important thresholds.
- Report point forecasts beside intervals, not instead of them.
- Widen or suspend intervals when the case lies outside the validated operating domain.
Read the Mechanism in Three Directions
Forward: model + current state → predictive distribution → interval → future observation → calibration check. Backward: start from a missed future outcome and ask whether the interval was too narrow because of noise, parameter uncertainty, misspecification or regime change. Across: compare forecaster, operator and receiver; each may prefer a different trade-off between sharpness and coverage.
A point forecast tells you where to look. A prediction interval tells you how much of the future you still need to be ready for.
Continue through How Epistemic and Aleatoric Uncertainty Work, How Extrapolation Works and the master How X Works hub. Next: out-of-distribution prediction — when the next case is not simply uncertain, but comes from a region of the world the model barely knows.