How is it different from a fact?

A fact is observed; a prediction is inferred from resemblance to past cases and carries error, known or not.

What makes a prediction trustworthy?

Evidence: validation accuracy on unseen data, a case similar to training coverage, and an honest uncertainty statement.

So outcomes can later audit accuracy, catching drift and calibrating how much the model's numbers deserve trust.

AI Glossary / AI and machine learning / Beginner

Prediction

Q: What is a prediction?

The model's answer for a new case, produced by applying patterns learned from training examples.

Definition

A prediction is a model's answer for a case it was not told the answer to: the estimated price of a house, the class of an email, the demand for next month, produced from the patterns learned in training.

What it means

Prediction is the moment a model stops studying and starts working. Training consumed examples with known answers; prediction applies whatever was learned to fresh cases where the answer is genuinely unknown.

A prediction is not a fact about the future but a statement from the model's learned viewpoint: given cases like this in the training data, the answer usually looked like that. Its reliability depends entirely on how much the new case resembles what the model learned from.

Predictions come with shape: a single number, a class label, a probability per class, or a distribution of possible outcomes. The shape matters, because a bare number hides the uncertainty that a probability or a range would have confessed.

For a business, the discipline is simple: a prediction is an input to a decision, never the decision itself, and its value depends on knowing how wrong it typically is.

How it works

Feed the trained model a new case's features, exactly as they were prepared in training.

Produce the output in the shape the decision needs: point estimate, class, probability or distribution.

Attach what is known about typical error: validation performance, confidence or prediction intervals.

Compare the case to training coverage, since predictions outside the data's territory are extrapolations, not estimates.

Log predictions with their inputs, so later outcomes can score how good the predictions actually were.

Three practical examples

A house-price model predicts three hundred and ten thousand for a new listing; the eventual sale at three hundred and two thousand lands inside its typical error band.
A spam filter predicts junk with ninety-nine per cent confidence, and the message never touches an inbox.
A demand model predicts next month's volume per warehouse, and the logistics plan is built from the predictions plus their known error margins.

Think of it this way

Think of a weather forecast: not a promise of rain but a statement from patterns in past weather, useful exactly in proportion to how honestly its uncertainty is reported. You plan the picnic with the forecast and the umbrella odds together.

Worked example

the confident newcomer

Suppose a lender's model predicts default risk for a new applicant type, gig-economy couriers, barely present in training data. The figures are illustrative.

The model returns predictions with its usual confidence, but those confidences were calibrated on cases unlike these, so they overstate how well the model knows this territory.

The fix is a coverage check: the scoring pipeline flags when a case sits outside training coverage, and flagged cases route to manual underwriting with the prediction demoted to advisory.

The general lesson: a prediction's quoted confidence means something only inside the territory the model learned; outside it, the number arrives with false assurance.

Illustrative case study

An airline revenue team once trusted a demand model's predictions through a schedule expansion into routes the training data barely covered. Forecasts looked precise, actuals disappointed, and the post-mortem drew a thick line between prediction and extrapolation in every subsequent review. The lasting changes were procedural: every forecast now ships with its error history on comparable routes, predictions on new routes carry an explicit low-evidence flag, and a forecast without its error band is returned to sender. Prediction quality, the team concluded, is a property of the match between the question and the evidence, not of the model's confidence.

Common misconceptions and mistakes

A prediction is the model knowing the future. It is the model applying past patterns; the future is under no obligation to resemble the past.
High confidence means high accuracy. Confidence reflects the model's internal consistency, which can be high and wrong, especially outside training territory.
A prediction is the decision. It is one input; the decision adds costs, constraints and judgment the model never saw.

Business use

Require every prediction used in decisions to carry its error band or confidence, and train staff to read both.

Score predictions against outcomes routinely; a model nobody audits drifts from useful to decorative.

Flag predictions outside training coverage and route them to humans, because that is where errors concentrate.

Keep the decision rule separate from the prediction, so changes in risk appetite do not require retraining.

Check your understanding

What is a prediction, and what three things decide whether one deserves trust?

Answer: A prediction is a model's inferred answer for a case whose true answer is unknown, produced by applying learned patterns. Trust depends on three things: measured accuracy on unseen data, whether the new case resembles the training territory, and an honest statement of uncertainty. A prediction missing any of the three is a guess wearing a number's clothing.

Related terms

Inference
Model
Validation
forecast
extrapolation
confidence

Learn next

Be Productive Using AI builds the habit of asking what a prediction rests on before acting on it: https://aiforbusinessandlife.com/academy/be-productive-using-ai [C1]

Sources

[S12] Aalto University, Dictionary of Machine Learning; [S2] Google, Machine Learning Glossary; [S3] NIST AI Resource Center glossary. Full URLs are in the source ledger. Verified 11 October 2026.

Previous: Predicate (logic) Next: Prediction bias