What is a prediction?
A prediction is the value a model returns when given an input it has not seen before. The form depends on the task. A classifier returns a class, often with a confidence score. A regression model returns a number, such as the hours a repair will take. A forecasting model returns a series of future values. A ranking model returns an ordered list. The word applies even when the model is describing the present, as when it predicts which product appears in a photo.
Producing predictions from a trained model is called inference. The same model can serve predictions one at a time through real-time inference or across millions of records at once through batch inference.
Why prediction matters
A prediction only has value when someone or something acts on it. A hotel revenue manager who receives a predicted occupancy for each night over the next 60 days can adjust room rates in advance. The prediction is the input to that decision, and the decision is where the value lands. Projects that stop at producing predictions, with no process to consume them, produce dashboards nobody reads.
Every prediction carries uncertainty. A good system reports it, through a confidence score or a prediction interval, so that the consumer can treat a confident prediction differently from a marginal one. Predictions should also be logged with their inputs, because comparing them later with what actually happened is how model monitoring detects that a model has gone stale.