Base Rate

How often a candidate cause, such as a software update or a schedule change, occurs in general without producing the failure under investigation. A correlation in time with a rare event is stronger evidence than a correlation with something that happens constantly and usually causes no harm; checking the base rate before treating a correlated change as a likely branch guards against overweighting the first coincidence noticed.

How common an outcome is in a dataset (for example, the fraction of cases that are actually fraud). A high accuracy figure can be meaningless if the base rate is very low, because a model can score well simply by always predicting the common outcome, which is why the base rate is a required companion to any accuracy claim.

What usually happens in situations of a given type, used as the anchor against which a specific prediction is judged. A prediction that departs sharply from the base rate must justify the departure; ignoring base rates for a vivid specific claim is a classic judgment error.

The historical frequency or cost of an event (a fraud loss, an error, a stockout) used to anchor a risk-reduction claim. Without a base rate, a claimed reduction is a guess; with one, it is a defensible number.

How common an outcome is across the whole population; essential context for interpreting any AI risk score or probability, because a high-accuracy flag for a rare event can still be wrong most of the time it fires.