Base rate
Brainteasers & GuesstimatesHow common something is before any evidence arrives, and the input that decides what a positive test result is actually worth.
Also written: prior probability, unconditional probability
The base rate is the unconditional frequency of an event: the share of borrowers who default, of expense claims that are fraudulent, of companies that breach a covenant. It is the starting point that evidence moves you away from, and no test result can be interpreted without it.
Base rate neglect is the reliable failure. Told that a model catches 90% of defaults, most people conclude that a flagged borrower is very likely to default. That conclusion is only available if defaults are common. Where they are rare, the flags generated by the very large healthy population swamp the flags generated by the small defaulting one.
The practical consequence for anyone building or buying a screening tool is that the false positive rate matters more than the hit rate whenever the event is rare. Pushing detection from 90% to 95% adds a handful of true flags. Halving the false positive rate removes hundreds of false ones.
It also explains why a screen that looks useless can still be valuable. A flag that raises the probability from 4% to 32% has multiplied the odds eightfold even though most flags are wrong. The right conclusion is that a flag is a reason to look, not a verdict.
Worked example
A book of 10,000 borrowers where 4% default: 400 defaulters, 9,600 sound.
A model catching 90% of defaults flags 360 of the 400. Flagging 8% of the sound borrowers adds 768 more, so 1,128 flags produce 360 genuine ones, about 32%.
Change nothing but the base rate. At 20% defaults the same model gives 1,800 correct flags against 640 false ones, and a flag now means about 74%. The model did not improve, the population did.