forecasting//base rate

A base rate is the frequency with which an outcome occurs in a reference class of similar cases, and it is the starting point of a forecast before any evidence about the specific case is considered. If one drug candidate in ten that enters clinical trials reaches approval, a new candidate starts at about 10%, whatever its story.

A base rate is the frequency with which an outcome occurs in a reference class of similar cases, and it is the starting point of a forecast before any evidence about the specific case is considered. If one drug candidate in ten that enters clinical trials reaches approval, a new candidate starts at about 10%, whatever its story.

People overweight vivid, specific evidence and underweight the base rate, a bias called base-rate neglect. In Bayes' rule the base rate is the prior, and a forecast that ignores it is one that has skipped the prior.

The reference class is a judgement, and it changes the answer. Predictions that AI progress would hit a wall have failed repeatedly, which gives a low base rate for such claims; a forecast about one physical input (chip packaging, power plants) belongs to a different class, with its own record.

A base rate is a starting point, never a verdict. Strong specific evidence should move the estimate far from it; weak evidence should barely move it.