Originally posted by Old Greg
View Post
The tricky thing with probability forecasts is that you can never verify them retrospectively for one-off events. The weather provides a reasonable (contrasting) analogy. If you collect together 1000 cases where a weather forecast suggested a 30% chance of rain (on average), you could at least have some confidence that the probability forecasts were reliable if, when collecting the observed outcomes retrospectively, rain occurred on ~300 of those 1000 occasions. But there are some implicit assumptions here (stationarity, edogicity) that don't apply in one-off referendums, so we'll never know whether the probabilities were any good, retrospectively, for this referendum.

Leave a comment: