What "1 in 14 million" actually feels like, why expected value is always negative, and how to enjoy a ticket without fooling yourself.
A treatment can look worse overall and better in every single group at the same time. It sounds impossible. It happens constantly.
Peeking, stopping early and testing twenty metrics until one looks good are the three classic ways to ship a "winner" that isn't. How to run a test you can defend.
When a billionaire walks into a café, the average income in the room becomes enormous and completely useless. Why skewed data needs the median.
Ice cream sales and shark attacks rise together. Banning ice cream will not save swimmers. A friendly guide to confounders and how not to embarrass yourself in a meeting.
One train/test split can be lucky. Cross-validation asks the same question five times so you can trust the answer, with a few traps for time series and grouped data.
No posts here yet.