The leadership journal
A highest-value summary looked like a convenient presentation choice. It had actually removed the test history needed to understand why a result existed and who could interpret it.
Read the essayA held-out row is not necessarily evidence about an unfamiliar batch. The validation design should reflect the decision a predictive product is being asked to support.
Read the essayA collection can be useful without being fully reusable or representative. A data product needs an honest account of both the permissions behind it and the selection that produced it.
Read the essay