Three questions
What drives it?
Permutation importance on a 30% holdout, with the holdout fit quoted. Arithmetic identities (revenue → margin) are called out, not sold as insight.
Is this one unusual?
Not against everyone — against its peers: the narrowest group with enough members where it still stands out.
Did the change work?
Difference-in-differences: the change group against a comparable group over the same months, with the assumption stated.
Questions
How is a column’s influence measured?
On rows the model never saw. A 30% holdout is set aside, each column is shuffled in turn, and the drop in accuracy is its influence. The holdout fit is quoted with the result, so a weak model is visible.
Why does it say revenue explains margin?
It does not sell that as insight. When one column is arithmetic on another, the answer calls it an identity and moves on to the drivers you can act on.
Is “did the change work” a causal answer?
It compares the changed group with a comparable group over the same months. That holds if both groups would have moved alike without the change, and the answer states that assumption every time.