Incremental lift
Also known as: Uplift · Incremental lift · Incremental validity · Gain over the current model
How much a new score adds to the model already in use, measured by the improvement in ordering power when the two run together — not by what the new score does on its own.
Legal basis
Wikipedia — Incremental validity
The question is not whether Zarv's score is good on its own, but whether it adds to what you already have. A signal can have a high Gini and still contribute nothing if it only repeats what the current model already captures. Incremental lift isolates exactly that: it combines your model with the new score and measures how much the ordering power rises over your model alone.
Because it is a difference between two models measured on the same base, it calls for a paired test to tell whether the gain is real or sampling noise — that is what the DeLong test on the two AUCs, with a confidence interval, is for. Without that pairing, a gain of a few Gini points is indistinguishable from luck.
Frequently asked questions
How is a score's incremental lift measured?
By comparing the ordering power (Gini/AUC) of the current model alone against the current model combined with the new score, on the same base, and testing the difference with the DeLong test and a confidence interval.