CUPED (Controlled-experiment Using Pre-Experiment Data) reduces a metric's variance by subtracting off the part of it that was already predictable from a user's own pre-experiment behavior — before the experiment ever started. The statistics behind it are a single formula; the actual difficulty in production is almost always the join structure that builds the covariate correctly.
The Core Idea
Users who spent more before an experiment tend to spend more during it too...