CUPED

What it means

CUPED reduces the noise in an experiment by adjusting each user’s outcome for how they behaved before the test began. Someone who already spent a lot last month is expected to spend a lot this month regardless of your change, so CUPED subtracts that predictable part out. What remains is a cleaner signal of the treatment effect. Mechanically it uses a pre-period covariate, correlated with the metric, to strip out variance that has nothing to do with the variant. The point estimate stays unbiased; the confidence interval around it gets narrower.

Why it matters

Variance is what forces long tests and large samples. By cutting it, CUPED can shrink the traffic or time needed to reach statistical significance, often by a meaningful margin on metrics with strong pre-period signal like revenue or engagement. For teams without huge traffic, that difference decides whether an experiment is feasible at all.

In practice

CUPED is a marker of a serious stats engine rather than a toggle with a chart attached. GrowthBook lists CUPED alongside sequential testing, Bayesian analysis, bandits and SRM checks as part of its warehouse-native engine, and Statsig ships CUPED variance reduction in its experimentation platform. Both need clean, well-instrumented pre-period data for it to help, which is the real cost. See best experimentation platforms for tools with CUPED built in.