Skip to main content
Two users rarely behave the same. A lot of the variation you see during a test was already visible before it started — a high spender keeps spending, a heavy user keeps engaging. CUPED (Controlled-experiment Using Pre-Experiment Data) uses that history to cancel out the noise it can explain, so what’s left is a cleaner view of the effect your change actually had. The result: narrower confidence intervals and faster significance, with no change to the lift in expectation.

​
The intuition

For each unit, Traffical looks at a pre-assignment covariate — typically the same metric measured over a window before the unit’s first assignment. It then removes the part of the unit’s measured outcome that was predictable from that covariate, comparing allocations on what remains. Because the covariate comes from before first assignment, it can’t have been affected by the treatment. Adjusting for it is therefore unbiased: CUPED only removes noise, it never invents or hides an effect.
How much you gain depends on how well the past predicts the present. The stronger the correlation between the pre-period covariate and the outcome, the more variance comes out and the tighter the interval — a strongly predictive covariate can cut interval width substantially.

​
Turning it on

CUPED is off by default and controlled by a setting at two levels:
  • Project default — set a default for every policy in the project (for example, “use CUPED everywhere we can”).
  • Per policy — each policy can turn CUPED on or off explicitly, overriding the project default. Left unset, a policy inherits the project default.
The effective rule is simply: policy setting if set, otherwise the project default, otherwise off. Turning the project default on is the easy way to get variance reduction across the board; the per-policy switch is there for the policies where you want to opt in or out individually.

​
What it measures

When CUPED is on, it adjusts by a pre-assignment covariate — the metric’s own value over a window before each unit’s first assignment (for example, the two weeks before it). A unit with no pre-period history simply carries a covariate of zero — the metric’s true pre-period value. CUPED only ever removes noise — enabling it never makes an interval wider or changes the lift in expectation.

​
Configuring it on a metric

CUPED needs two things to take effect, by design:
  1. A covariate on the metric. On the metric itself, enable variance reduction and set the lookback window (how many days before first assignment to measure the covariate). The covariate defaults to the metric’s own value over that window. Available on Conversion and Sum metric types.
  2. CUPED enabled on the policy (or project default), as described above.
If either is missing, the metric is analysed normally — so turning the project default on does nothing until your metrics also carry a covariate, and a metric’s covariate is dormant until a policy enables CUPED.

​
Where it applies

CUPED works on Conversion and Sum metrics — a Sum metric is analysed as a per-unit average, which is what the adjustment operates on. Count and Ratio metrics aren’t adjusted. It composes cleanly with the rest of the engine:

With sequential testing

Variance reduction narrows the confidence sequence too — you keep the anytime-valid guarantee and reach a conclusion sooner.

With winsorization

Outlier capping and CUPED stack: cap the extreme tail, then remove the predictable variation from the rest.

​
When it helps most

  • Metrics with high natural variance between units — revenue, engagement time, order value.
  • Tests where users have history in your product, so the pre-period covariate is meaningful.
  • Any time you want a decision sooner without lowering your confidence level.
For brand-new users with no pre-period history, the covariate is simply zero, so there’s little predictable variation to remove — CUPED helps where history exists and costs nothing where it doesn’t. For the underlying math — the pooled θ estimator, the adjusted mean and variance formulas, and precisely which metric types are adjusted — see the technical companion.