CUPED Variance Reduction Estimator

CUPED shrinks the variance of a numeric metric using pre-experiment data, so your test reaches significance with less traffic. Enter how strongly a user’s pre-experiment metric correlates with their in-experiment metric to estimate how much traffic and time you save.

Your inputs

0.50
0 · none0.3 · weak0.5 · moderate0.7 · strong0.9+

Estimated impact

25%
variance removed (ρ²)
~25%
less traffic / time to significance
A strong correlation of 0.50 removes 25% of the variance. Because required sample size scales with variance, you need roughly the same fraction less traffic — and, at a steady arrival rate, less time.

How to read this

  • Variance reduction = ρ². A correlation of 0.7 removes ~49% of the variance; 0.3 removes only ~9%. The relationship is quadratic, so higher-correlation metrics benefit disproportionately.
  • CUPED only helps numeric metrics (revenue per visitor, pages per session, order value) where users have pre-experiment history. It does nothing for conversion-rate metrics or brand-new users.
  • This is an estimate. Optimizely does not publish a guaranteed number; actual reduction depends on your data. Correlations of 0.3–0.7 are typical for engagement and revenue metrics.

Read the full method in CUPED variance reduction explained, then translate the savings into a concrete sample size and runtime with the traffic & duration estimator.