More feature flags related terms



CUPED is a statistical technique used to improve the efficiency of controlled experiments, particularly in the context of online A/B testing. The method involves using pre-experiment data to reduce variance in the experiment outcome, thereby increasing the sensitivity and reliability of the experiment results. CUPED adjusts for fluctuations in the metric of interest that are unrelated to the experiment itself, focusing the analysis more directly on the effects of the experimental manipulation.

How It Works

The key to CUPED lies in the utilization of historical data prior to the experiment. By accounting for individual baseline levels, CUPED can effectively reduce random variance in the experimental data that comes from sources other than the experimental manipulation.

  1. Baseline Data Collection: Gather data on the metric of interest from a period before the experiment.
  2. Variance Calculation: Calculate the variance of the pre-experiment metrics and determine a covariance adjustment factor.
  3. Adjustment: Adjust the experimental data using the covariance adjustment factor to control for the pre-experiment variance.
  4. Analysis: Analyze the adjusted data to assess the impact of the experimental changes with reduced variance.



CUPED is particularly useful in online A/B testing where businesses continuously test improvements to web pages, products, or services. It's applied in scenarios where high variability in user behavior can mask the effects of changes, such as in user engagement metrics, conversion rates, or time spent on a page.



CUPED is a powerful tool for enhancing the efficiency and sensitivity of controlled experiments, particularly in environments with high variability. By intelligently incorporating pre-experiment data, researchers and practitioners can significantly improve the reliability of their findings, making more confident decisions based on the outcomes of A/B tests and other experimental methodologies.


Correlation quantifies the degree to which two variables are related, ranging from -1 to +1. It's useful for analysis but doesn't imply causation and has limitations like not capturing non-linear relationships and being sensitive to outliers.

Learn about Correlation

Multivariate testing

Testing multiple variables in a controlled environment to determine which combination yields the best outcomes.

Learn about Multivariate testing

User segmentation

Dividing users into groups based on behavior or attributes for targeted feature releases.

Learn about User segmentation

Start your controlled experiment today

No credit-card required - 30 day trial included