Outlier Capping Calculator

Test an outlier cap on a simulated A/B test before applying it to your own revenue metric:

1. Describe your customers. Set the share of customers who buy anything during the experiment, and the share of revenue that comes from your top 1% of buyers. That second number decides most of what capping can gain.
2. Describe your A/B test. Set the number of users and the true effect on revenue per customer. Then choose whether your feature adds the same dollar amount to every buyer’s spend or the same percentage. A percentage effect puts more of the gain on your biggest buyers, which a cap removes.
3. Set the outlier cap and compare. Pick the cap percentile and whether zero-spend customers count toward it. Then compare the precision gained, the effect given up and the chance of detecting a real effect, with and without the cap. Try a deeper cap to see where the trade-off stops paying off.

Winsorization is a way to cap outliers in an A/B test metric. Every value above a chosen percentile, such as the 99th percentile of buyer revenue, is replaced with that percentile before the experiment is analyzed. It trades a small, deliberate bias for much less noise, so revenue experiments reach significance faster. GrowthBook supports percentile and absolute caps on capped metrics. For when capping helps and when it hurts, read Should you cap outliers? Winsorization in A/B testing.

‍

Check what share of your total revenue comes from your top 1% of buyers, counting only customers who bought something. That number most decides how much capping outliers (winsorization) helps. If the top 1% hold only a few percent of revenue, a cap changes almost nothing. If they hold 10–30%, capping at the 99th percentile can cut the noise in a revenue metric enough to be worth doubling your sample size or more. The GrowthBook Outlier Capping Explorer shows the exact trade-off for your numbers.

‍

Usually not. When most customers buy nothing, including their zero values pushes the same percentile cap much deeper into your buyers. For example, if 10% of customers buy, a 99th percentile cap calculated over all customers caps the top 10% of buyers rather than the top 1%. That removes far more of the treatment effect than intended. GrowthBook lets you set capped metrics either way. The Outlier Capping Explorer shows the difference with zeros excluded or counted.

The explorer simulates thousands of A/B tests using the revenue distribution, effect type and cap percentile you choose, so its numbers describe a typical experiment with those settings, not your own data. Use it to judge the size of the trade-off between precision gained and effect removed before you choose a cap. Then compare the raw and capped results in your real experiment. In GrowthBook, the Debug tab of the metric drilldown shows both.