Hey Junqi. Thanks for the comment and sorry for the late response!

This point refers to randomization. When you run an AB test, randomization makes group A and group B "identical" on average. That means that potential confounding variables are distributed identically between each group. That's why A/B tests allow for causal conclusions - the only difference between the groups is the treatment.

The key point is that if we are predicting the variance, we can't use something that will be impacted by the experiment (because we may not have evenly distributed confounders). So, the easiest thing to do is use data prior to the experiment which, by definition, can't be impacted by the treatment.

Theoretically, the predictor isn't limited to data prior to the experiment, but our predictor can't be a confounder.