Variance Reduction Below the Randomization Grain

What this post added

Developed and implemented an 'Order-Level CUPED' methodology that trains an order-level prediction model using pre-treatment features. These predictions are then aggregated to the region-day grain, serving as the covariate in the CUPED adjustment. This method demonstrated a 18-40% variance reduction, leading to an average one-third reduction in experiment runtimes, compared to standard region-day CUPED.