Faster Anytime-Valid A/B Testing via Gradient-Boosted Covariate-Adjusted E-Processes

Faster Anytime-Valid A/B Testing via Gradient-Boosted Covariate-Adjusted E-Processes E-commerce companies continuously A/B test deployed ML systems (recommenders, ranking models, pricing policies) and want to monitor results live, stopping as soon as a decision is clear. But naive continuous monitoring ("peeking") inflates false-positive rate well above nominal, a well-documented failure mode in production experimentation platforms. Anytime-valid inference via e-processes solves this exactly: a nonnegative betting martingale gives Type-I error control at any stopping time, with no correction needed. We combine this with covariate adjustment, the standard variance-reduction tool in industrial A/B testing (CUPED), but replace the usual linear adjustment with a flexible, nonlinear ML model trained on pre-experiment customer data, in the spirit of CUPAC-style adjustment. This combination has not previously been proposed: the closest prior work either pairs anytime-valid testing with linear adjustment only, or pairs ML-based adjustment with fixed-horizon (non-anytime-valid) testing. In a realistic simulated e-commerce A/B test with a nonlinear customer-behavior model, we first confirm Type-I error control by brute-force simulation (naive peeking: 41.8% false-positive rate at 1000 replications; all three e-process variants: ≤ 3.4%, versus a nominal α = 0.05). We then show that ML-based covariate adjustment reduces outcome variance substantially more than linear adjustment (11.5% vs. 3.5% relative to unadjusted), which translates into detecting a true conversion-rate lift both more often (76.0% vs. 72.8% vs. 40.4% detection rate for ML-adjusted, linear-adjusted, and raw e-processes, respectively) and sooner (median 1192 vs. 1266 vs. 1997 paired customers). The method requires no new infrastructure beyond what CUPED-based platforms already have: an offline model trained once on historical data, frozen before the live experiment starts.