Data flow · Causal / Uplift

CUPED variance cut and uplift targeting

CUPED variance cut and uplift targeting Data-flow diagram of the causal-uplift pipeline: a deterministic synthetic generator (data/generate_data.py, seed 42) produces a 20,000-user randomized experiment; its pre-period covariate feeds CUPED variance reduction while the post-treatment outcome feeds T-learner and S-learner uplift models; null references and bootstrap intervals score the segment recovery, and the report states the 26% standard-error cut (about 5,500 users per arm) and the roughly eight-fold new-versus-returning gap. 01GENERATE 02EXPERIMENT 03CUPED 04UPLIFT 05SCORE 06REPORT SOURCEDATA PRE-PERIODCOVARIATE ESTIMATORSCUPED · UPLIFT EVIDENCEREFS · REPORT POST-TREATMENT PRE-PERIOD SCRIPT Data Generator data/generate_data.py DB DATA Experiment Data seed 42 · 20,000 users DB TB DATA Pre-period x_pre covariate · ρ ≈ 0.67 TB MODEL CUPED Adjust y_adj = y − θ(x − x̄) TB TB MODEL Uplift Learners T-learner · S-learner TB TB EVAL ATE + 95% CI -26% SE · 5.5k/arm TB EVAL Null & Bootstrap 200 draws · 95% CI TB EVAL Segment Recovery new 0.111 · ret 0.013 TB OUT Evaluation Report ~8× new vs returning TB FL NODE SCRIPT DATA MODEL EVAL OUT DATA DB DATASET TB TABLE FL REPORT LEFT CHIP = INPUT · RIGHT CHIP = OUTPUT FLOW STANDARD HANDOFF PRE-PERIOD COVARIATE DASHED = PRE-TREATMENT INPUT · ACCENT = FOCAL HANDOFF