An end-to-end causal machine learning system for identifying customers whose purchasing behavior can be influenced by marketing treatment.
The project uses a 200K-customer randomized retail experiment and builds a leakage-safe feature pipeline from 45.8M transaction records using SQL and DuckDB.
Highlights
- Causal inference and uplift modeling
- T-Learner and EconML CausalForestDML
- 47 leakage-safe features
- 15.5 percentage-point purchase uplift in the top 5% targeted customers
- XGBoost propensity targeting achieved only 1.1 percentage-point uplift
- Qini AUC comparison between causal and propensity targeting
- 500-repetition bootstrap confidence intervals
- Budget-constrained targeting optimizer
- Interactive Streamlit simulator
- Dockerized application
- 157 pytest tests, including a test-set firewall against data leakage