A production-ready causal inference project focused on measuring incrementality using Bayesian methods and geo-based experiments.
This repository demonstrates how to estimate true causal lift from marketing interventions using Bayesian synthetic control and geo-experimentation.
Standard attribution models often overstate marketing impact by confusing correlation with causation.
This project answers the real business question:
What would have happened if we had NOT run the campaign?
- Bayesian Synthetic Control
- Geo-based experiment design (test vs control regions)
- Counterfactual outcome estimation
- Posterior credible intervals for lift
- Incrementality (ATT / Causal Lift) estimation
- Counterfactual vs observed time series
- Incremental lift curves with uncertainty
- Cumulative and marginal impact estimates
- Decision-ready summaries for stakeholders
bayesian-incrementality-geo-lift/
├── notebooks/ # Exploratory and methodological notebooks
├── src/causal/ # Core causal modeling logic
├── src/viz/ # Lift & impact visualizations
├── data/ # Simulated geo-level dataset
pip install pymc arviz numpy pandas matplotlibRun the causal model:
python src/causal/geo_experiment.py-
01_synthetic_geo_data.ipynb
- Simulate geo-level time series data
- Define treatment and control regions
-
02_bayesian_synthetic_control.ipynb
- Fit Bayesian synthetic control model
- Estimate counterfactual outcomes
- Compute incremental lift
The dataset provided is fully simulated and included for demonstration purposes only.
No proprietary or confidential data is used.
- PyMC
- ArviZ
- NumPy / pandas
- Matplotlib
- Bayesian causal inference
MIT License