Combine BG/NBD for purchase frequency with Gamma‑Gamma for transaction value to project customer lifetime value and guide acquisition and retention spend.
Data Requirements
- Transactions:
customer_id,timestamp,amount; optional channel, product/category, acquisition date - Define calibration and holdout windows (e.g., 6M + 3M) to validate forecasts
Model Overview
- BG/NBD: models repeat purchasing with dropout (churn) after any purchase; parameters capture heterogeneity across customers
- Gamma‑Gamma: models average transaction value assuming value is independent of purchase frequency
- LTV: expected number of future transactions × expected value per transaction (optionally discounted)
Validation
- Use a holdout period to check calibration (purchase counts, spend) and error metrics
- Inspect cohort retention curves and compare predicted vs actual in holdout
- Re‑estimate periodically to adapt to behavior shifts
Assumptions & Caveats
- Stationarity in purchasing probabilities; independence of value and frequency in Gamma‑Gamma
- Non‑contractual settings; for subscription churn, use survival models instead
- Segment by channel or product when behavior is heterogeneous
Applications
- Bid and budget setting by predicted LTV
- Target lifecycle programs and win‑back to high‑value prospects
- Explain growth through cohort LTV and retention composition