💰 Customer Lifetime Value (CLV) Analysis

An end-to-end Customer Lifetime Value pipeline using the UCI Online Retail dataset. This project demonstrates two approaches to CLV — progressing from a simple RFM-based model to the industry-standard BG/NBD + Gamma-Gamma probabilistic model — and combines both to create actionable strategic customer segments.

---

🎯 Business Questions

• Which customers are most valuable over the long term?
• How much should we spend to acquire or retain a customer?
• Which customers are at risk of churning despite high historical spend?
• Who are our rising stars worth investing in now?

---

📁 Project Structure

customer_lifetime_value/
├── data/                        # Place OnlineRetail.xlsx here
├── outputs/                     # Charts, CLV predictions, executive summary
├── src/
│   ├── data_prep.py             # Load, clean, build RFM+ customer summary
│   ├── clv_simple.py            # Simple RFM-based CLV (historical average method)
│   ├── clv_bgnbd.py             # BG/NBD + Gamma-Gamma (industry standard)
│   ├── segments.py              # Strategic 2×2 segmentation + recommendations
│   └── visualizations.py       # All charts and summary dashboard
├── main.py                      # Run the full pipeline
├── requirements.txt
└── README.md

---

📊 Dataset

UCI Online Retail Dataset — real-world UK e-commerce transactions (Dec 2010 – Dec 2011).

Download: https://archive.ics.uci.edu/dataset/352/online+retail
Place OnlineRetail.xlsx in the /data/ folder before running.

The data is split into:

• Observation period (75%): used to fit the models
• Validation period (25%): used to evaluate prediction accuracy

---

📐 Methodology

Approach 1: Simple RFM-Based CLV

CLV = Average Order Value × Orders per Year × Years × Gross Margin

Pros: Intuitive, easy to explain, no special libraries required.
Cons: Assumes all customers will keep buying at their historical rate — ignores churn.

Approach 2: BG/NBD + Gamma-Gamma (Industry Standard)

BG/NBD (Beta-Geometric / Negative Binomial Distribution) models two simultaneous processes:

1. Transaction process: while active, a customer purchases at rate λ (Gamma-distributed across customers)
2. Dropout process: after each transaction, a customer has probability p of churning permanently (Beta-distributed across customers)

Gamma-Gamma model estimates expected average transaction value per customer.

Combined CLV:

CLV = E[future purchases | BG/NBD] × E[avg order value | Gamma-Gamma] × gross margin

Discounted to Net Present Value over a 3-year horizon.

Strategic 2×2 Segmentation

	High BG/NBD CLV	Low BG/NBD CLV
High Simple CLV	🏆 Champions	⚠️ At Risk
Low Simple CLV	🌟 Rising Stars	😴 Dormant

Segment	Strategy
Champions	VIP rewards, early access, loyalty programs
At Risk	Personalized win-back campaigns, time-limited offers
Rising Stars	Onboarding sequences, product discovery, habit building
Dormant	Low-cost reactivation; suppress if unresponsive

---

📈 Output Files

File	Description
01_clv_distribution.png	CLV histogram + segment pie chart
02_model_comparison.png	Simple vs BG/NBD CLV scatter plot
03_probability_alive.png	P(customer still active) distribution
04_model_validation.png	Predicted vs actual purchases (validation)
05_strategic_segments.png	2×2 strategic segmentation matrix
06_top_customers.png	Top 20 customers by predicted CLV
07_summary_dashboard.png	Full summary dashboard
clv_predictions.csv	Full customer-level CLV predictions
clv_executive_summary.txt	Business-ready summary with strategies

---

🚀 Getting Started

git clone https://github.com/yourusername/customer_lifetime_value.git
cd customer_lifetime_value
pip install -r requirements.txt
# Add OnlineRetail.xlsx to /data/
python main.py

---

🛠 Tech Stack

• Python 3.14+
• lifetimes — BG/NBD and Gamma-Gamma model fitting
• pandas / numpy — data wrangling
• scipy — statistical utilities
• matplotlib — visualizations
• scikit-learn — preprocessing utilities

---

💡 Key Concepts Demonstrated

• RFM feature engineering from raw transaction data
• Simple CLV with NPV discounting
• Probabilistic CLV modeling (BG/NBD + Gamma-Gamma)
• Train/validation split for time-series model evaluation
• P(alive) estimation and churn probability
• Strategic 2×2 customer segmentation
• Executive-ready reporting and business recommendations

---

👤 Author

Built as part of a data science portfolio project.
Background: 15+ years in Marketing Analytics | SQL | Python | Statistical Modeling

customer_lifetime_value

A CLV model that uses public data and compares the simple method based on AOV with the industry standard BG/NBD + Gamma-Gamma models.

413ce37a5c6498dd747d510bbc4c0597e0504d72