E-commerce Churn Intelligence

Business Problem

In e-commerce, customer inactivity often leads to silent churn and future revenue loss.
The key challenge is not predicting churn itself, but deciding which customers should be prioritized for retention actions.

This project addresses the question:

Which valuable e-commerce customers are at risk of churn, and how should retention efforts be prioritized?

---

Solution Overview

I built an end-to-end Customer Churn Intelligence pipeline using transactional e-commerce data.

The system:

• defines churn using a business proxy (customer inactivity)
• engineers behavioral features from transaction history
• predicts churn risk using a supervised ML model
• combines customer value and churn risk
• outputs clear retention recommendations

This approach reflects how real-world CDP / MarTech systems support marketing and CRM decision-making.

---

Churn Definition

Since e-commerce data does not include an explicit churn label, churn is defined as:

Customer is considered churned if no purchase was made in the last 90 days.

This inactivity-based definition is a common business proxy used in retail retention analytics.

---

Data

• Source: Kaggle – Online Retail (UK)
• Type: Transactional e-commerce data
• Key fields:
  • CustomerID
  • InvoiceDate
  • InvoiceNo
  • Quantity
  • UnitPrice

All processing is performed locally.

---

Feature Engineering

Customer-level features include:

Core behavioral features

• Recency (days since last purchase)
• Frequency (number of orders)
• Monetary value (total and average spend)

Time-window activity features

• Orders and spend in last 30 / 60 / 90 days
• Activity decay indicators (trend features)

These features capture early warning signals of churn, not just static behavior.

---

Modeling

• Problem type: Binary classification (churn vs non-churn)
• Model: Logistic Regression (baseline, interpretable)
• Output: Churn probability score per customer

The model is treated as a scoring component, not the final product.

---

Decision Layer (Key Value)

The final output is a decision-oriented customer table, combining:

• Churn risk (probability score)
• Customer value segment
• Priority score (Value × Risk)
• Recommended retention action

Example actions:

• Priority retention (high value + high risk)
• Targeted discount campaign
• Monitoring or no action

This layer translates analytics into clear business decisions.

Decisioning Visuals (MarTech / Retention)

Value × Risk Matrix

This matrix helps prioritize retention spend by combining customer value with churn risk.

How to use it (campaign thinking):

• High Value + High Risk → priority retention (personal offer / stronger incentive)
• Mid Value + High Risk → scalable retention campaign (email/ads)
• Low Risk → regular communication, no aggressive discounts

---

Top Retention Targets (Campaign List)

A ready-to-activate list of customers with the highest priority score (Value × Risk).

This output is designed to be used as an activation list for CRM / email / paid retargeting workflows.

Project Structure

ecommerce-churn-intelligence/
├── data/
│   └── raw/                  # raw transactional data (ignored in git)
├── src/
│   ├── load_data.py           # data loading & cleaning
│   ├── churn_label.py         # churn definition logic
│   ├── features.py            # RFM & behavioral features
│   ├── modeling.py            # churn model & scoring
│   ├── decisioning.py         # value-risk decision logic
│   ├── plots.py               # decisioning visuals
│   └── run_pipeline.py        # end-to-end pipeline
├── outputs/
│   ├── customer_features.csv
│   ├── churn_priority_table.csv
│   └── model_metrics.json
├── reports/
│   └── figures/
│       ├── value_risk_matrix.png
│       └── top_retention_targets.png
└── README.md

Key Takeaways

• Churn prediction alone is not enough — decision context matters
• Combining customer value with churn risk enables effective retention prioritization
• This project reflects a production-oriented mindset, not exploratory analysis

---

Next Steps

• Test multiple churn windows (60 / 90 / 120 days)
• Replace value proxy with predictive CLV
• Integrate campaign simulation or uplift modeling