🚀 Customer Analytics Case Study

Segmentation + Churn Prediction + Next-Month Spend Forecast (End-to-End ML System)

An industry-grade data science project that simulates a real-world subscription-based e-commerce business and builds three production-relevant machine learning systems:

• 🎯 Customer Segmentation (Unsupervised Learning)
• ⚠️ Churn Prediction (Classification)
• 💰 Next-Month Spend Forecasting (Regression)

---

🏢 Business Context

Modern e-commerce platforms generate large volumes of customer behavioral and transactional data. However, without structured analytics, businesses fail to:

• Identify high-value customers
• Prevent customer churn
• Forecast revenue accurately

This project builds a data-driven decision system that enables businesses to:

• Segment customers intelligently
• Predict churn risk proactively
• Forecast revenue at a granular level

---

🎯 Business Objectives

1️⃣ Customer Segmentation

• Identify distinct customer groups
• Enable targeted marketing strategies
• Improve campaign efficiency

2️⃣ Churn Prediction

• Predict probability of customer churn
• Enable early retention intervention

3️⃣ Spend Forecasting

• Predict next-month customer spending
• Support revenue planning and personalization

---

🧠 Data Strategy (Simulation with Realistic Assumptions)

Dataset Scale

• 👥 Customers: 5,000
• 🛒 Transactions: ~40,000
• 📅 Simulated over multiple months

---

📊 Features Generated

👤 Customer Attributes

• Age (18–70)
• Income ($5K – $150K)
• Country (US 60%, UK 40%)
• Tenure (1–60 months)

📱 Behavioral Features

• Visits per month (Poisson distributed)
• Engagement intensity

🛒 Transaction Features

• Total orders
• Total spend
• Average order value (AOV)
• Discount usage patterns

---

🔎 Exploratory Data Analysis (EDA)

Key Statistics

• Average monthly visits: 5.8 visits/user
• Average order value: $72.4
• Average spend per customer: $1,180
• High-income segment contributes ~48% of total revenue

---

Observations

• Income distribution is right-skewed
• Engagement strongly correlates with spending
• Low-activity users show early signs of churn

---

🧾 Feature Engineering

Aggregated Features

• Total Spend
• Total Orders
• Avg Order Value
• Visit Frequency

Derived Features

• Spend per visit
• Engagement ratio
• Tenure-adjusted activity

---

🔷 Part A — Customer Segmentation (KMeans + PCA)

⚙️ Approach

• Feature Scaling using StandardScaler
• PCA → Reduced to 2 components (~81% variance explained)
• KMeans clustering

---

🔍 Optimal Cluster Selection

• Elbow Method → K = 5

---

📊 Cluster Summary

Cluster	Size	Avg Spend	Avg Visits	Segment Type
0	1,120	$1,950	8.2	High-value engaged
1	980	$420	2.1	Low-value inactive
2	1,050	$1,200	5.4	Mid-tier steady
3	890	$300	3.0	Price-sensitive
4	960	$2,400	9.5	Premium users

---

💡 Insights

• Top 20% customers generate ~55% revenue
• Low-engagement users (<3 visits/month) have 2.5x higher churn risk
• Premium segment shows highest retention

---

🔷 Part B — Churn Prediction (Classification)

⚙️ Models Evaluated

• Logistic Regression
• Random Forest ✅ (Best)

---

📊 Model Performance

Metric	Logistic Regression	Random Forest
Accuracy	0.78	0.86
Precision	0.74	0.83
Recall	0.69	0.81
F1 Score	0.71	0.82

---

🔍 Key Drivers

• Low visits/month
• Low spend
• Short tenure
• High discount usage

---

💡 Insights

• Users with <3 visits/month → 68% churn probability
• Users with tenure >24 months → <15% churn probability

---

🔷 Part C — Spend Forecasting (Regression)

⚙️ Model Used

• Linear Regression

---

📊 Model Performance

• RMSE: $142.6
• R² Score: 0.72

---

🔍 Key Drivers of Spend

1. Historical spend
2. Visit frequency
3. Avg order value
4. Tenure

---

💡 Insights

• Users with ≥8 visits/month spend 2.3x more
• High AOV users dominate revenue growth

---

🏗️ End-to-End Pipeline

Data Simulation
   ↓
EDA & Validation
   ↓
Feature Engineering
   ↓
Segmentation (K-Means + PCA)
   ↓
Churn Prediction (Random Forest)
   ↓
Spend Forecasting (Linear Regression)
   ↓
Business Insights

---

---

📊 Key Business Insights

• Top 20% customers generate >50% of revenue
• Engagement is the strongest predictor of churn
• Segmentation enables targeted marketing
• Spend prediction improves revenue planning

---

🧠 Key Learnings

• Built multi-model ML pipeline (unsupervised + classification + regression)
• Applied feature engineering for behavioral analytics
• Connected ML outputs to business decisions
• Learned trade-offs between model types

---

🧰 Tech Stack

• Python
• Pandas / NumPy
• Matplotlib / Seaborn
• Scikit-learn
• KMeans
• PCA
• Random Forest
• Linear Regression

---

📁 Project Structure

Customer-Analytics-Case-Study/
│
├── CASE_STUDY_DS.ipynb
├── README.md

---

---

🔥 Future Improvements

• XGBoost / LightGBM models
• Time-series forecasting
• FastAPI deployment
• Model monitoring
• Cloud deployment (AWS/GCP)

---

👨‍💻 Author

Birjung Thapa
Master’s in Data Science
University of Colorado Boulder

---

⭐ Support

If you found this project useful:

⭐ Star the repo
📢 Share with others
🤝 Connect for collaboration