Reducing Churn in E-Commerce: My End-to-End Capstone Project in Predictive Modeling

Customer churn isn’t just a marketing problem - it’s a business survival issue. In competitive industries like e-commerce, losing one customer often means losing several revenue streams, especially when one account can represent multiple users.

This post is a breakdown of my churn prediction capstone project for the postgraduate data science program at UT Austin - also tied to my master’s in data science at Deakin U. The project was closed-source, so I can’t release the full notebook, but I’ll walk you through everything I did including code snippets, results, charts, what I learned, and where this project fits in my larger journey into machine learning and MLOps.

🧩 The Problem: Why Churn Matters So Much

The e-commerce company in question faced intense competitive pressure and a rising churn rate. Each lost account didn’t just represent a single user - it impacted groups of customers. That meant multiplicative revenue loss.

Our goals were:

• Predict churn at the account level
• Identify top drivers of churn
• Segment customers for targeted retention
• Recommend cost-effective, revenue-positive strategies

🧪 Dataset Overview & Initial Processing

The dataset had ~11,000 rows and 19 columns, combining both categorical and numerical variables. A few key points:

• Heavy class imbalance: Only ~17% of the accounts had churned
• Several columns had missing values which were handled via KNN imputation
• Mix of behavioral data (logins, payments, complaints), service scores, and engagement metrics

🔍 Exploratory Data Analysis (EDA)

Key univariate and bivariate findings:

• Device Used: Desktop users churned more (~20%) than mobile (~16%)
• Revenue Patterns: Lower revenue/month was linked to higher churn
• Complaints: Past complaints were strongly correlated with churn
• Tenure: Shorter tenure customers churned more
• Payment Method: Debit and UPI users churned more than Credit Card users

Interestingly, some expected predictors (like cashback or service score) didn’t show strong correlation individually - indicating multi-factorial churn behavior.

📊 Clustering for Segmentation

Rather than just modeling churn, we wanted actionable segmentation. Using the elbow method and KMeans, we identified 4 customer clusters.

Each cluster showed different churn profiles, behavior patterns, and required tailored business strategies.

📌 Cluster Highlights & Strategic Takeaways

Cluster 1 – Large Segment, High Churn (~20%)

• Male-dominated, married, mobile-first
• Mid-to-high revenue, but frequent complaints
• Strategy: Premium customer care, non-discount loyalty programs, debit-card reward partnerships

Cluster 3 – Smallest, but Worst Churn (~21%)

• Newer users with high cashback dependency
• Low service scores, high complaints, revenue decline
• Strategy: Early engagement triggers, reduce cashback reliance, introduce tiered perks

Cluster 0 – Mid-Tier Churn (~13%)

• HNI and Super users, higher spenders
• Credit card and UPI users, moderate complaints
• Strategy: VIP retention plans, gamified reward systems, re-engagement emails

Cluster 2 – Best Segment (~5% Churn)

• Long-tenure, high-revenue customers
• Highest satisfaction, lowest complaints
• Strategy: Just keep them happy! Introduce premium perks and incentivize digital payments

🤖 ML Model Evaluation & Results

This wasn’t just a segmentation project - I evaluated 6 different machine learning models for churn prediction:

• Random Forest & XGBoost were the top performers with high precision and AUC scores, and no signs of overfitting, so I ended up tuning the random forest further.

• I also experimented with SMOTE for handling class imbalance, but performance dropped significantly across models so oversampling isn’t always the solution.

🔍 Feature Importance (XGBoost)

Top predictors of churn from the model:

• Tenure and Complaints were the most influential features.
• Other strong indicators: customer care interaction history, cashback dependency, and service score.

🎯 Business Outcomes & Learnings

This wasn’t just a modeling exercise - it was a strategy recommendation engine.

Key takeaways:

• Churn is complex: No one variable dominates, but multiple small signals do.
• Clustering works: It lets us build segment-specific retention tactics.
• Complaints matter: Past complaints + low satisfaction are gold for early churn signals.
• Cashback ≠ loyalty: High discount reliance often signals at-risk behavior.

🧭 Business Recommendations by Cluster

Cluster 1 (Large Segment, High Churn ~20%)

• Premium support: Faster complaint resolution and priority escalation.
• Loyalty rewards without discounts: e.g., early access perks.
• Debit-card partner offers & wallet payment nudges.
• Predictive churn alerts + personalized retention bonuses.

Cluster 3 (Smallest, Highest Churn ~21%)

• High-touch onboarding: Loyalty bonuses in first 3–6 months.
• Reduce cashback reliance: Use membership perks instead.
• VIP service model for high revenue churners.
• Improve care response time: Add real-time channels.

Cluster 2 (Lowest Churn ~5%)

• Maintain delight: Premium memberships, early sales access.
• Encourage digital payments with UPI/wallet incentives.
• Reward repeat purchases via tiered cashback.

Cluster 0 (Moderate Churn ~13%)

• HNI support: Personalized account managers.
• Gamified engagement: Spend & Earn reward mechanics.
• Proactive re-engagement emails.
• Introduce premium subscription loyalty tier.

🚀 What’s Next?

This project helped me:

• Think beyond models and into business strategy
• Practice handling imbalanced data
• Understand how to build insights*with both EDA and clustering

It’s part of a broader set of real-world projects I’ll be tackling throughout my second year of the master’s program - with more blogs and eventually open-sourced templates (when possible).

If you’re interested in just nerding out - feel free to connect or drop a note.

Thanks for reading! 🧠📊