🎵 KKBox Subscription Retention: A 360° Survival Analysis

Predicting Subscriber Lifespan and Identifying Churn Drivers in 2.4M+ Users

Project Overview

Subscription churn is the silent killer of SaaS revenue. Using the KKBox dataset (3.1M+ transactions), this project moves beyond simple churn rates to perform a deep-dive longitudinal analysis. By triangulating Cohort Analysis, Kaplan-Meier Survival Curves, and Cox Proportional Hazards, I identified not just when users leave, but the specific financial and behaviuoral "hazards" that cause them to exit.

Tech Stack

• Language: Python, SQL
• Libraries: Pandas, Lifelines (Survival Analysis), Matplotlib/Seaborn (Visualisation), NumPy

Dataset: 2.4M unique users | 3.1M subscriptions

Source: WSDM Churn Prediction Challenge on Kaggle

• transactions.csv 21.5 million rows - Historical transaction data from January 2015 through February 2017
• transactions_v2.csv 1.3 million rows - Refreshed transaction log including March 2017
• members_v3.csv 6 million rows - User demographis

---

Business Insights (TL;DR)

• Churn nearly doubles in Month 2 (22.3%) compared to Month 1 (12.4%), identifying a critical renewal friction point.
• A typical subscriber stays for 184 days (~6 months).
• Users on Payment Method 20 are 3x more likely to churn (Hazard Ratio = 3.0) than those on the baseline method.
• 1-year payment plans show near-infinite survival, while 7-day trials are churn magnets with a median survival of only 8 days.

---

Phase 1: The "When" - Cohort Analysis

I analysed 27 monthly cohorts over a 24-month observation window to establish retention benchmarks.

• Retention stabilises significantly for users who survive the first 90 days.
• The Jan 2015 "Launch Cohort" (549k users) exhibited significantly higher long-term loyalty (78.9% Year 1 retention) compared to subsequent steady-state cohorts.
• Identified a major drop-off in the June 2015 cohort (only 25.8% Month 1 retention), likely signaling a technical failure in the signup flow or a low-intent marketing campaign.

Figure 1: Retention Heatmap identifying the 60-day churn spike and June 2015 anomaly.

---

Phase 2: The "How Long" - Kaplan-Meier

Using Kaplan-Meier estimates, I quantified the probability of a user reaching specific milestones.

• Only 42.5% of users reach the 1-year mark, but those who do represent the highest-LTV (Lifetime Value) segment.
• Standard 30-day plans dominate the step-down churn pattern, whereas long-term commitment plans (180+ days) maintain high survival throughout the lifecycle.

Figure 2: Survival probability across all subscriptions with a median survival of 184 days.

---

Phase 3: The "Why" - Cox Proportional Hazards

I used a Stratified Cox Model to identify which variables significantly increase the risk (Hazard) of churn.

Variable	Hazard Ratio ($HR$)	Impact
Payment Method 20	3.01	200% Increase in Churn Risk
Registration Channel -1	2.08	108% Increase in Churn Risk
Registration Channel 7	0.68	32% Reduction in Churn Risk
Payment Method 34	0.34	66% Reduction in Churn Risk

Technicality: Validated model assumptions using Proportional Hazard tests. While the massive scale of the data (3M+ rows) flagged violations, stratification by plan duration was used to ensure robust directional estimates.

Figure 3: Hazard Ratios for Payment Methods and Registration Channels.

---

Recommendations

1. Target the 45-60 Day Window: Launch automated "Value-Add" email campaigns 15 days before the second renewal to mitigate the observed Month 2 spike.
2. Financial Migration: Incentivise users on high-hazard methods (Method 20) to migrate to Method 34 (e.g., through a one-time discount), as Method 34 correlates with a 66% risk reduction.
3. Product Fit: Shift acquisition spend away from 7-day trials toward 30-day or 90-day introductory offers to improve the initial "Success Probability."

---

📂 Project Structure

• notebooks/kkbox_survival_data.ipynb: Initial data cleaning and processing of 22M+ rows.
• notebooks/kkbox_cohort_retention_analysis.ipynb: Implementation of Cohort Analysis
• notebooks/kkbox_survival_exploratory_analysis.ipynb: Implementation of Kaplan Meier
• notebooks/kkbox_cox_regression.ipynb: Implementation of Cox Regression
• visuals/: Plots from notebooks

---

How to Run

1. Clone the repo.
2. Install requirements: pip install pandas lifelines matplotlib seaborn.
3. You can access the data via the WSDM Churn Prediction Challenge on Kaggle.
4. Run the notebooks