← Portfolio · Notebook
Customer Survival Analysis — MTN Nigeria Telecom
Dataset note. This Kaggle snapshot (974 MTN Nigeria customer rows) is approximately random-uniform in the covariates we tested; Date of Purchase is Jan–Mar 2025 for every row while tenure ranges 1–60 months (logically impossible), and chi² on Satisfaction × Churn returns $p = 0.48$. This notebook is a teaching demo of the mechanics of survival analysis — Kaplan–Meier, Cox PH, Weibull AFT, log-rank, C-index. The numerical hazard ratios and PH diagnostics are not substantively interpretable. For a real benchmark see lifelines.datasets.load_lung(), the SUPPORT dataset, or the IBM Watson Telco Customer Churn dataset (7,043 customers).
Business question
Beyond a binary “will they churn?”, how does churn risk change with tenure? Survival analysis lets us:
- Estimate the survival function S(t) — probability a customer is still active at month t
- Identify high-hazard tenure windows (e.g., months 1–6 after onboarding) for targeted retention
- Quantify how covariates (subscription plan, age, satisfaction) shift the hazard
Data
- Source: MTN Nigeria Customer Churn (Kaggle) — 974 customers, 17 attributes including tenure (months), subscription plan, satisfaction, churn outcome.
- Survival framing:
duration = Customer Tenure in monthsevent = 1 if Customer Churn Status == 'Yes', else 0 (right-censored, still active)
Run python download_data.py to fetch (Kaggle CLI required).
EDA targets
- Median time-to-churn vs. censoring
- Hazard differences by subscription plan, age band, satisfaction
- Reasons for churn (free-text categorization)
Modeling
- Kaplan–Meier survival curves stratified by subscription plan and satisfaction
- Cox Proportional Hazards with covariate-adjusted hazard ratios
- Weibull AFT for direct interpretability of expected lifetime
Validation
- Concordance index (C-index) on a 20% held-out test set
- Log-rank tests across strata
Deployment
- API
POST /customer-survival returning expected remaining tenure + churn-risk band given customer attributes
- Retention dashboard ranking customers by 90-day churn probability
Business outcome
- Tenure-aware retention targeting beats blanket “anyone classified as risky”
- Hazard ratios drive product/pricing decisions (e.g., does plan X retain longer than plan Y, controlling for age?)
- Same pipeline applies to any subscription business with tenure + churn signal