Data Science Portfolio Project

Telecom Churn Intelligence
Revenue-at-Risk Modeling & Retention Effectiveness

Predicted customer churn, quantified revenue at risk using CLV modeling, and statistically validated a retention campaign via A/B testing — delivering a complete business intelligence system.

 GitHub Repo ↗ Dataset

Dataset Size

7,032

customers · 21 features

Churn Rate

26.58%

1,869 customers at risk

Best Model ROC-AUC

0.83

Random Forest

Revenue at Risk

₹3.56L

identified portfolio risk

Campaign ROI

1123%

p < 0.0001 · statistically proven

Phase 1

Exploratory Data Analysis

Key patterns identified across contract type, payment method, tenure, and monthly charges.

42.7%

Month-to-month contract customers churn — 15x higher than two-year contracts

45.3%

Electronic check users churn — highest among all payment methods

41.9%

Fiber optic internet users churn — despite paying the highest charges

Churn rate by contract type

Churn rate by payment method

Tenure distribution by churn status

Monthly charges by churn status

Phase 3

Statistical Analysis

Hypothesis testing confirms all key drivers are statistically significant (p < 0.0001).

Chi-Square Tests — Categorical vs Churn

Contract type vs Churn ✓ p < 0.0001

Electronic check vs Churn ✓ p < 0.0001

Fiber optic vs Churn ✓ p < 0.0001

Fiber optic Chi² statistic 663.36 (strongest)

T-Test & Mann-Whitney U — Numeric vs Churn

MonthlyCharges T-Test ✓ T = 16.48

Churners avg MonthlyCharges ₹74.44

Non-Churners avg MonthlyCharges ₹61.31

Tenure Mann-Whitney U ✓ p < 0.0001

ANOVA — MonthlyCharges across Tenure Groups

New customers (≤12 mo) ₹56.17 avg

Mid customers (13–36 mo) ₹63.25 avg

Loyal customers (36+ mo) ₹72.01 avg

F-Statistic ✓ F = 187.50

Cramér's V — Association with Churn

TenureGroup V = 0.34 ▲ strongest

Fiber optic service V = 0.31

Two-year contract V = 0.30

Electronic check V = 0.30

Phase 4

ML Modeling

SMOTE applied on training set only (4,130 vs 4,130). Two models trained and compared.

Logistic Regression

VIF-cleaned dataset · 10 features · Odds Ratios interpreted

ROC-AUC

0.79

Churn Recall

81%

Churn F1

0.57

Accuracy

67%

Key Odds Ratios

Electronic check: 2.27x more likely to churn

Two-year contract: 0.03x (97% less likely)

Random Forest Best Model

Full 27 features · max_depth=10 · 100 estimators

ROC-AUC

0.83

Churn Recall

74%

Churn F1

0.62

Accuracy

76%

Top features by importance

Tenure 16.9% TotalCharges 9.8% MonthlyCharges 9.3% Electronic check 9.2% Two-year contract 8.9%

ROC-AUC comparison — Logistic Regression vs Random Forest

Phase 6

A/B Testing — Retention Campaign

297 high-risk customers (churn prob > 0.7) split into treatment and control groups. Treatment offered a 20% discount.

Campaign effectiveness — statistical validation

27.15%

Churn Reduction

<0.0001

P-Value (Z-Test)

-5.03

Z-Statistic

24.06

Chi² Statistic

H₀: Retention campaign has no effect on churn rate

H₁: Campaign significantly reduces churn rate

✓ Reject H₀ — Campaign effectively reduced churn (p < 0.0001)

95% Confidence Interval: churn reduction between 17.02% and 37.28%

Churn rate — treatment vs control

Revenue impact breakdown

Campaign Return on Investment

Revenue Saved: ₹23,863 | Campaign Cost: ₹1,951 | Net Profit: ₹21,912

41 customers retained × avg CLV ₹582

1123%

ROI

Stack

Tech Stack

Tools and libraries used across all phases.

Python · Pandas, NumPy

Statistics · SciPy, Statsmodels, Pingouin

ML · Scikit-learn, Imbalanced-learn

Visualization · Matplotlib, Seaborn

Dashboard · Streamlit

Models · Logistic Regression, Random Forest

Sampling · SMOTE