📞 Will They Stay or Will They Go? Telco Customer Churn Analysis

Can we predict which customers will cancel their telecom subscription and why?
Logistic regression in both R and Python on 7,043 customer records.

❗ROC Curve

---

📋 Table of Contents

• Problem
• Data
• Methods
• Key Results
• Business Insights
• How to Reproduce
• What I'd Do Next

---

Problem

Customer churn is one of the most costly problems in subscription businesses. This project builds a logistic regression model to:

1. Predict whether a customer will churn (binary outcome: Yes / No)
2. Identify the strongest drivers of churn using odds ratios
3. Evaluate model performance using AUC, confusion matrix, and classification report
4. Deliver actionable business recommendations

The same analysis is done in both R and Python comparing coefficients across both implementations.

---

Data

Property	Detail
Source	Kaggle Telco Customer Churn (IBM)
Size	7,043 customers × 21 columns (11 blanks dropped → 7,032 clean rows)
Target	Churn Yes (26.6%) / No (73.4%)
License	IBM Sample Data via Kaggle

Key Variables

Variable	Type	Description
tenure	numeric	Months as a customer
MonthlyCharges	numeric	Monthly bill ($)
TotalCharges	numeric	Total amount billed ($)
Contract	factor	Month-to-month / One year / Two year
InternetService	factor	DSL / Fiber optic / No
TechSupport	factor	Yes / No / No internet
OnlineSecurity	factor	Yes / No / No internet
PaperlessBilling	factor	Yes / No
SeniorCitizen	binary	1 = senior citizen
Churn	target	Yes / No

---

Methods

R (R/churn_analysis.R)

Step	Function	Package
Data cleaning	drop_na(), as.numeric()	tidyverse
EDA	ggplot2 bar + box + histogram	ggplot2
Train/test split	createDataPartition()	caret
Logistic regression	glm(family = binomial)	base R
Odds ratios + CI	tidy(exponentiate = TRUE)	broom
Confusion matrix	confusionMatrix()	caret
ROC / AUC	roc(), auc()	pROC
Hypothesis test	Anova(type = "II")	car

Python (Python/churn_analysis.py)

Step	Function	Package
Data cleaning	pd.to_numeric(errors="coerce")	pandas
EDA	boxplot, hist	matplotlib
Preprocessing	ColumnTransformer + StandardScaler + OneHotEncoder	scikit-learn
Logistic regression	LogisticRegression(class_weight="balanced")	scikit-learn
Cross-validation	cross_val_score(scoring="roc_auc")	scikit-learn
ROC curve	RocCurveDisplay	scikit-learn
Odds ratios	sm.Logit().fit() + np.exp(params)	statsmodels

---

Key Results

Model Performance

Metric	R	Python (sklearn)
AUC	~0.84	~0.84
Accuracy	~0.80	~0.77
Precision (Churn)	~0.65	~0.60
Recall (Churn)	~0.55	~0.70

Both implementations converge to the same AUC ≈ 0.84, validating the analysis.

Top Churn Drivers (Odds Ratios)

Predictor	Direction	Interpretation
Two-year contract	↓ churn	~75% lower odds vs month-to-month
One-year contract	↓ churn	~40% lower odds vs month-to-month
Fiber optic internet	↑ churn	Higher churn vs DSL
No TechSupport	↑ churn	Customers without support churn more
No OnlineSecurity	↑ churn	Security add-on linked to retention
Tenure (per month)	↓ churn	Longer customers are more loyal
Paperless billing	↑ churn	Slight increase in churn risk

6-Step Hypothesis Test: Contract Effect

1. α = 0.05
2. H₀: β_One-year = β_Two-year = 0 (no contract effect on churn)
3. Decision rule: Reject H₀ if p-value ≤ α
4. Test statistic: χ² (from Type II likelihood-ratio test)
5. p-value < 0.001
6. Conclusion: Reject H₀ contract type significantly predicts churn

---

Business Insights

1. Promote long-term contracts at signup. Month-to-month customers churn at ~43% vs ~3% for two-year customers. Even a small incentive to sign a one-year contract could meaningfully reduce attrition.

2. Bundle TechSupport and OnlineSecurity. These services are among the strongest protective factors. Consider offering them free for the first 3 months.

3. Target new customers aggressively. Churn risk is highest in the first 12 months of tenure. A proactive outreach program in months 1–6 could significantly improve lifetime value.

4. Fiber optic customers are high-risk. Despite paying more, fiber customers churn at higher rates possibly due to price sensitivity or competition. Investigate service quality and pricing.

---

How to Reproduce

Download the data

1. Go to: https://www.kaggle.com/datasets/blastchar/telco-customer-churn
2. Download WA_Fn-UseC_-Telco-Customer-Churn.csv
3. Place it in the data/ folder

R

install.packages(c("tidyverse","broom","caret","pROC","car"))
source("R/churn_analysis.R")

Python

pip install pandas numpy matplotlib seaborn scikit-learn statsmodels
python Python/churn_analysis.py

---

What I'd Do Next

• Try Lasso logistic regression (glmnet in R / LogisticRegression(penalty="l1") in Python) for automatic variable selection
• Compare against a Random Forest baseline to see if non-linear methods outperform logistic regression
• Build a Shiny / Streamlit dashboard where a business user can input customer attributes and get a churn probability score
• Address class imbalance more rigorously with SMOTE (themis in R / imbalanced-learn in Python)

---

Data: IBM Sample Dataset via Kaggle (blastchar). 7,043 customers, 21 variables.