Telecom Customer Churn — Predictive Modeling & Retention Strategy

---

The Business Problem

A telecom company losing 25–30% of customers annually needs to know two things before it can act: which customers are about to leave, and what does it cost to keep them versus replace them?

This project answers both. Five classification models were built, compared, and then the best one was deployed against a new batch of 500 customers to produce actual revenue figures — not just accuracy metrics.

---

Results

74 customers flagged as churn risk out of 500 analyzed.

Metric	Value
Monthly revenue at risk	$6,158
Incentive cost (short-term)	$2,220
Net benefit if retained	$3,938
ROI on incentive spend	177%
Loyalty program net benefit	$128,574

---

What Drives Churn

Five factors from logistic regression odds ratios:

Factor	Effect
Fiber optic internet	3x more likely to churn than DSL
Month-to-month contract	Highest risk segment
Electronic check payment	1.42x higher churn odds
Two-year contract	80% less likely to churn
Tenure (each additional month)	Reduces churn probability

Contract type is the single most actionable lever. Moving a customer from month-to-month to a two-year contract reduces their churn probability by 80%.

---

Model Performance

Model	AUC	Best Accuracy	Threshold
Logistic Regression	0.8355	79.57%	0.6
LDA	0.8289	79.11%	0.6
QDA	0.8180	78.10%	0.9
Naive Bayes	0.8112	77.10%	0.9
Decision Tree (Pruned)	0.8006	78.95%	0.5

Logistic regression was selected — it leads on both AUC and accuracy, and its coefficients read directly as odds ratios, which gives a retention team something concrete to act on.

---

Screenshots

ROC Curves — All Five Models

Churn by Contract Type

Tenure Distribution

---

Project Structure

telecom-churn-prediction/
├── analysis/
│   ├── churn_analysis.Rmd     ← full analysis with business framing
│   ├── churn_analysis.html    ← rendered output (open in browser)
│   └── README.md
├── docs/
│   └── screenshots/
│       ├── roc_curves.png
│       ├── churn_distribution.png
│       ├── tenure_vs_churn.png
│       └── contract_churn_rate.png
├── data/
│   └── README.md
└── README.md

---

Setup

1. Install R and RStudio

R: https://cran.r-project.org

RStudio: https://posit.co/download/rstudio-desktop/

Install R first, then RStudio.

2. Get the Data

See data/README.md for the Kaggle download link. Place telecom_churn.csv (or the .rda files if available) in data/.

3. Install R Packages

install.packages(c(
  "tidyverse", "caret", "ggplot2", "pROC",
  "e1071", "MASS", "naivebayes",
  "rpart", "rpart.plot"
))

4. Run

Open analysis/churn_analysis.Rmd in RStudio. Click Knit → Knit to HTML.

Or open analysis/churn_analysis.html directly in a browser — no R installation required to read the results.

---

Troubleshooting

Problem	Fix
there is no package called 'naivebayes'	install.packages("naivebayes")
cannot open file 'Model_Building_Data.rda'	Place the .rda files in the same folder as the .Rmd
Knit fails	Check the Console — usually a missing package

---

Tech Stack

R · tidyverse · ggplot2 · caret · pROC · MASS · naivebayes · rpart

---

Author

Adarsh Shukla MS Business Analytics · University of Dayton LinkedIn · GitHub