Telco Customer Churn — Analysis & Prediction

Analysis of 7,000+ telecom customer records to understand what drives churn and to build a baseline model that predicts churn risk. The work combines exploratory data analysis and a logistic-regression classifier in Python, with an interactive Power BI dashboard for the business view.

---

Project Overview

The problem. Telecom companies lose revenue when customers leave. Understanding why customers churn — and flagging who is at risk — supports proactive retention.

What this project does:

• Exploratory Data Analysis (EDA) to surface churn patterns across contract, service, payment, charges, and tenure
• A logistic-regression baseline model for binary churn classification
• A Power BI dashboard summarising churn KPIs and driver analysis

---

Key Results

• Model accuracy: 78.7% on a held-out 20% test set (logistic regression)
• Churn class — precision 0.62, recall 0.52, F1 0.56; retained class — precision 0.83, recall 0.89, F1 0.86
• EDA shows churn concentrated among month-to-month contracts, fiber-optic internet, electronic-check payments, higher monthly charges, and shorter tenure

(Metrics come directly from the classification report in the notebook. See Limitations for why recall on the churn class is the main area for improvement.)

---

Dataset

Attribute	Details
Source	IBM Telco Customer Churn (Kaggle)
Size	7,043 records (7,032 after cleaning)
Features	20 features used (demographics, services, contract, billing)
Target	Churn (Yes / No — binary)
Class balance	~26% churn, ~74% retained

---

Approach

Data cleaning

• Dropped customerID (identifier, not predictive)
• Converted TotalCharges to numeric and dropped the 11 blank rows (~0.15% of data)

Feature engineering

• One-hot encoded categorical variables with drop_first=True to avoid multicollinearity
• Standardised numerical features with StandardScaler

Modelling

• 80/20 train–test split (random_state=42)
• Logistic regression (max_iter=5000) as an interpretable baseline
• Evaluated with accuracy and a per-class classification report

---

Repository Structure

Telco_Customer_Churn/
├── data/
│   └── Telco-Customer-Churn.csv        # Raw dataset
│
├── notebooks/
│   └── telco_churn_analysis.ipynb      # EDA + modelling
│
├── dashboards/
│   ├── Telco_Churn_Dashboard.pbix      # Power BI dashboard
│   └── screenshots/
│       ├── page1.png
│       └── page2.png
│
├── requirements.txt
├── LICENSE
└── README.md

---

Quick Start

# 1. Clone
git clone https://github.com/Harshitharam25/Telco_Customer_Churn.git
cd Telco_Customer_Churn

# 2. (Optional) virtual environment
python -m venv venv
source venv/bin/activate        # Windows: venv\Scripts\activate

# 3. Install dependencies
pip install -r requirements.txt

# 4. Run the notebook
jupyter notebook notebooks/telco_churn_analysis.ipynb

The dataset is already included in data/. The original source is linked above.

---

Dashboard

The Power BI dashboard (dashboards/Telco_Churn_Dashboard.pbix) covers churn KPIs and driver analysis, with filters for segment-level exploration.

---

Business Recommendations

Finding	Recommendation
Month-to-month contracts churn most	Incentivise 1–2 year contracts
Fiber-optic customers churn heavily	Review fiber service quality and pricing
Electronic-check payers churn more	Promote auto-pay / card options
Short-tenure customers churn most	Strengthen early-onboarding experience
Higher monthly charges track with churn	Revisit pricing tiers

---

Limitations & Future Work

This is a baseline project, and the following are known limitations:

• Single model (logistic regression) — no ensemble comparison yet
• No hyperparameter tuning — default settings
• Single train–test split — no cross-validation
• Class imbalance (26/74) not explicitly handled, which lowers recall on the churn class

Planned improvements:

• Compare tree-based models (Random Forest, XGBoost)
• Add cross-validation and hyperparameter tuning (GridSearchCV)
• Address class imbalance (class weights / SMOTE) to raise churn recall
• Add SHAP for feature-importance interpretation
• Add ROC-AUC to the evaluation

---

Tools & Libraries

pandas · NumPy · scikit-learn · seaborn · matplotlib · Jupyter · Power BI

---

License

MIT License — see LICENSE.

---

Author

Harshitha Ram

• GitHub: @Harshitharam25
• LinkedIn:@harshitharam25
• Email: harshitharam5623@gmail.com

---

AI Disclosure

AI tools assisted with documentation wording and formatting. The data analysis, modelling decisions, feature engineering, and insights are my own work.