Telco Customer Churn Prediction

Predict which telecom customers are likely to churn using SQL-based exploratory analysis and machine learning — enabling targeted retention campaigns before customers leave.

---

Key Highlights

• 26.5% overall churn rate identified across 7,043 customers
• Customers in their first 6 months churn at 52.94% — over 3× higher than long-tenure customers
• Fiber optic + month-to-month contract customers are the highest-risk segment (41.89% churn rate)
• Senior citizens churn at 1.8× the rate of non-senior customers
• LR + SMOTE selected as best model: Recall 0.797, F1 0.637, ROC-AUC 0.845

---

Project Structure

telco-churn-prediction/
├── data/
│   ├── raw/
│   │   └── Telco_Customer_Churn.csv
│   ├── processed/
│   │   ├── telco_feature_engineered.csv
│   │   └── customer_ids.csv
│   └── result/
│       └── top_100_customer_highest_churn_rate.csv
│
├── docs/
│   └── DATA_DICTIONARY.md
│
├── notebooks/
│   ├── data_preparation_and_loading.ipynb
│   ├── preprocessing.ipynb
│   └── scale_modeling.ipynb
│
├── sql/
│   ├── 01_who_is_churning.sql
│   ├── 02_how_services_affect_churn.sql
│   └── 03_how_customer_behavior_affects_churn.sql
│
└── README.md

---

Pipeline Overview

1. Data Preparation (data_preparation_and_loading.ipynb)

• Load raw CSV, inspect data types, missing values, and target distribution
• Convert TotalCharges from object to float; encode SeniorCitizen as Yes/No
• Push cleaned data to MySQL for SQL-based analysis

2. SQL Exploratory Analysis (sql/)

01 — Who is churning? Segment churn rates by gender, senior citizen status, partner, and dependents. Key finding: senior citizens and single customers are significantly higher-risk groups.

SeniorCitizen	Total Customers	Churned	Churn Rate (%)
Yes	1,142	476	41.68
No	5,901	1,393	23.61

02 — How do services affect churn? Analyze churn across phone service, internet type, add-on services (security, backup, tech support), and streaming. Key finding: fiber optic users without protective add-ons show the highest churn, and early-tenure customers are most at-risk.

Tenure Group	Total Customers	Churn Rate (%)
0–6 months	1,481	52.94
7–12 months	705	35.89
13–24 months	1,024	28.71
24+ months	3,833	14.04

03 — How does customer behavior affect churn? Examine contract type, payment method, and monthly/total charges. Key finding: month-to-month customers have both the highest churn rate and the highest monthly costs.

Contract Type	Avg Monthly Charges ($)	Churn Rate (%)
Month-to-month	66.40	42.71
One year	65.05	11.27
Two year	60.77	2.83

3. Preprocessing & EDA (preprocessing.ipynb)

• Handle missing values: TotalCharges nulls filled with 0 for new customers (tenure = 0)
• Save customerID separately before dropping, for use in final output
• Visualize class imbalance (73% No / 27% Yes → handled with SMOTE)
• Outlier detection via IQR on numerical features
• Churn rate bar charts across all categorical features

Feature Engineering

Feature	Description
tenure_group	Tenure bucketed into 4 lifecycle stages (0–6, 7–12, 13–24, 24+ months)
TotalServices	Count of active services per customer
AvgChargesPerService	Monthly charges normalized by service count
ChargesToTenureRatio	Total charges relative to tenure
IsSeniorSingle	Flag: senior citizen with no partner (high-risk segment)
HighRiskPayment	Flag: month-to-month contract + electronic check payment
HasFamily	Flag: customer has partner or dependents

4. Modeling (scale_modeling.ipynb)

• Stratified 80/20 train-test split to preserve class distribution
• Preprocessing pipeline: StandardScaler for numerical, OneHotEncoder for categorical
• Baseline models: Logistic Regression, Random Forest, Gradient Boosting
• Resampling: SMOTE applied inside pipeline on training data only
• Hyperparameter tuning: GridSearchCV with 5-fold stratified cross-validation on Logistic Regression
• Evaluation metrics: Accuracy, Precision, Recall, F1-Score, ROC-AUC

---

Model Performance

Model	Accuracy	Precision	Recall	F1-Score	ROC-AUC	Category
Logistic Regression	0.8070	0.6796	0.5160	0.5866	0.8463	Baseline
Random Forest	0.7913	0.6449	0.4759	0.5477	0.8167	Baseline
Gradient Boosting	0.7977	0.6519	0.5107	0.5727	0.8401	Baseline
LR + SMOTE (tuned)	0.7594	0.5312	0.7968	0.6374	0.8452	SMOTE
RF + SMOTE	0.7821	0.5954	0.5588	0.5766	0.8137	SMOTE
GB + SMOTE	0.7871	0.5869	0.6684	0.6250	0.8416	SMOTE

Selected model: LR + SMOTE (tuned) — highest Recall (0.797) and F1-Score (0.637), most effective at identifying customers likely to churn.

---

Results — Top 100 Highest Churn Risk Customers

Customers are scored by churn probability and ranked. The top 100 are exported to data/result/top_100_customer_highest_churn_rate.csv for use by the retention team.

CustomerID	Contract	Tenure	MonthlyCharges	ChurnProbability	Risk Segment
5178-LMXOP	Month-to-month	1	95.10	0.9861	High
7216-EWTRS	Month-to-month	1	100.80	0.9860	High
9497-QCMMS	Month-to-month	1	93.55	0.9853	High
9300-AGZNL	Month-to-month	1	94.00	0.9853	High
5419-JPRRN	Month-to-month	1	101.45	0.9832	High
...	...	...	...	...	...
1143-NMNQJ	Month-to-month	2	85.70	0.9488	High

All top 100 customers share the same pattern: month-to-month contract, tenure ≤ 3 months, and high monthly charges — consistent with SQL findings.

---

Key Business Findings

Finding	Churn Rate	Recommendation
Customers in first 6 months	52.94%	Onboarding support program in months 1–3
Fiber optic + no security/backup	~46%	Bundle protective add-ons with fiber plans
Month-to-month + electronic check	42.71%	Incentivize auto-pay and annual contract upgrades
Senior citizens (no partner)	41.68%	Simplified plans + dedicated support line
Customers without dependents	~2× average	Personalized retention offers for single users

---

How to Run

# 1. Clone the repository
git clone https://github.com/your-username/telco-churn-prediction.git
cd telco-churn-prediction

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

# 3. Run notebooks in order
jupyter notebook
# → notebooks/data_preparation_and_loading.ipynb  (loads data to MySQL)
# → notebooks/preprocessing.ipynb                 (EDA + feature engineering)
# → notebooks/scale_modeling.ipynb                (modeling + evaluation → outputs top_100 CSV)

Note: MySQL must be running locally. Update credentials in data_preparation_and_loading.ipynb before running.

---

Dataset

• Source: Kaggle — IBM Telco Customer Churn
• Size: 7,043 rows × 21 columns
• Target: Churn (Yes / No)
• Features: 3 numeric (tenure, MonthlyCharges, TotalCharges) + 18 categorical

---

Tech Stack

Category	Tools
Language	Python 3.x
Data processing	pandas, NumPy
Visualization	Matplotlib, Seaborn
ML & tuning	scikit-learn, imbalanced-learn
Database	MySQL, SQLAlchemy
Environment	Jupyter Notebook

---

Future Improvements

• Add SHAP values to explain individual predictions
• Deploy a Streamlit dashboard for interactive risk scoring
• Experiment with XGBoost / LightGBM for potential performance gains