Telco Churn Prediction with Advanced SQL & CatBoost

This project implements an end-to-end churn prediction pipeline. It simulates a telecom company's relational database to perform feature engineering using SQL Window Functions and trains a CatBoost classifier to identify high-risk customers.

The primary goal is to demonstrate the ability to handle data transformation at the database level (ELT) rather than relying solely on in-memory processing, followed by state-of-the-art machine learning modeling.

📋 Project Overview

1. Data Simulation: Generates a relational database (SQLite) with realistic patterns for customers, call logs, and complaints.
2. SQL Feature Engineering: Uses CTEs, Aggregations, and Window Functions (e.g., calculating trend changes over time) to extract features directly from the raw database tables.
3. Machine Learning: Trains a CatBoost Classifier to predict customer churn, optimizing for Recall to capture as many potential churners as possible.

🛠 Tech Stack

• Language: Python 3.x
• Database: SQLite
• Data Manipulation: SQL (Window Functions, Joins), Pandas
• Machine Learning: CatBoost, Scikit-learn
• Version Control: Git

📊 Key Results

The model achieved high performance in distinguishing between loyal and churning customers.

• ROC-AUC Score: 0.9367
• Recall (Churn Class): 0.78 (Correctly identified 78% of leaving customers)
• Precision (Churn Class): 0.70

Top Predictive Features:

1. calls_last_30_days (Derived via SQL: Significant drop in usage)
2. total_complaints (Customer dissatisfaction signal)
3. total_calls (General usage volume)

📂 Project Structure

├── data/                  # Stores generated database and model artifacts
├── src/
│   ├── db_generator.py    # Simulates customers, calls, and complaints data
│   ├── feature_store.py   # Extracts features using complex SQL queries
│   └── train_model.py     # Trains, evaluates, and saves the CatBoost model
├── requirements.txt       # Project dependencies
└── README.md              # Project documentation