PowerCo Customer Churn Prediction

Real-world ML analysis predicting customer churn for a European energy provider using 50+ engineered features and XGBoost

Table of Contents

• Overview
• Key Findings
• Methodology
• Model Performance
• Technical Stack
• Getting Started

---

Overview

Advanced data science analysis for PowerCo using real-world dataset of 10,000+ customers with 12-month transaction history. Engineered 50+ predictive features across price dynamics, profitability, tenure, and contract signals. Trained and compared three ML algorithms (Random Forest, XGBoost, Gradient Boosting) to identify at-risk customers.

Key Finding: Customer profitability (5.1%) & segmentation (9.0%) are primary churn drivers—not price sensitivity (2.4%)

🎯 Business Problem

PowerCo initially hypothesized price sensitivity as the primary churn driver. This project validates that hypothesis using data:

• Contract terms (tenure, renewal timing)
• Customer profitability (margin per customer)
• Pricing patterns (volatility, rate changes)
• Geographic segmentation (origin, channel)

Result: Price hypothesis partially rejected in favor of profitability-driven insights.

Model Performance

Best Model: XGBoost

• Test ROC-AUC: 0.677 | Precision: 95.7% | Recall: 6.0%
• Feature Importance Top 3:
  1. Net margin on electricity (5.1%)
  2. Customer geographic origin (9.0%)
  3. Price volatility (2.4%)

---

Methodology

Feature Engineering (50+ Features)

• Price Dynamics (14): Monthly deltas, volatility, trend slopes
• Tenure Signals (5): Months active, contract end date, modification history
• Consumption Patterns (6): Log-space ratios, forecast deviations, usage trends
• Financial Metrics (5): Profitability scores, customer lifetime value, cost per kWh
• Contract Signals (6): Stability scores, renewal indicators, change flags

Technical Note: All skewed variables transformed via log-space arithmetic to prevent extreme outliers.

Model Comparison

Model	CV ROC-AUC	Test ROC-AUC	Precision	Recall
Random Forest	0.7006	0.6678	90.91%	2.73%
XGBoost	0.7133	0.6773 ✅	95.65%	6.01%
Gradient Boosting	0.7081	0.6760	100%	1.91%

---

🔍 Key Findings

1. Profitability > Price: Low-margin customers churn at higher rates
2. Segmentation Matters: Geographic origin + sales channel = 9% importance
3. Price Not Primary: Price volatility ranks #3, contradicting initial hypothesis

---

Technical Stack

Languages & Libraries:

• Python 3.12.3 | Pandas 3.0.1 | NumPy 2.4.3
• Scikit-learn 1.8.0 | XGBoost | GridSearchCV
• Matplotlib 3.10.8 | Seaborn 0.13.2

Data Pipeline: Raw Data → Cleaning → Feature Engineering (50+) → Log Transformation → Train/Test Split → GridSearchCV Tuning → Model Evaluation

---

Getting Started

Clone & Setup

git clone https://github.com/kpatc/PowerCo-Churn-Prediction-BCG.git
cd "PowerCo Churn Prediction "
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Run Analysis

# EDA
jupyter notebook eda/eda_analysis.ipynb

# Features
jupyter notebook feature_engineering/feature_engineering_v2.ipynb

# Models
jupyter notebook modeling/modeling.ipynb

---

Key Learnings

• Log-space arithmetic: Prevented extreme outliers in skewed data
• Class imbalance: 90% no-churn required ROC-AUC focus over accuracy
• Feature importance: Profitability > Segmentation >> Price

---