Churn Prediction Model for PowerCo
Muskan
PowerCo Project
BCG's client is PowerCo - a major gas and electricity utility that supplies to small and medium-sized enterprises.
🚀 Project Overview
Business Context:
Customer churn is a significant challenge in subscription-based services like utilities. Retaining existing customers is often more cost-effective than acquiring new ones.
Goal:
To analyze client and pricing data, engineer meaningful features, and build a machine learning model to predict churned customers. This helps the business focus on retention strategies.
🔍 Project Workflow
1. Exploratory Data Analysis (EDA)
In this phase, we:
Analyzed the churn distribution: Found ~10% churn rate — relatively healthy.
Explored relationships between churn and features like:
Sales Channel
Contract Type
Number of Products
Number of Years
Consumption and Power Usage
Discovered that some sales channels had 0% churn, while others had varied churn levels.
Identified skewed distributions and outliers in numerical features using histograms and boxplots.
2. Feature Engineering
This step included:
Price Sensitivity Features:
avg_price_diff: Average price change across peak, mid-peak, and off-peak.max_price_diff: Maximum price difference across all months and periods.Transformations:
Converted
date into month-based numerical features.Converted booleans to binary.
Categorical features → One-hot encoding (dummies).
Skewness Treatment:
Applied logarithmic transformation to skewed features (e.g.,
consumption, forecast, etc.)Correlation Analysis:
Identified multicollinearity and highlighted features that might be redundant.
3. Predictive Modeling
Used Random Forest Classifier for prediction.
Performed Train-Test Split to avoid overfitting and simulate real-world predictions.
Model Evaluation based on:
Accuracy: Overall correct predictions.
Precision: Correct churn predictions among predicted churn.
Recall: Correct churn predictions among actual churn.
Findings:
The model predicts non-churners well (high accuracy), but struggles to detect actual churners (low recall).
Indicates a need for better feature engineering or balancing techniques (e.g., SMOTE).
📊 Key Insights
Net Margin and 12-month Consumption were among the most important features.Price sensitivity features did not contribute significantly to predicting churn in their current form.
Skewed distributions and class imbalance affected model performance.
🛠 Tech Stack & Tools
Python 3.10+
Jupyter Notebook
Pandas & NumPy – Data manipulation
Matplotlib & Seaborn – Visualization
Scikit-learn – Machine learning and evaluation
Log Transformation – Skewness treatment
OneHotEncoder / pd.get_dummies – Categorical encoding
Like this project
Posted May 13, 2025
Developed a churn prediction model for PowerCo using machine learning and data analysis.
Likes
0
Views
0
Clients
PowerCo
