E.ON: 70% Higher Sales Potential for Natural Gas

E.ON, supplying energy to more than 1.5 million customers and operating 66,000 kilometers of power lines and 4,000 kilometers of gas pipelines, was looking for ways to optimize its product offering, streamline customer acquisition, and expand gas supply to selected customers. The focus was on reaching those with a high likelihood of purchasing gas services.

Identifying the Right Customers to Target

To enable timely and effective customer engagement, we designed and automated a process running on Microsoft Azure. Our scope included:

1. Designing an analytical datamart containing all data necessary and relevant for building a predictive machine learning (ML) model.
2. Automating the transformation and loading of data from the data warehouse into the datamart.
3. Using the datamart as the foundation for ML model development on the Databricks platform, where we trained, evaluated, interpreted, and automated the model to:

• Identify customers at risk of switching to a competitor.
• Determine the key factors driving customer dissatisfaction, enabling E.ON to proactively target these customers with the most appropriate retention offers.

Advanced Analytics and Machine Learning on the Databricks Platform

In collaboration with E.ON, we designed and implemented a solution on the Databricks platform, running on Microsoft Azure. The key steps included:

1. Building an Analytical Datamart

• Developed a datamart containing all necessary data for predictive model development.
• Automated the process of transforming and loading data from the data warehouse (DWH) into the datamart.

2. Advanced Feature Engineering and Modeling

• On Databricks, we expanded 300 core attributes into over 7,500 advanced predictors.
• Developed a machine learning model to identify customers without a gas connection from E.ON who are likely to establish one or switch from another provider.
• Mapped the key factors influencing customer switching behavior, including the transfer of gas and electricity contracts between suppliers, the relevant regulatory framework, and E.ON’s internal processes.
• Trained the model on a dataset containing 1.2 million records and 600 attributes.

3. Automation and Integration

• Regularly exported model outputs (churn probability scores) back to the DWH for automated use in customer outreach.
• Fully automated the entire process with a monthly refresh cycle.

The ML Model Identifies 70% More Potential Customers Willing to Switch to E.ON Compared to the Original Expert Model

When measured by lift in the top decile—customers E.ON can prioritize through higher-cost acquisition channels—the ML model consistently achieves a lift of 3 to 3.5 compared to random targeting.

This represents approximately a 100% improvement over the first version of the model and a 70% improvement over targeting based solely on business or expert rules.

Key Results:

• 70% improvement in identifying potential customers compared to the original expert-based manual selection approach.
• 3–3.5× higher targeting efficiency in the top customer decile compared to random selection.