Pharmaceutical wholesaler
Training a decision model for churn prediction
Starting point
- 600 million data records of sales activities and purchase transactions
- 26 tables, ~3,700 attributes
- Churn definition:
- Manually marked (notification)
- Revenue decline > 40%"
Project result
- Decision model with multiple components and an overall model accuracy of 86.49%
Role in the project
- Analysis of the data base using statistical methods and domain expertise
- Text mining on sales reports
- Time series analysis (trends)
- Classification on customer master data
Customer benefits
- Identification of potential churn candidates
- Enabling the initiation of countermeasures
- Foundation for further analyses (more domain expertise)
- Insights into current data quality and the need for data management
Jens Schnettler | Member of Executive Board