CREDIT MANAGEMENT
BUSINESS QUESTION
A CAC40 industrial company, with more than 100 000 customers is facing overdue. In the past years, a tier1 collection process was implemented with more than 30 FTE calling customers to secure on-time payments. This lead to -5% overdue but the head of credit management was willing to leverage available data to go one step further.
DATA
- 3 years worth SAP data, extracted from several tables in the order-to-cash process and CRM systems
- More than 1 million invoices to analyze, with various complexities in clearing mechanisms
- Data were collected in .csv format and uploaded to a PostgreSQL database
APPROACH
- 18 interviews with referees all across the globe, to understand the data structure and collection processes
- Pages of SQL queries to filter the invoices scope and prepare features (intragroup and the most complex deals where excluded)
- Ensemble methods to automatically learn customers habits in a specific context (customer sector, size, payment history, FX rates, holidays, invoicing pattern, etc.)
- Several presentations to top executives to explain data science principles and deep dive in the methodology
RESULTS
- The model was able to predict overdue invoices at the posting date, with a good performance.
- 20% productivity gains where demonstrated on the POC perimeter, when compared to the current methodology to prioritize collection efforts
- A project is planned to industrialize the model and benefit from daily predictive alerts / ranking