Three months of churn prediction modeling in Budapest

As part of the 3-year program of the Aegon Analytical Academy, participants complete two exchange projects of 3 months at another Aegon Unit, in many cases abroad. In Paula Genovés’ case, she went to Budapest to work on a churn prediction project, putting all her new skills and learnings into practice. Read on to learn about her experience building a model that could save the company more than €1,000,000.00 a year!

Can you tell us something about yourself?

My name is Paula Genovés and I grew up in Valencia (on the East Coast of Spain), where I also stayed to go to university. After finishing my bachelor’s degree in Business Administration I decided to do a master’s degree in Actuarial Science, because I wanted to do more with statistics and numbers (topics that I really like).

After graduating, I started my career as an underwriter for personal lines at AIG (a multinational insurance company) in 2012, and in 2013, I decided to start as a business analyst in Madrid at AEGON, a multinational life insurance, pensions and asset management company with offices in over 20 countries. I decided to switch, because I wanted to work in the actuarial field. From 2015-2017, I was a participant in the first International AEGON Analytical Academy. I currently work for the Analysis and Control team, where we study the evolution of the portfolio in terms of profitability. My focus is health, where I apply predictive modelling to obtain the optimal increase of premiums when customers renew their contracts.

What kind of project are you working on/have you worked on?

For the final part of the Academy, I went on a 3-month assignment to a AEGON-department in Budapest. The goal of this project was to build a model which predicts which customers are likely to churn. These predictions can then be used to target those customers who are likely to churn with retention programs.

To build this model, I worked closely with the Data Mining Department, because they have a lot of knowledge on customer behavior. We used many different types of information to feed the model. Naturally, we used information on the contracts and personal characteristics of the policyholder, but we also used contact information, payment history, other contracts, financial information etc. The model was built using a combination of a CHAID model and Logistic Regression.


How was the model received by the business?

The model turned out to work very well. For example for pension products, the group of 10% contracts with the highest probability of churning (according to the model), turned out to have a churn rate which was 6 times higher than the average contract. This result was well-received, because just a 5% reduction of the number of customers that churns, can already save over €1.000.000 a year.

Apart from building the model, I also provided insights on the characteristics of loyal customers. This was also well-received by the product and sales team, because it helped them to design the retention efforts.

What are the biggest challenges you have encountered in your project?

The biggest challenge was at the start of the project. We wanted to use many different types of information in the model, which meant that I had to combine many different types of data sources. Moreover, many variables were in Hungarian, which made it quite difficult for me to navigate through the information on my own. Luckily, my colleagues were very helpful.

A second challenge was to work in R. I did learn some basics during the Academy, but I usually work in SAS and Emblem, so I had to do a lot of googling. My colleagues from the Academy were also very helpful!

What advice would you want to give to your colleagues in the field?

In my case it helped me a lot to start the project just doing the plan: setting the goal, the phases and tasks, thinking about the timing etc. This helped me to get through the first couple of weeks and helped preventing side-tracking too much. Moreover, it helped to schedule meetings in advance