Optimization of a mathematical model for churn prediction and customer segmentation to improve cross and up sell policies for a B2B operator.

It is now widely accepted that firms should direct more effort into retaining existing customers than to attracting new ones, since the cost for getting a new customer is usually high. To achieve this, customers likely to defect need to be identified so that they can be approached with tailored incentives or other bespoke retention offers. Such strategies call for predictive models, capable of identifying customers with higher probabilities of defecting in the relatively near future. In addition, not all users have the same added value to the business. That's why it's just as interesting to set up customer segmentation strategies, to better understand customers' needs and provide them with offers that suit them. Egis, an international player in construction engineering and mobility services creates and operates, intelligent infrastructures and buildings capable of responding to the climate emergency and the major challenges of our time, enabling more balanced, sustainable and resilient land use planning. It has a tremendous amount of data collected and stored, in order to be able to efficiently deliver good services to its users. This data is full of potentialities that could be exploited to provide better services to end customers, or to create new services. However, they are confronted with the difficulty of exploiting this data which is very heterogeneous and distributed in different systems, and the lack of data science skills to take advantage of this big data heritage. It is in this context that the internship took place. The aim of the internship was to exploit the big data assets of Easytrip, which is an entity of the Egis group providing services to heavy goods vehicles, in order to achieve a double objective: to build a predictive algorithm allowing to predict one month in advance the users who are going to churn in order to allow the commercial forces to take the necessary actions, to segment the customers by using different variables in order to improve the policies of cross and up sell of the marketing department. Given the fact that it was a POC (Prof Of Concept) to validate the feasibility of the project from an IT point of view (availability and quality of the data) and from a business point of view (verification of the gain linked to the solution), the project was developed following the 3 main stages of a Data Science project: Scoping, data collection and analysis, mathematical modelling through machine learning and putting the project into a pilot environment. After testing different data analysis and processing techniques, and then different predictive models, gradient boosting produced the best performance on the churn prediction task. The unsupervised machine learning techniques did not produce the best performance. However, a better understanding of the business need allowed the definition of a data processing pipeline allowing the sales teams to derive different business rules from a custom built reporting. At the end of the POC phase, the cross sell project was launched in pilot phase, while for the churn project, we managed to build a predictive model that works from a theoretical point and we still need to solve some IT related issues before moving it to a pilot phase.