CHURN PREDICTION FOR ONLINE SPORTS BETTING PLATFORMBack Case Studies
Client: Online Sports Betting
Length: 6 months (ongoing)
Goal: Utilize a vast amount of user data, including user’s past behaviour, in order to build a
prediction model that would assign high churn probabilities to real churners and low
churn probabilities to users who are there to stay.
A client approached us to propose advanced projects for user behaviour analysis and predictions. After suggesting several possible projects, we agreed to do churn prediction. Customer retention is one of the fundamental aspects of Customer Relationship Management (CRM), driven by the fact that retaining an existing consumer is much less expensive than acquiring a new consumer. Detection of attrition or customer churn is one of the standard CRM strategies. Customer churn prediction is the process of assigning a probability of future churning behaviour to each user by building a prediction model based on the available user information, such as past behaviour and demographics.
Our approach in this manner involves multiple stages. The Idea Generation phase (phase 0) was completed free of charge and there were 3 proposed advanced analytics projects. After a thorough discussion, we decided to proceed with the “Late Churn Prediction” project.
- Phase 1: A fixed-cost engagement where we got familiar with available data and created a data model outline.
- Phase 2: An ongoing, iterative process of implementation of the predictive model.
- Phase 3: Data Model integration into daily workflow (building a pipeline accessible through API).
- Phase 4: Analysis of effectiveness of different re-engagement campaigns.
During the first phase, we used historical data to analyze patterns and trends in user behaviour. Different statistics and visualizations were implemented in R. Before the start of the modeling phase, we proposed an accuracy that might be achieved for the given domain and behaviour patterns. During this phase, we extracted many features that were used as input to train several machine learning models. The accuracy of models was measured using precision and recall for churners (because of imbalanced labels in the dataset), and compared with the client’s baseline model. After overperforming the baseline model, we continued with the integration phase, which is ongoing.
Familiarization with data has provided the customer with valuable insights into their customers but it also yielded a model outline (machine learning algorithms that can be used to create a model, and list of modeling features). A predictive algorithm is being trained with historical data and optimized as we strive for our defined goal of prediction accuracy. Comparing to the client’s baseline mode, for the same recall values, our final model had a 5-10% higher precision. Currently working in “offline mode”, this model will be applied to live data for testing of prediction accuracy and eventually integrated with the live system and marketing department initiatives to re-engage those customers with a high risk of leaving / churning.