A food delivery platform in the UK.
No business likes to lose customers and experience customer churn. The food delivery platform asked Faculty to build a data model that would help it identify which customers were most likely to churn, so that it could put in place measures, such as special offers, to try and stop them from leaving.
Separating loyal customers from customers likely to churn is a common ‘classification’ problem, suitable for a wide range of machine learning models. In this example we needed to be aware of ‘imbalanced classes’. With an annual churn rate of about 6%, there were more people in the ‘will not churn’ class than in the ‘will churn’ class, with a proportion of about 17:1.
We investigated several models to determine the customers who were likely to churn. We fed in data such as order frequency, the length of time a customer had been with the client, their average order value, and how recently they had placed an order. We chose a random forest model, which not only gave the best accuracy, but is also notable for computational reasons. A random forest is a combination of many decision trees, which means it is highly parallelisable and so was quick to train on the company’s data, which is many terabytes in size.
The random forest algorithm was able to identify 92% of customers who went on to churn, a four-fold improvement on the 22% previously achieved by the food delivery platform. If the business could retain these identified customers, it could prevent £5.9 million of revenue being lost.