Churn Prediction Modelling in B2b E-commerce: a Case Study of Twiga Foods

Abstract

The informal structure of African retail, which is mostly fueled by layers of brokers and middlemen, is fundamentally changing as a result of technology. Selling locally grown fresh produce to Kenyans should be simple, but this wasn't the case at the time due to the disjointed distribution networks and ineffective supply chains. Twiga foods started in 2014 has managed to build efficiency though technology. The digital plat-form helps to on- board farmers and aggregate demand of the produce by the retailers.The farmers then supply to Twiga Foods who then deliver to retailers for free. Evidently, acquiring a new customer is costlier than retaining a customer and therefore, one of the strategic and cost optimization initiatives by such digital platforms also known as e-commerce platforms is to retain customers. To retain the customers, such businesses must build mechanisms that allow them to predict the probability of a customer churning immediately a customer is acquired so that they are able to undertake business measures that would prevent a customer from churning. In this research, we build a machine learning model (Logistic Regression (LR) model) that predicts if a retailer will be retained or not, test the model, put it in production and identify different use cases the model could be applied.Logistic regression is a binary classification model; the goal of the model is to predict yes or no in Twiga’s case churn or retained. It utilizes a sigmoid function to assign a probability of between 0 and 1 to each retailer. We then evaluated the model performance by comparing two scenarios. Scenario 1; we split the data randomly. Scenario 2; we arrange the data in ascending order and split the data using delivery date. Scenario 2 had a higher percentage accuracy of 81% compared to Scenario1 at 52% accu- racy, Scenario 2 had 76% precision/specificity compared to scenario 1 at 71% and Scenario 2 had a 75% Recall (sensitivity) compared to scenario 1 at 62%. Scenario 2 is an improve- ment from the conventional models and has a higher performance.