Customer Churn in Subscription Business Model—Predictive Analytics on Customer Churn

Abstract

There is a growing tendency for more companies to develop towards a subscription business model. Under such a trend, it is important to learn about the customer churn rate within the business, learn from it and adjust business strategies accordingly. This paper aims to predict customer churn rate in subscription business models using a variety of machine learning algorithms. Through comparing the results from the different algorithms, the best algorithms can be identified so that it provides an insight on which algorithm a subscription business should choose in order to predict customer churn most effectively. In this work, a total of 21 features and 9 algorithms are taken into account. Through a set of rigorous procedure including data preparation, feature engineering, feature selection, model building, and finally, model evaluation, three algorithms, namely Logistic Regression, Gradient Boosting (SMOTE) and Neural Network outperformed other 6 algorithms. The best performing algorithm being Logistic Regression with its 79.6% prediction accuracy, thus the conclusion that when subscription business predicts customer churn rate, Logistic Regression is the most preferable algorithm. During the process of feature engineering, SMOTE did not improve the model performance as it supposed to, so it is not recommended during the model building process.