This study investigates the use of machine learning models for customer churn prediction in subscription businesses. The study investigates the influence of significant features such as tenure, monthly charges, and contract on churn prediction and how the two models classify churn and non-churn customers. The information contained considerable class imbalance with non-churn customers well outpacing churn customers, which proved difficult for accurate prediction. Despite this, Random Forest and Boost both exhibited strong classification performance with an accuracy rate of 79% and AUC of 0.83 and clearly showed that they were capable of identifying successfully the churn. The results show that Boost outperforms Random Forest with a slightly better recall and precision values, showing that it does better on the churn case detection without compromising with the balance offered with the non-churn customers. Analyzing feature importance, tenure, contract, and monthly charges appeared at the top rank among all predictors of churn, as suggested by previous literature based on customer interaction and monetary concerns. The study also found that customer support-related features, such as Tech Support, were critical to the churn prediction mechanism. The results suggest that companies are able to use machine learning models to identify which customers will churn and formulate targeted retention efforts. Additional studies would make these models more streamlined and analyze additional customer behavior data for even better churn prediction, eventually aiding customer retention programs in subscription companies.
Bello et al. (Thu,) studied this question.