Predicting Fitness Centre Dropout

The phenomenon of dropout is often found among customers of sports services. In this study we intend to evaluate the performance of machine learning algorithms in predicting dropout using available data about their historic use of facilities. The data relating to a sample of 5209 members was taken from a Portuguese fitness centre and included the variables registration data, payments and frequency, age, sex, non-attendance days, amount billed, average weekly visits, total number of visits, visits hired per week, number of registration renewals, number of members referrals, total monthly registrations, and total member enrolment time, which may be indicative of members' commitment. Whilst the Gradient Boosting Classifier had the best performance in predicting dropout (sensitivity = 0.986), the Random Forest Classifier was the best at predicting non-dropout (specificity = 0.790); the overall performance of the Gradient Boosting Classifier was superior to the Random Forest Classifier (accuracy 0.955 against 0.920). The most relevant variables predicting dropout were "non-attendance days", "total length of stay", and "total amount billed". The use of decision trees provides information that can be readily acted upon to identify member profiles of those at risk of dropout, giving also guidelines for measures and policies to reduce it.