Customer churn prediction in hotel firms

Hotel firms should be able to maintain long-term and lasting relationships with their customers to continue their activities profitably by standing out from their competitors. Thus, an accurate estimation of which customers will maintain their relations with the hotel firm and which ones will leave, in other words predicting the customer churn, will help hotel firms to carry out faster and more effective retention activities. The main purpose of this study is to estimate the future customer churn for hotel customers by applying machine learning methods and to develop recommendations for retaining prospective churners. In this context, customer churn prediction has been applied by comparing Logistic Regression and Random Forest algorithms using data belonging to the repeat customers of a chain hotel firm with three five-star resort hotels operating in Belek region, Antalya. According to the results of the study, the Random Forest algorithm has shown the best performance predicting the prospective churners that are likely to leave in the next three years 80% correctly (AUC 0.80). Besides, the Random Forest algorithm found the traditional variables based on RFM (recency, frequency, monetary) to be more important (respectively: recency, inter-purchase time, total number of overnight stays in the last three years, length of relationship, change in average expenditure amount in the last three years, total frequency). On the other hand, the most important variables suggested by the Logistic Regression method were mostly hotel industry-oriented (respectively: accommodation at the other hotels of the hotel chain, total frequency, accommodation with children, accommodation as a couple, being a Turkish citizen). Finally, the probability of churn for the next period was calculated through the Random Forest model. A priority matrix was created by taking into consideration the churn probability and customer value. Suggestions for customer segments called "hesitant", "economic", "alternative seeker" and "opportunist" were shared.