An Integrated Machine Learning Pipeline for the Hospitality Industry: Customer Segmentation, Churn Prediction, and Hybrid Revenue Forecasting

This study presents an end-to-end machine learning pipeline that integrates customer segmentation, churn prediction, and revenue forecasting to support data-driven decision-making in hotel revenue management. The proposed framework was validated using approximately 1,060 days of operational data, encompassing 26,729 guest profiles from a 485 -room resort hotel in Turkey. K-Means clustering identified five distinct and interpretable guest segments, revealing that family travelers generate the highest average expenditure while high-frequency VIP guests constitute the most loyal cohort. A Random Forestbased churn prediction model achieved an ROC-AUC of 0.8702 and a weighted$F_{1}$-score of 0.9148 under stratified 5-fold crossvalidation, demonstrating robust discriminative power despite an 88.65 % class imbalance. Feature importance analysis confirmed that directly observable behavioral metrics-total stays, guest count, and average revenue per stay-serve as the strongest predictors of churn. For revenue forecasting, a comparative evaluation of SARIMA, XGBoost, LightGBM, and LSTM models revealed that LightGBM achieved the best performance with a MAPE of 6.14 %, while a performance-based hybrid ensemble maintained a MAPE of 7.78 % as a scalable production solution. The integrated pipeline was deployed as a FastAPIbased service, enabling real-time segment assignment and churn probability computation within the hotel's Property Management System. Scenario-based revenue forecasts further support proactive staffing, procurement, and dynamic pricing decisions, effectively transforming raw operational data into actionable business intelligence for the hospitality industry.