Predicting Hotel Booking Cancellations with Explainable Machine Learning: Implications for Revenue Management

Abstract Booking cancellations are a routine and costly feature of hotel demand: 37.0 percent of the bookings in the dataset used here were cancelled before arrival. Hotels respond with overbooking and with deposit and cancellation policies, but both responses are usually set in aggregate, while the decisions a revenue manager faces concern individual bookings on specific arrival dates. This study uses the Hotel Booking Demand dataset of Antonio, de Almeida and Nunes (2019c), covering 119,390 bookings at a Portuguese resort hotel and a city hotel, to predict individual booking cancellations under a protocol designed for decisions rather than for headline accuracy. Post-outcome variables are excluded, models are evaluated out of time, and a recording audit tests three fields whose values accumulate over a booking's life. A logistic regression baseline is compared with random forest and XGBoost on discrimination, calibration, and classification at cost-based thresholds. The best-performing model is explained with SHAP values at global, local and segment level, separating drivers a manager can act on, such as deposit type, distribution channel terms and offers targeted by lead time, from drivers that can only be priced for, such as a guest's own cancellation history. The predicted probabilities are then translated into expected cancellations and net demand by arrival date, an uncertainty range around expected cancellations, and a threshold-based rule for retention intervention, with the costs of false positives and false negatives stated explicitly. The study's contribution is a transparent benchmark, a decision-oriented explanation, and a quantified translation on a widely used public dataset; it proposes no new algorithm. Out of time, XGBoost performed best at both properties. On the twenty predictors as recorded, ROC-AUC was 0.840 at the city hotel and 0.849 at the resort hotel (logistic regression: 0.817 and 0.833); in an audited specification that excludes three fields whose recording is entangled with the outcome, it was 0.794 and 0.784 (logistic regression: 0.778 and 0.761). Deposit type was the strongest recorded driver at the city property, subject to a data caveat set out in Section 4.3.