From if to when: Timing hotel booking cancellations to improve demand management

Booking cancellations distort hotel demand forecasts, pricing, and overbooking decisions. The dominant approach in the literature predicts whether a reservation will cancel. We move the question forward by asking when a cancellation is most likely to occur. Using reservation data from four Portuguese hotels, we fit survival-analysis models, with Random Survival Forests as our primary specification, to estimate, for each booking, a day-by-day cancellation risk profile over the booking horizon. We translate this profile into two operational outputs: a Predicted Cancellation Day (PCD), the single day on which cancellation is most likely, and a Predicted Cancellation Window (PCW), a compact interval around it. The best PCW setting captures 33–72% of cancellations within a median window of 5–7 days. The approach complements existing classifiers by indicating when interventions are most likely to matter, supporting retention, overbooking, and short-term staffing decisions.