Will this booking cancel?
A decision aid for the hotel revenue manager: should we resell this room? Enter a booking. The tool gives the probability that it cancels, a recommendation based on the real cost of each mistake, and the three reasons that weigh most.
Try an example:
Booking
Assessment
Loading the model…
Probability the booking is cancelled
Why — the top 3 reasons for this booking
Money at stake
Booking worth
Walking this guest
Expected value of reselling
Decision support only. The revenue manager makes the final call. Never used to treat guests differently at arrival.
About the model — and how far to trust it
- Data: 119,390 real bookings of a city hotel (Lisbon) and a resort hotel (Algarve), 2015–2017. Antonio, de Almeida & Nunes (2019), Data in Brief, CC BY 4.0.
- Model: tuned LightGBM. It only uses information known when the booking is made. Country, parking requests, special requests, room assigned and booking changes were removed because they are filled in later (leakage).
- Honest test results on 23,030 bookings never seen in training (grouped so that rooms of one group booking are never split): ROC-AUC 0.85, accuracy 79% (doing nothing: 64%). At the threshold of 0.58, 91% of flagged bookings really cancel.
- Threshold 0.58 maximises net euros. A correctly resold room recovers the booking's revenue. A wrongly resold one costs a walked guest: a comparable hotel for the stay plus €100.
- Limits: weaker on online-agency and short-notice bookings. Accuracy drops to an AUC of 0.81 when predicting a later period, so retrain monthly. Families are wrongly flagged more often, hence the manual review.
- Reasons are exact SHAP values of the model. They describe what drives the prediction, not what causes cancellations.
- Runs entirely in your browser. No data leaves this page.