AAUBS/hotel-cancellation-model

Hotel cancellation model (teaching)

Predicts the probability that a hotel booking will be cancelled, from what is known when the booking is made. Built in the Business Data Science programme, Aalborg University, session 10.

• Data: Antonio, Almeida & Nunes (2019), Hotel booking demand datasets, Data in Brief 22. Two Portuguese hotels, 2015–2017.
• Model: scikit-learn pipeline (imputation, scaling, one-hot) + XGBoost, tuned with Optuna on validation log loss.
• Split: by arrival date. Train to Nov 2016, validation Dec 2016–Mar 2017, test Apr–Aug 2017.
• Test performance: AUC 0.8102, log loss 0.5004, Brier 0.1694.
• Removed as leakage: required_car_parking_spaces, total_of_special_requests, booking_changes, days_in_waiting_list, room_changed, is_portugal (recorded or updated after the booking).
• Limits: two hotels, one country, 2015–2017; cancellations rose in 2017 and the model under-forecasts them slightly. Probabilities describe patterns in this data, not causes. Not for real pricing or staffing decisions.

Load with joblib.load("model.joblib") using the versions in config.json.