What the paper studied
Most hotel AI studies test a single application and report prediction accuracy rather than business value. This paper tests five connected modules inside a four-layer framework (data and sensing, AI analytics, decision and automation, value): booking cancellation prediction, occupancy forecasting, risk-based overbooking, guest segmentation and smart-room HVAC control. Chatbots and service robots are reviewed but not quantified.
Because the public Hotel Booking Demand dataset could not be loaded in the study's computing environment, the author generated a synthetic dataset with the same schema: a 215-room resort and a 200-room city hotel with arrivals from July 2015 to August 2017, 121,279 recorded bookings and an overall cancellation rate of 35.7% (close to the roughly 37% in the public data). Models were trained on arrivals before 2017 and tested on January to August 2017, mirroring how a deployed system would work. The code switches to the real dataset when it is supplied.
Key findings
- Cancellation prediction. Of five classifiers, histogram gradient boosting performed best (AUC 0.845, 95% CI 0.841–0.848), narrowly ahead of a neural network (0.844) and logistic regression (0.840). Lead time and deposit type were the strongest predictors, followed by market segment and special requests. For online travel agency bookings, cancellation probability rose from 18% at zero lead time to 66% at 300 days, while direct bookings stayed below 36%.
- Leak-free testing matters. These scores sit below the AUCs above 0.9 often reported in studies that use random splits and post-booking variables, which the author notes give models information about the future that a live system would not have.
- The best forecast depends on the horizon. Within a week of arrival, simple pickup methods were most accurate (MAPE 1.6% for the resort, 4.0% for the city hotel). From 30 days out, gradient boosting led: at 90 days its MAPE was 4.4% against 7.4% for additive pickup at the resort, and 7.6% against 9.3% at the city hotel. Holt-Winters, which ignores reservations on the books, deteriorated quickly with horizon.
- Risk-based overbooking. Turning cancellation probabilities into an overbooking policy raised net room revenue by 0.55% (resort) and 0.32% (city) over the historical 5% rule, capturing 57–59% of the gain available with perfect information. A naive limit based only on cancellation rates lost 20.4% of resort revenue to walked guests.
- Four guest segments. K-means clustering produced leisure couples (51%), corporate travelers (26%, booking 14 days ahead on average), family holidays (17%, booking 141 days ahead and staying 6.9 nights) and loyal repeat guests (6%). The segments overlap considerably, which is typical of behavioral data.
- Smart-room energy. Linking room controls to the PMS to set back unsold rooms cut HVAC energy by 18.1%. Reactive IoT occupancy control saved the most (23.4%) but left rooms outside the comfort band in 0.50% of occupied hours. AI-predictive pre-conditioning saved 21.2% while holding discomfort to 0.14%.
Why it matters for hospitality
The paper's core message is that a small accuracy gain can be worth a lot or very little, depending on the decision it drives. A model only half an AUC point better than logistic regression, once plugged into overbooking, captured more than half the value of perfect information. It also shows that "AI beats traditional methods" is too simple: pickup forecasting stayed best close to arrival, so a hybrid that switches by horizon outperforms any single method.
Practical takeaways
- Judge AI tools by the revenue, cost or energy outcomes of the decisions they drive, not by accuracy scores alone.
- Ask vendors whether reported accuracy comes from time-based testing; random splits inflate results.
- Use horizon-based hybrids: pickup close in, machine learning further out.
- Consider channel-specific deposit and reconfirmation policies for long-lead OTA bookings, where cancellation risk climbs steeply.
- In smart rooms, start with PMS-linked setbacks for unsold rooms, then weigh predictive pre-conditioning to protect guest comfort.
Limitations
All results come from simulated data, so the absolute figures (AUCs, error rates, revenue gains and energy savings) depend on the generator's assumptions and need validation with operational data. The energy model uses a single-zone thermal model with synthetic weather, the overbooking replay ignores upgrades, multi-room bookings and competitive pricing, and segmentation uses individual bookings rather than guest histories.