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

Every hotel room faces a unique kind of emptiness. Unlike a factory that can store unsold goods for tomorrow, a hotel room is a perishable product; if a guest does not arrive, that night's revenue vanishes forever. To protect against this loss, hotels often accept more reservations than they have rooms, betting that some guests will cancel or simply not show up. This practice, known as overbooking, is a delicate balancing act. If a hotel guesses too conservatively, it leaves money on the table with empty beds. If it guesses too aggressively, it risks turning away guests who hold valid reservations, damaging their trust and the hotel's reputation. The core challenge for hotel managers is knowing exactly which specific reservations are likely to disappear so they can adjust their numbers with precision, rather than relying on broad guesses about the entire market.

For years, researchers have tried to solve this puzzle using computer programs that learn from past data. However, many of these attempts have been flawed because they tested their ideas in ways that did not reflect real life. Some studies mixed up past and future data, allowing the computer to use information that would not be available at the time of prediction. Others focused only on getting the highest possible score on a test, ignoring whether the computer's reasoning made sense to a human manager. A manager cannot trust a system they cannot understand, especially if the system makes a mistake that costs the hotel money. The goal, therefore, is not just to build a smart computer, but to build one that is honest, transparent, and useful for the daily decisions a revenue manager must make.

In a new study, researchers set out to create such a system using data from two Portuguese hotels: one located in a city and the other in a resort area. They analyzed nearly 120,000 individual bookings to see if they could predict which ones would be cancelled before the guest arrived. To ensure their results were trustworthy, they designed a strict test that mimicked real-world conditions. They trained their computer models on data from the past and then asked them to predict the future, using only information that would be available at the moment a booking was made. They also carefully removed any data points that might have been recorded only after a guest had already cancelled, ensuring the computer was not simply using information that would not be available at the time of prediction.

The researchers compared three different types of computer models: a standard statistical method that has been used for decades, and two more advanced machine learning systems that are better at spotting complex patterns. The most successful model was a sophisticated algorithm known as XGBoost, which proved slightly better than the others at distinguishing between bookings that would be kept and those that would be cancelled. However, the improvement was modest, not a massive leap. The study found that while advanced computers can find subtle clues, a well-designed statistical model still captures most of the important information. The real value of the study was not in finding a new, magical algorithm, but in proving that a rigorous, honest test yields much lower and more realistic success rates than many previous studies claimed. When the researchers removed the data points that would not be available at the time of prediction, the success rate dropped from the high 90s seen in other papers to a more honest range in the high 70s to mid-80s.

Beyond simply predicting cancellations, the researchers wanted to understand why the computer made its predictions. They used a technique that breaks down each prediction to show which factors pushed the risk up or down. This revealed a crucial difference between what a hotel can change and what it can only price for. At the city hotel, the strongest signal came from the type of deposit a guest was required to pay. Bookings with non-refundable terms were far more likely to be cancelled in the data than those with flexible terms, a finding that suggests deposit policies are a powerful tool for managing risk. At the resort hotel, the most important factors were how far in advance the booking was made and the type of guest, such as whether they were part of a travel group.

The study also showed how to turn these predictions into actual business decisions. Instead of just saying a booking is "risky," the researchers translated the computer's confidence into a specific number: the expected number of cancellations for a given day. This allows a manager to calculate exactly how many extra rooms they can safely sell. They also established a clear rule for when to intervene. If a booking is flagged as high-risk, the manager can choose to contact the guest for a confirmation or ask for a deposit. The study provided a menu of options based on how much the hotel values a lost room versus the annoyance of contacting a guest who wasn't going to cancel anyway. This approach respects the manager's expertise, giving them the numbers they need to make a judgment call rather than forcing them to follow a rigid computer command.

One of the most important findings was that the computer's advice changes depending on the type of hotel. For the city property, the data suggested that the hotel could actively manage risk by adjusting its deposit rules for specific groups, such as large parties or bookings made through travel agents. For the resort, the risk was driven more by the nature of the guest and the timing of the booking, meaning the hotel's best response was to adjust its pricing and overbooking levels rather than trying to change the guest's behavior. The researchers also noted that the data had some quirks; for instance, the way non-refundable bookings were recorded in the system made them look almost certain to cancel, which likely exaggerated the importance of that factor. By identifying these quirks, the study prevented managers from drawing false conclusions.

Ultimately, this research offers a practical guide for hotel managers who want to use artificial intelligence without being misled by hype. It demonstrates that the most valuable tool is not the most complex algorithm, but a system that is tested honestly, explained clearly, and translated into concrete numbers. By separating what a hotel can control, like its cancellation policies, from what it cannot, like a guest's personal history, the study provides a roadmap for making smarter decisions. It shows that while computers can help predict the future, the final judgment on how to handle a risky booking still belongs to the human manager, armed with clear information and a clear understanding of the costs involved.