Dear Tomke, congratulations on recently graduating from the Business Informatics programme and on producing an excellent bachelor’s thesis with clear practical relevance. What’s next for you? Will you be staying with your partner company, Deloitte?
Yes, I’ll be staying with Deloitte and will be joining the Performance Improvement & Restructuring team as a Consultant after graduation. I’m looking forward to building on what I learned during my dual-study programme while also moving into a new field.
How did you come to join HSBA?
I knew quite early on that I wanted to study in Hamburg and pursue a dual-study programme at the same time. Since friends of mine were already studying at HSBA and spoke very positively about their experience, the university quickly became an obvious choice for me. Once I had also found the right partner company, the combination simply felt like the right fit.
Why did you choose Business Informatics?
Before starting university, I wasn’t entirely sure whether I wanted to focus more on traditional business administration or move further towards IT. Business Informatics was the ideal combination for me because it brings both areas together. At the same time, it was important for me to develop a solid understanding of technology, as IT plays an increasingly central role in many business challenges. The programme allowed me to gain insight into both worlds and helped me work out where I might want to specialise in the future.
How did you come up with the topic for your bachelor’s thesis?
I particularly enjoyed the Data Science minor during my studies. It therefore became clear quite quickly that I wanted to focus on a technical topic for my bachelor’s thesis and apply what I had learned in practice. What I found especially appealing was working with real-world data, as it reflects actual challenges far better than prepared sample datasets.
Through my family’s hotel business, I had access to genuine booking data, which ultimately led to the topic of my thesis: “Evaluation of Machine Learning Models for the Early Detection of Hotel Booking Cancellations.”
What exactly was your research about?
At its core, the project addresses a very practical challenge for hotels: a reservation in the system does not necessarily mean that a room will ultimately be occupied. Cancellations therefore make it more difficult to estimate actual demand and plan room capacity effectively. Short-notice cancellations are particularly problematic, as hotels may not have enough time to resell the rooms.
I therefore investigated whether it is possible to predict, at the time of booking, whether a reservation is likely to be cancelled later on. To do this, I analysed the data in three consecutive stages. The first stage focused solely on booking information, such as booking lead time, length of stay, and booking channel. The second stage added aggregated guest information, while the third incorporated contextual factors such as public holidays, school holidays, and local events.
Because I identified irregularities in the earliest years of data, I also created two time-filtered versions of each stage. This resulted in a total of nine dataset variants, which were then analysed systematically.
Based on these datasets, I compared six different machine learning methods: Logistic Regression, Support Vector Machines, Decision Trees, Random Forest, XGBoost, and CatBoost. The best-performing manually selected model was then benchmarked against an AutoML approach using AutoGluon-Tabular.
The results showed that the available data does indeed contain meaningful patterns that can be used to predict cancellations. Tree-based methods delivered consistently strong results across the different dataset versions. CatBoost ultimately emerged as the strongest-performing model, which was also well suited to the predominantly categorical structure of the data.
The final CatBoost model was further refined through hyperparameter optimisation and, in particular, through threshold tuning to prioritise reliable cancellation predictions. On the holdout test year, around 88 per cent of bookings predicted as cancellations were in fact cancelled. At the same time, however, the model identified only around 39 per cent of all actual cancellations.
The findings are also noteworthy in relation to previous research, as they demonstrate that useful risk signals can be derived even from a comparatively small dataset.
“This is an excellent piece of work. It combines a highly interesting use case with a practical benefit that is immediately apparent. At the same time, it demonstrates very clearly that success in machine learning depends not only on optimising models, but equally on preparing the underlying data with great care.”
Prof. Dr. Sönke Hartmann, Professor of Operations Research & Logistics
How would you assess the potential impact of your work for hotels?
The findings show clear potential for practical application. The model could serve as an additional risk indicator, helping hotels identify and monitor bookings with an elevated likelihood of cancellation and, where appropriate, contact guests at an earlier stage.
This could make expected occupancy levels more predictable and support capacity and room-planning decisions. Operational decisions should, of course, continue to be taken by hotel staff rather than being triggered automatically by the model.
The next challenge is moving from academic modelling into day-to-day operations and assessing how the model could be integrated into an existing IT landscape from a technical perspective.
Looking back on your studies, what will stay with you most?
Most of all, I’ll probably remember the people and the experience of learning together. Over the course of three years, you work towards a shared goal, support one another, and try to make the most of the time you have together. Along the way, I also made many new friends.
