Forecasting Hotel Demand Using Stacked Generalization

Demand forecasting is one of the most critical functions of hotel revenue management. It serves as the bedrock for strategic and operational decision-making across all facets of a hospitality organization, including pricing, inventory management, staffing, risk management, and purchasing. In fact, the success of a hotel is largely predicated on its ability to accurately and consistently predict and adapt to future trends. Now, with the proliferation of artificial intelligence and advances in information technology, demand forecasting in hotels stands to benefit significantly. Yet, the literature has only recently started to explore these solutions. Accordingly, there is a lack of research that rigorously develops and applies cutting-edge machine learning methodologies. In this dissertation, I fill this gap by rigorously validating and evaluating the accuracy and interpretability of machine learning models as well as introducing the concept of meta-learning in the field hotel revenue management using hotel occupancy data from a large casino hotel. The models were evaluated over various forecasting horizons and accuracy metrics, resulting in a series of experiments that showed that model performance varied across forecasting horizon and between model types. In doing so, I found that the suggested meta-model had the highest performance for further-out time horizons, but that the traditional models yielded better performances at the shorter time horizons. These results suggest that forecasting is not a one-size-fits-all exercise, but rather a search for the best forecasting method for the particular market, forecasting horizon, or dataset. This phenomenon is particularly important for both academics and practitioners, as it implies that models must be built with methodological rigor and contextual awareness in order to achieve the most accurate results.