An interpretable machine learning method for hotel demand forecasting

This paper proposes a new interpretable two-step approach to forecast daily hotel demand based on advanced booking data. First, this approach applies a Principal Components Analysis (PCA) to booking curves of a hotel. Second, a pickup forecasting model is used to forecast daily hotel occupancy. The forecasting performance of the proposed approach was evaluated using real booking data of three European hotels (2018-2022) and using the additive pickup and a clustering-based method as benchmarks. The proposed approach was also evaluated using additional information (i.e., average daily rate) included in orthogonal components: the PCA(ADR) method. Empirical results show that the PCA methods perform better than the benchmarks for all hotels and all forecasting horizons (7, 14, 21 and 45 days). Furthermore, the results indicate that the PCA(ADR) method generates forecasts that are slightly more accurate than the basic PCA, which enhances forecasts with additional business information in a low-dimensional space.