Improving Hotel Demand Forecasting Accuracy by Identifying Seasonality-Adjusted Outliers

Forecasting accuracy determines the effectiveness of revenue optimization. Although researchers have evaluated demand forecasting models applied to various types of hotel, their results have differed considerably because the forecasting methods tested were based on varying hotel demand patterns. The objectives of this thesis are to improve forecasting accuracy and to provide general guidance on method selection. To achieve the objectives, the author proposes two stages. In the first stage, the author tests the performance of each of the selected forecasting models in 13 scenarios classified by location and scale. The selected forecasting methods are (1) Same Day Last Year (SD), (2) Moving Average (MA), (3) Double Exponential Smoothing (ES), (4) Triple Exponential Smoothing (TES), and (5) the Winters model (WT). The ANOVA model and the Tukey Pairwise Comparison procedure are used to compare the performance of selected models. The best model identified in Stage 1 is further improved by reducing model redundancy, and a two-sample t-test is conducted to evaluate the reduced model. In the second stage, this research proposes a new procedure, namely, Seasonality-Adjusted Monthly Range (SAMR) to detect and handle the outliers in hotel demand. The results from Stage 1 suggest that the best method identified in all scenarios was the Winters method. In particular, the Winters method significantly outperforms other methods in the scenarios with strong day-of-week demand patterns. In scenarios (i.e., urban, airport, and luxury hotels) that experience less monthly variations, the reduced Winters method, which only considers day-of-week seasonality, is identified as the best model based on the principle of parsimonious model selection. The results from Stage 2 suggest that the best model (i.e., the Winters model) was further improved by SAMR, a proposed outlier handling procedure. As a result, the forecasting accuracy of the Winters method was improved in most scenarios