Forecasting Hotel Daily Occupancy for High-Frequency and Complex Seasonality Data

Abstract

Accurate hotel daily occupancy forecasting is an important input for hotel revenue management. This research presents forecasting models both time series and causal methods for a case study 4-star hotel in Phuket, Thailand. Holt-Winters, Box-Jenkins, Box-Cox transformation, ARMA errors, trend and multiple seasonal patterns (BATS), Trigonometric BATS (TBATS), back-propagation Neural Network (BPNN), and Support Vector Regression (SVR) are explored in this work, where Same Day Last Year is considered as benchmark forecast. For causal method, independent variables used as regressor inputs are transformed data observed in the past periods, the number of tourist arrivals from main countries to Phuket, Oil prices, exchange rate, etc. Accuracy for Aggregated data are tested using Mean Absolute Percentage Error (MAPE) and Mean Absolute Error (MAE). Findings suggested that back-propagation Neural Network (BPNN) outperforms other models with the lowest MAPE of 8.96%. However, SVR which gives second least error of 12.30% are not statistically different compared to BPNN, while all others are significantly different. It shows that Machine Learning techniques studied in this research outperform the sophisticated time series methods designed for complex seasonality data like BATS and TBATS. The results obtained can be applied to the case study hotel’s future planning about the forecasted number of left-over rooms so that they effectively allocate to their discounted online travel agent (OTA) more effectively.