Forecasting hotel demand for revenue management using machine learning regression methods

This paper compares the accuracy of a set of 22 methods for short-term hotel demand forecasting for lead times up to 14 days ahead. Machine learning models are compared with methods ranging from seasonal naive to exponential smoothing methods for double seasonality. The machine learning methods considered include a new approach based on arbitrating, in which several forecasting models are dynamically combined to obtain predictions. Arbitrating is a metalearning approach that combines the output of experts according to predictions of the loss that they will incur. Particularly, the dynamic ensemble method is used. The methods were compared using a real time series of daily demand for a four-star hotel located in the south of Europe. The forecasting performance of those methods was assessed using three alternative accuracy measures. Results from extensive empirical experiments led us to conclude that machine learning methods outperform traditional hotel demand forecasting methods. We found that the use of machine learning models can reduce the root mean squared error up to 54% for a 1-day forecast horizon, and up to 45% for a 14-days forecast horizon, when compared with traditional exponential smoothing methods.