Recurrent and decomposed neural network-based hotel occupancy forecasting

Occupancy forecasting is one of the most important decisions a hotel's management must make. If a more accurate form of forecasting is available, the hotel managers can increase their profits through the efficient management of rooms based on the forecasts. However, there are two major problems in applying the traditional time series model to hotel occupancy forecasting. One of the reasons is due to the upper and lower bounds which exist in the hotel occupancy data. The other is that there are many judgmental and exogenous factors which strongly affect the hotel occupancy. However, it is difficult to consider these factors in the forecasting model by using a time series model or causal model. Therefore, a recurrent and decomposed neural network-based hotel occupancy forecasting model has been proposed to resolve such problems. To validate our proposed neural network model, experiments have been performed using the real world data from a hotel. Based on the experiment's results, the proposed neural network model outperforms not only the ARIMA approach, but also other possible neural network models. This study also includes an interesting discussion about the issues that should be considered in applying a neural network to the forecasting domain through various comparative experiments between the time series model and the neural network, and among various types of neural networks.