Forecasting inbound tourism in light of data structure: A case study of Australia using deep learning

Abstract

Tourism is an important socioeconomic sector for many countries worldwide. The perishable nature of this industry requires highly accurate forecasts to support decision-makers with their strategies and planning. This study explores the relationship between time series data characteristics and the forecasting performance of the cutting edge Long Short-Term Memory (LSTM) neural network, along with benchmark methods. Such analyses are important to provide practical recommendations based on empirical evidence to support the development of more accurate forecasts. We analyze the case of inbound tourism in Australia from several country sources, including developed and developing economies from five continents. Findings from this study reveal that the LSTM deep learning approach achieves superior performance in most cases. However, we find that data characteristics, mainly unit root and structural breaks, are related to poor performance of LSTM forecasting model and, in such cases, the deep learning method is not recommended. The results reveal insights that can lead to a forecasting error reduction of around 40% in some cases. Further, more accurate results are found using univariate time series compared to models that employ regressor variables.