Unlocking air cargo demand: A novel predictive model selection framework based on data characteristics

Abstract

Air cargo transportation plays a key role in global trade, especially for time-sensitive and high-value products. Accurate prediction of air cargo demand is essential for informed decision-making on infrastructure planning, capacity management, and resource allocation across the air transport sector. While previous studies have largely advanced air cargo demand forecasting through model development and performance comparison, this study introduces a framework to explain and compare established predictive models based on data characteristic analysis (DCA). By examining key time series characteristics, such as stationarity, seasonality, and complexity, across statistical, machine learning, and deep learning approaches, this research examines how intrinsic properties of demand data are associated with forecasting performance. A rolling horizon design is also employed to evaluate how dynamic changes in data characteristics influence model performance over time. The findings reveal that statistical models are particularly sensitive to the mutability and complexity of air cargo demand data, whereas machine learning and deep learning models demonstrate stronger adaptability under diverse demand data conditions. Overall, this study shifts the emphasis from developing new forecasting models to explaining model suitability and supporting model selection, offering both theoretical insights and practical guidance for stakeholders by highlighting which models are best suited under specific data conditions.