Air transportation direct share is the ratio of direct passengers to total passengers on a directional origin and destination (O&D) pair. Direct share is an essential factor of passenger flow distribution and shows passengers’ general preference for direct flight services on a certain O&D. A better understanding and a more accurate forecast of direct share can benefit air transportation planners, airlines, and airports in multiple ways. In most of the previous research and applications, it is commonly assumed that direct share is a fixed ratio, which contradicts the air transportation practice. In the Federal Aviation Administration (FAA) Terminal Area Forecast (TAF), the O&D direct share is forecasted as a constant based on the latest observation of direct share on the O&D. To find factors which have significant impacts on O&D direct share and to build an accurate model for O&D direct share forecasting, both parametric and nonparametric machine learning models are investigated in this research. We propose a novel category-based learning method which can provide better forecasting performance compared to employing the single modeling method for O&D direct share forecasting. Based on the comparison, the developed category-based learning model is a promising replacement for the model used for O&D direct share forecasting by the FAA TAF.

Air transportation direct share is the ratio of direct passengers to total passengers on a directional origin and destination (O&D) pair during a period of time. The O&D direct share shows the air passengers' general preference for direct flight services under a certain market state. The O&D direct share time series shows the changing process of a direct flight market. A better understanding and more accurate forecasting of direct share can benefit air traffic planners, airlines, and airports in multiple ways. In previous studies and practice, the direct share is usually assumed as a fixed average, which is not true for air transportation practice. The characteristics of O&D direct share time series are investigated in this research. Time series models and machine learning models are explored for the direct share time series forecasting. 1295 busiest O&D pairs across the U.S. are considered in this research. A framework for direct share time series modeling and forecasting is proposed in this research, which is a promising replacement to the model used for the direct share forecasting by the Federal Aviation Administration Terminal Area Forecast.

No abstract