Assessment of the transit ridership prediction errors using AVL/APC data

The disparity between actual and forecasted transit ridership has been an important area of study and a concern for researchers for several decades. In order to decrease the discrepancy caused by model property errors, a number of studies focus on better representation of difficult-to-measure cost functions and incorporation of behavioral variables in mode choice models. In spite of the improvement, some gaps still remain in practical applications, particularly for large-scale regional travel forecasting models which are zone-based and aggregated. With automated data collection systems including Automatic Vehicle Location/Automatic Passenger Count (AVL/APC), modellers have great potential to use these technologies as new or complementary data sources to reliably estimate system performances and observed transit ridership. In particular, an opportunity exists to explore model prediction errors at a more disaggregate spatial scale. In this paper, using AVL/APC data, a method to effectively identify and evaluate the source of transit ridership prediction errors is proposed. Multinomial regression models developed in this research produce equations for mode choice prediction errors as a function of: measurable but omitted market segmentation variables in current mode choice utility functions; and newly quantifiable attributes with new data sources or techniques including quality of service variables. Further, the proposed composite index can systematically evaluate and prioritize the major source of prediction errors by quantifying total magnitudes of prediction error and a possible error component. The outcomes of the research can serve as foundation towards more reliable and accurate mode choice models and ultimately enhanced transit travel forecasting.