ABSTRACT
Count-based bicycle demand models have traditionally focused on estimation rather than prediction and have been criticized for lacking a direct causal relationship between significant variables and the activity being modeled. Because they are not choice-based models, they are doubted for their ability to forecast well. The rise of machine learning techniques has given researchers tools to build better predictive models, and the tools to evaluate predictiveness. Extensive previous work in the statistics and machine learning field has shown that the best predictive model is not synonymous with the most true (or explanatory) model. The non-motorized demand modeling community could leverage these lessons learned to develop better count-based predictive models. The rise of the COVID-19 pandemic has clearly affected travel patterns, and the broad data collection has opened-up an opportunity to leverage machine learning techniques to build a predictive bicycle demand model. This study uses bicycle count data, COVID-19 data, and weather data to develop a LASSO regression model for three facilities in Austin, TX. The COVID-19 variables included both state- and local-level data between March 15, 2020, and January 31, 2021. The final model selects six variables out of 28 variables and reveals that the increase of statewide COVID-19 fatalities, statewide molecular positivity rate, and local precipitation cause a decrease in bike ridership, meanwhile maximum temperature causes an increase. The LASSO model also has a lower prediction MSE during cross validation compared to the full model. This paper aims to bring to light that our present-day demand and volume forecasting efforts would benefit tremendously from a predictive modeling approach rather than valuing the most explanatory models as the only strong forecasters of demand. In the end, modelers can use this approach to improve the forecasting ability of any count-based bicycle demand model. © 2023