Passenger train delay (PTD) prediction is an effective solution not only for rail operators to plan and solve their operation problems, but also for passengers to better plan or alter their travel schedules. As an endogenous variable, train delay profile or accumulated delay is normally used as a key variable to model train delay. In addition, several exogenous variables, especially ridership or station activities and population seem to influence train delay. This study develops and compares the performance of PTD prediction models using data-driven linear regression and ensemble learning methods with and without exogenous variables; the two exogenous variables examined are ridership and population. Three Amtrak passenger trains with different delay profiles and host-railroad performances from 2008 to 2019 are used as case studies to build the prediction models. The results indicate that the PTD prediction models using ensemble learning methods outperform the ones using linear regression (LR). Moreover, the results also indicate that the two exogenous variables had little to no impact on the prediction models using ensemble learning methods.