A COMPARISON OF TIME SERIES METHODS FOR POST-COVID TRANSIT RIDERSHIP FORECASTING

Transit agencies conduct system level ridership forecasting for planning, budgeting, and administrative purposes. The COVID-19 pandemic introduced substantial changes to transit ridership levels and patterns, which has impacted the performance of ridership forecasting. Although time series methods are commonly used for forecasting transportation demand, they have received limited use in practice for ridership forecasting. This thesis compares the performance of seven time series methods for univariate forecasts of system-wide, monthly transit ridership for heavy rail agencies in the continental US. The forecasting methods are: ETS, ARIMA, STL-ETS, STL-ARIMA, TBATS, neural network, and a hybrid model. Ridership was forecasted for pre- and post-COVID periods (pre- and post- March 2020) as well as for the full series (January 2002 to December 2023). The MAPE and MASE were used to compare forecast performance. About half of the pre-COVID forecasts produced a MAPE below 5% and about 90% produced a MAPE below 10%. Most of the full-series and post-COVID models underperformed by comparison, with only about 10% producing a MAPE below 5% and under half producing a MAPE below 10%. The hybrid method performed relatively well for each time period. The neural network underperformed for the pre-COVID and full series periods, but performed relatively well for the post-COVID period. Based on these findings, time series forecasting is an efficient method to forecast ridership with stable seasonality, periodicity, and trends, which was typical of pre-pandemic ridership. However, the results suggest that these stable ridership patterns have not yet returned at many US heavy rail agencies.

The R functions used in this study have been made publicly available on Github at https://github.com/ashley2876/forecasting_repo