Modeling transportation time series using bayesian dynamic linear models

Sudden large-scale changes like the recent COVID-19 pandemic make the management and planning of transport systems difficult, despite the ever-increasing availability of data. The primary goal of this work is to model transportation data time series using a dynamic model that is interpretable and can be used for long-term forecasting. The Bayesian Dynamic Linear Model (BDLM) is chosen because it can account for complex data and can be easily adapted to the data. A component for the BDLM is introduced to recognize the underlying patterns using temporal control points. A moving-event component is also added to take into account events that do not occur on the same date every year such as sports games. The proposed model is parsimonious and can learn from the data. After providing a brief summary of the theory of the model, experimental results are shown for transport demand data obtained from smart card transaction data for the Montreal subway system. The proposed model is compared to different time series models and shows superior accuracy.