Space-time modeling of traffic variables with adaptive LASSO

• Activity Number: 408
• Type: Contributed
• Date/Time: Tuesday, August 5, 2014 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Yiannis Kamarianakis*+
• Keywords: traffic forecasting ; real-time predictions ; threshold regressions ; adaptive LASSO

Abstract

This article presents a parametric time-series model for short-term traffic forecasting, which accounts for space-time dependencies and cross-correlations of traffic variables. In the adopted framework, a separate model is built for each measurement location in the network. Each model may contain some hundreds of potentially useful predictors, that contain information from other locations in the network; the influential ones are chosen via a 2-step, penalized estimation scheme, namely, adaptive LASSO. In the Athens data, the method achieves approximately 90\% 1-step forecast accuracy on 3-minute volumes, which declines slowly as the forecast-horizon increases; the observed forecasting performance for occupancies is not as satisfactory though. An advantage of the proposed approach is that it is readily able to be automated.