Short-Term Prediction of Urban Traffic Variability: Stochastic Volatility Modeling Approach Theodore Tsekeris 1 and Antony Stathopoulos 2 Abstract: This paper addresses the problem of modeling and predicting urban traffic flow variability, which involves considerable implications for the deployment of dynamic transportation management systems. Traffic variability is described in terms of a volatility metric, i.e., the conditional variance of traffic flow level, as a latent stochastic low-order Markov process. A discrete-time parametric stochastic model, referred to as stochastic volatility SV model is employed to provide short-term adaptive forecasts of traffic speed variability by using real-time detector measurements of volumes and occupancies in an urban arterial. The predictive performance of the SV model is compared to that of the generalized autoregressive conditional heteroscedasticity GARCH model, which has been recently used for the traffic variability forecasting, with regard to different measurement locations, forms of data input, lengths of forecasting horizon and performance measures. The results indicate the potential of the SV model to produce out-of-sample forecasts of speed variability with significantly higher accuracy, in comparison to the GARCH model. DOI: 10.1061/ASCETE.1943-5436.0000112 CE Database subject headings: Traffic models; Forecasting; Traffic flow; Intelligent transportation systems; Urban areas; Stochastic processes. Author keywords: Traffic models; Forecasting; Traffic flow; Intelligent transportation systems. Introduction Traffic variability provides a measure of the fluctuations observed in the level mean of traffic flow over time. It reflects the behav- ior of the random unforeseen component of a traffic flow time series along a specific link or path of the transportation network. The capacity to provide short-term forecasts of urban road traffic variability involves considerable implications for the agents of transportation system. From the side of network operators, the fast and accurate prediction of traffic variability can support the identification of emergent anomalies unusual behavior in the traffic flow process, by constructing confidence regions out of which structural breaks and outliers can be captured see Statho- poulos and Tsekeris 2006. In turn, efficient operating policies can be selected to enhance system performance during congestion pe- riods at particular network locations. These policies encompass alternative guidance, information, and control actions, such as rerouting to paths with less “volatile” traffic conditions as well as dissemination of real-time information about the fluctuations of travel speed or time along specific paths in the immediate future, i.e., within the next few minutes Helbing et al. 2002. From the users’ side, such information can assist risk-sensitive travelers to suitably adapt their trip strategies, by reconsidering their departure, route or modal decisions according to their own perception of the distribution of possible outcomes and their gen- eralized travel cost. Also, it can support the design of better just- in-time supply chain strategies of firms in delivering goods in urban areas. From the side of information providers, the increase of variability in traffic conditions typically enhances the users’ willing-to-pay for “better” information, in terms of reducing un- certainty, and the value of information services Bonsall 2004. In addition, the increased traffic variability, which adversely affects the reliability of traffic forecasts, may induce information provid- ers to allocate their resources e.g., on surveillance equipment at specific locations and time periods in order to promote the quality of their services. Despite the growing importance of determining the variability of traffic conditions for passenger and freight transportation, par- ticularly within congested urban road networks, the development of formal approaches for estimating and predicting traffic vari- ability has received rather limited attention in the current litera- ture. Existing approaches mostly focus on measuring present or historical levels of traffic variability based on archived flow mea- surement data sets. Such approaches include the multivariate sta- tistical quality control method Turochy and Smith 2002 and ANOVA Weijermars and van Berkum 2004 for describing traffic variability in urban freeways and roadways, respectively. More- over, Tsekeris and Stathopoulos 2006a used a principal compo- nent analysis for the description and long-term prediction over weekly periods of the spatiotemporal variations of traffic in urban arterials. A few approaches have also been considered to deal with fu- ture changes in the variability of the traffic flow time series, such as nonparametric methods, including the Kernel estimators see Washington et al. 2003. In particular, Cho and Rilett 2003 used an artificial neural network ANN model with a bootstrap proce- 1 Research Fellow, Centre of Planning and Economic Research KEPE, 11 Amerikis, 106 72 Athens, Greece corresponding author. E-mail: tsek@kepe.gr 2 Professor, Dept. of Transportation Planning and Engineering, School of Civil Engineering, National Technical Univ. of Athens, 5 Iroon Poly- techniou, 157 73 Athens, Greece. E-mail: astath@transport.ntua.gr Note. This manuscript was submitted on September 2, 2008; approved on September 22, 2009; published online on October 1, 2009. Discussion period open until December 1, 2010; separate discussions must be sub- mitted for individual papers. This paper is part of the Journal of Trans- portation Engineering, Vol. 136, No. 7, July 1, 2010. ©ASCE, ISSN 0733-947X/2010/7-606–613/$25.00. 606 / JOURNAL OF TRANSPORTATION ENGINEERING © ASCE / JULY 2010 Downloaded 16 Jun 2010 to 147.102.154.11. Redistribution subject to ASCE license or copyright. Visit http://www.ascelibrary.org