Dynamic factor model for network traffic state forecast

We propose a dynamic factor model to forecast traffic state for groups of locations. The model decomposes the grouped traffic time series into the latent common factor component and idiosyncratic component. It uses a few latent factor series to represent the comovement of the underlying dynamics of grouped traffic flows, and idiosyncratic component to represent location-specific traffic characteristics. The dynamic factor model is estimated by the maximum likelihood method via an iterative EM (expectation maximization) algorithm. The traffic state forecast for each location is a combination of the respective forecast from the common factor component and idiosyncratic component. The dynamic factor model exhibits four advantages. It provides an excellent way to (1) seamlessly incorporate spatial correlations among grouped traffic flows into forecast; (2) produce forecast simultaneously for group locations; (3) perform dimension reduction such that high-dimension grouped traffic time series can be modeled at a low-dimension space; (4) consider not only location-specific information but also global common dynamics in the forecast. Meanwhile, it also has capacity to accommodate typical characteristics of traffic flows including temporal correlation, seasonality, structural change in mean and/or covariance function, and cointegration. Forecast accuracy is significantly improved across highway network as well as urban road network in comparison with the Sparse VAR and ARIMA models. The proposed method is suitable for large-scale network traffic forecast in the context of big data environment. This research provides sufficient evidence that justified the importance and efficacy of spatial correlation for predictive accuracy and attempts to fill the gaps in literature of traffic forecast for groups of locations.