Predicting road system speeds using spatial structure variables and network characteristics

Spatial regression is applied to GPS floating car measurements to build a predictive model of road system speed as a function of link type, time period, and spatial structure. The models correct for correlated spatial errors and autocorrelation of speeds. Correlation neighborhoods are based on either Euclidean or network distance. Econometric and statistical methods are used to choose the best model form and statistical neighborhood. Models of different types have different coefficient estimates and fit quality, which might affect inferences. Speed predictions are validated against a holdout sample to illustrate the usefulness of spatial regression in road system speed monitoring. [PUBLICATION ABSTRACT]