Testing and Comparing Neural Network and Statistical Approaches for Predicting Transportation Time Series

Abstract

Univariate and multivariate neural network (NN) and autoregressive time series models are compared with regard to application to the short-term forecasting of freeway speeds. Statistical tests are used to evaluate the developed models with respect to temporal data resolution, prediction accuracy, and quality of fit. The results indicate that, by and large, NNs provide more accurate predictions than do classical statistical approaches, particularly for finer data resolutions. Evaluation of model fit indicated that, in contrast to vector autoregressive models, NNs may also provide unbiased predictions. Overall, the findings clearly suggest the need to jointly consider statistical and NN models to develop more efficient prediction models.