Should we use neural networks or statistical models for short-term motorway traffic forecasting

This article discusses the relative merits of neural networks and time series methods for traffic forecasting and summarises the findings from a comparative study of their performance for motorway traffic in France. Whilst it was possible to get a good performance with both neural networks and traditional Auto-Regressive Integrated Moving Average (ARIMA) models when forecasting up to an hour ahead using data supplied in 30-min intervals, a purpose-built pattern based forecasting model known as ATHENA, developed by INRETS, out-performed both these methods somewhat. The ways in which these models relate to the structure of traffic data are discussed and alternative paradigms are proposed.