Keep it simple stupid! A non-parametric kernel regression approach to forecast travel speeds

Abstract

The approach taken by the second place winner of the TRANSFOR prediction challenge is presented. The challenge involves forecasting travel speeds on two arterial links in Xi'an City in China for two five hour periods on a single day. Travel speeds are measured from trajectory information on probe vehicles from a fleet of vehicles for a large sub-area of the city. After experimenting with several deep learning methods, we settle on a simple non-parametric kernel regression approach. The method, borrowed from previous work in fixed route transit predictions, formalizes the intuition that in urban systems most failure patterns are recurrent. Our choice is supported by test results where the method outperformed all evaluated neural architectures. The results suggest simple methods are very competitive, particularly considering the high lifecycle cost of deep learning models.