Robust Regression Methods for Traffic Growth Forecasting

Abstract

Least-squares regression has been applied as a tool to understand traffic growth patterns and to predict future growth. Specifically, given a set of historical annual average daily traffic (AADT) values for a location, regression can be used to summarize traffic growth patterns and to predict growth. However, this technique is vulnerable to outliers because standard linear regression techniques can produce arbitrarily large errors in their results if points are badly placed. The situation is made worse when thousands of traffic sites are analyzed at once because it is infeasible to examine each set of regression results individually. In this paper two outlier detection and removal techniques and one robust regression technique are compared with simple least-squares regression for accuracy in traffic growth prediction, with both linear and log-linear models of traffic growth on historical AADT values for several thousand sites in the state of New York. Each method was evaluated by the median absolute error in predictions being computed for 1 year, 4 years, and 8 years beyond the modeled values and also by the mean percent error being computed, giving each site equal weight. When all sites were equally weighted, the robust regression technique produced significantly better results than either plain regression or outlier detection techniques. Using median absolute error, none of the robust techniques produced significantly more accurate results than ordinary regression.