Assumptions Inherent in Assessing Traffic Forecast Accuracy

Abstract

A dozen approaches such as extrapolation of previous traffic counts, trip rates based on expected future land development, and application of regional travel demand models can be used to forecast traffic volumes. Because of Virginia’s interest in knowing the accuracy of forecasting approaches, the researchers compared previous Virginia forecasts with observed volumes. These comparisons revealed five observations affecting how accuracy is evaluated. First, the statistic used to aggregate individual link forecast errors may influence the perception of accuracy; for one study, two percentage-based statistics—the median percentage error and the mean absolute percentage error—yielded errors of 2% and 43%, respectively. (An implication is that the appropriate error statistic depends on the forecast’s purpose.) Second, the accuracy of different types of traffic forecasts varies by magnitude; the median absolute percentage error for Virginia studies ranged from 12% (for a site-specific land development study) to 72% (for statewide forecasts based on historic traffic volumes). Third, because of differences in geometry and the frequency of counts, there may not be a direct relationship between a link’s forecast and its observed volume; in one study, the error ranged from almost 0% to 36%, depending on how the forecast and observed volumes were compared. Fourth, some methods require input data that must be forecast, and different indications of accuracy result (e.g., mean absolute percentage errors of 13% versus 34% for the Fratar method) depending on whether the input data are forecast with or without error. Fifth, depending on the chosen decision criterion, large errors may not necessarily affect the actions taken.