A Retrospective Evaluation of Traffic Forecasting Accuracy: Lessons Learned from Virginia

Understanding the accuracy of techniques for forecasting traffic volumes for a future year, such as extrapolation of previous years traffic volumes, use of regional travel demand models, and use of local trip generation rates, can aid analysts in considering the range of transportation investments for a given location. To determine this accuracy, forecasts from 39 Virginia studies (published from 1967-2010) were compared to observed volumes for the forecast year. The comparison enabled the identification of potential assumptions that might cause variations in the way forecast accuracy is assessed. Some of the assumptions include construction of proposed infrastructure (A new bridge in York river crossing), the appropriate error statistics (Average value or median value of error), chosen observed volume (Volume from continuous count or periodic count), anticipated alignment (relocation of Route 33 where Route 3 and Route 33 are signed together).
Excluding statewide forecasts, the number of roadway segments in each study ranged from 1 to 240 links. For each segment, the difference between the forecast and observed volume divided by the observed volume gives a percent error, such that a segment with a perfect forecast has an error of 0%. The analysis showed that based on 39 Virginia past studies, the median absolute percent error ranged from 1% to 134% with an average value of 40%; forecast volumes tended to be larger than observed volumes. The accuracy of different types of traffic forecasts varies by an order of magnitude: 12% (for a site-specific land development study) to 72% (for statewide forecasts based on historic traffic volumes). The importance of forecast accuracy is determined by whether such errors have any impact on decision taken (i.e. signal warrant, change of road alignment, no remedy for environmental impact).
Slightly more than one-fourth of the variation in such error (29%) was explained by three identifiable factors: the forecast method, the forecast duration (number of years between the base and forecast years), and the number of economic recessions in the same interval (p 0.04). Interaction effects matter: the first two factors have significant and expected impacts on accuracy only if economic changes are explicitly considered. Finally, link-by-link error in a study has sufficient variation (p = 0.02) such that if this variation is not controlled, explanatory factors are impossible to detect. Although no forecast is perfect, this study provides an indication of expected forecast error for future studies that might help decision makers to evaluate transportation needs.