Estimating the uncertainty of traffic forecasts from their historical accuracy

Traffic forecasters may find value in expressing the uncertainty of their forecasts as a range of expected outcomes. Traditional methods for estimating such uncertainty windows rely on assumptions about reasonable ranges of travel demand forecasting model inputs and parameters. Rather than relying on assumptions, we demonstrate how to use empirical measures of past forecast accuracy to estimate the uncertainty in future forecasts. We develop an econometric framework based on quantile regression to estimate an expected (median) traffic volume as a function of the forecast, and a range within which we expect 90% of traffic volumes to fall. Using data on observed versus forecast traffic for 3912 observations from 1291 road projects, we apply this framework to estimate a model of overall uncertainty and a full model that considers the effect of project attributes. Our results show that the median post-opening traffic is 6% lower than forecast. The expected range of outcomes varies significantly with the forecast volume, the forecast method, the project type, the functional class, the time span and the unemployment rate at the time forecast is made. For example, consider a 5-year forecast for an existing arterial roadway made in 2019 when the state unemployment rate was 4% using a travel model. If a travel model predicted 30,000 Average Daily Traffic (ADT) on this road, our results suggest that 90% of future traffic volumes would fall between 19,000 and 36,000 ADT. A forecaster can apply the resulting equations to calculate an uncertainty window for their project, or they can estimate new quantile regression equations from locally collected forecast accuracy data. Aided by decision intervals, such uncertainty windows can help planners determine whether a forecast deviation would change a project decision.

Automated summary

Researchers demonstrate a method to estimate uncertainty in traffic forecasts using past forecast accuracy measures, rather than relying on assumptions. They develop an econometric framework based on quantile regression to estimate expected traffic volumes and the range within which 90% of volumes are expected to fall. Results show a 6% lower post-opening traffic volume than forecast, with the expected outcomes varying significantly based on various factors such as forecast volume, method, project type, and time span.