Lessons Learned from Backcasting and Forecasting Exercises
Corresponding Author: Thomas Rossi, Cambridge Systematics, Inc.
Presented By: Tom Rossi, Cambridge Systematics, Inc.
Travel demand model validation involves running the model for a “base year” and comparing the outputs to observed data such as traffic counts, travel times and speeds, transit ridership, and other measures of travel demand. Sensitivity testing and temporal validation are also critical components of any model validation effort.
Many agencies’ model validation efforts include “backcasts,” where the model is run for a year prior to the base year of the model and compared to observed data from that year. The second point of reference besides the base year provides additional confidence that the model is reasonably sensitive to changes in conditions that affect travel demand. When the base year is far enough in the past, the model can be used to forecast conditions for a point in the recent past after the base year and to compare the model results to observed data.
The Federal Highway Administration (FHWA) undertook a project to provide information for agencies performing this part of the validation process. This research was intended to produce useful data on which model components are most stable over time and their sensitivities to the factors affecting travel demand that vary over time. The models for two U.S. metropolitan areas (Cincinnati and Baltimore) were chosen as case studies for this work. For each region, an earlier version of the model was run for the base year and a forecast year, and the current model version was run for the base year and a backcast year.
The presentation will compare the results from the four model runs for each region and will present the lessons learned from these comparisons. We will discuss how the model results address questions such as:
• How well does the model forecast/backcast for the scenario several years removed from the base year?
• Does the model perform appreciably better for forecasting or backcasting?
• Are there particular items where the model performs better for forecasting/backcasting? (These may be defined by geography, land use type, travel mode, levels of congestion, time periods, and other segmentations.)
• Does the model show reasonable sensitivity to the factors that changes between scenarios?