Validation of reference forecasts for passenger transport

Abstract

This paper compares a number of forecasts for aggregate passenger transport (by mode) produced in Sweden between 1975 and 2009 with actual outcomes. In addition to the comparison between forecasts and outcomes, we explore to what extent forecast errors are due to erroneous input assumptions, enabling us to give a more fair judgment of the models’ forecasting abilities.

We find substantial differences between forecasts and actual outcomes. In forecasts produced since the early 1990s, road and air traffic growth rates have generally been overpredicted. Aggregate railway growth has been fairly accurate, but commercial long‐distance railway growth has been overpredicted, and the growth of subsidized intra‐regional railway travel has been underpredicted (following vast unanticipated supply increases).

When discussing forecasting errors, it is useful to distinguish between different sources of errors: ‐ Model deficiencies: deficiencies in underlying theories, estimation methodologies, model implementation, data used for estimation and so on. ‐ Differences between cross‐sectional and intertemporal relationships: most transport models that are able to generate detailed forecasts are estimated based on cross‐sectional data. Using such models to predict future reference forecasts thus rests on the tacit assumption that cross‐sectional and intertemporal relationships are equal. ‐ Changes in preferences or behavior: using models to predict the future also tacitly assumes stability in preferences and behavior. ‐ Assumption errors: to produce a reference forecast, it is necessary to make a large number of assumptions about future transport supply and general societal and socioeconomic variables.

Focusing on car traffic forecasts, we show that a very large share of forecast errors can be explained by assumption errors, i.e. input variables turning out to be different than what was assumed when the forecasts were made. Even the original forecasts are much closer to actual outcomes than simple trendlines would have been, and once the input assumptions are corrected, the forecasts vastly outperform simple trendlines.

The models used to produce the forecasts from the last decades are state‐of‐the‐art (at the time) nested logit models, estimated on disaggregate cross‐sectional data, and producing detailed forecasts on network and origin‐destination levels. The present paper focuses on how well the models have managed to predict aggregate numbers of total passenger kilometers by mode some time into the future (typically more than a decade). The fact that they vastly outperform simple time‐series trendline forecasts of aggregate passenger transport is interesting: it indicates that the potential problems of using cross‐sectional models for forecasting intertemporal changes seem to be limited. This tentative conclusion is also supported by the finding that elasticities from the cross‐sectional models are consistent with those from a time‐ series model.