1981 年 12 巻 p. 85-103
Transport forecasts are subject to great uncertainty which is caused by a variety of errors and assumptions through the estimation process. The decision in the transport planning should be made in the context of uncertainty, but it has been ignored because of poor means to quantify it.
The purpose of this paper is to describe the error structure of transport prediction and to assess the magnitude of effect of each error sources on forecasts. After possible sources of forecast error are examined, three factors, i.e., the accuracy of data for model caliblation, the specification error of models and the accuracy of input to predictor variables at the target year, are identified as the main sources of forecast error. The unfitness of forecasts which are caused by different level of errors of these three factors is measured by numerical simulation. The outcomes of the research are outlined as below.
Firstly, the effects of these factors, defined as respectively variances of error of observed data, ones of residual of model and ones of error of input variables, are formulated in the case of multiregression model. It shows that Mahalanobis' generalized distance between means of data and forecasts as well affects on the accuracy of forecasts.
Secondly, the magnitude of effect of three factors on forecasts error are quantitatively assessed through numerical simulation in the case of four-step transport prediction. The data precision brings relatively small error to forcasts, while specification error and ones of input variables cause rather great forecast errors.
Thirdly, trade-off relationship between specification error and error of input variables is pointed out from the viewpoint of forecast error. This represents the trade-off between preciseness and complexity of a model. The more complex model is adopted in order to decrease specification error, the more forecast error is caused by the error of input variables. This suggests that the fitness of a model is not sufficient criteria to choose a prediction model and that a relatively simple model has advantage when input variables are rather vague.