Aggregation Level, Variability, and Linear Hypotheses for Urban Delivery Generation Models

Paper in proceeding, 2016

Category classification models are nowadays one of the main approaches in urban goods transport demand generation, and we see applications of those frameworks in different contexts, both in Europe and the USA. Those models are developed for a fixed category classification, which can be more or less disaggregated. However, those models depend on the data quantity and quality, and the category construction can have a real impact on the model quality. This paper aims to analyze how the data aggregation of category classifications affects the significance and the quality of the model deployed. To do this, a dispersion analysis is combined with the search of linear relations between the number of deliveries and the employment, for each category given. After motivating the proposed via a synthesis of the literature, this paper present the methodology used to assess the relations between data aggregation and model quality. The authors also present the data used to carry out the analysis and make a first description of sample size and dispersion, via the estimation of coefficients of variation for each category, regarding the number of deliveries and the number of deliveries per employee. Then, the results of the analysis are presented. The presented results are obtained from the assessment and comparison between a constant model, a linear model with constant and a linear model without constant. Each model prediction quality is estimated using the root mean square error (RMSE) as indicator. The best model is then selected. Finally, results are discussed and the choice of a category aggregation is proposed.