Comparison of Parametric and Non-Parametric Methods for Modeling Establishment-Level Freight Generation

Abstract

The traditional parametric modeling approaches used to predict freight generation (FG), such as ordinary least squares (OLS), suffer from limitations because of the realistic possibility of violating basic assumptions such as linearity or data distribution. This problem is multidimensional because of the need to model numerous industry sectors in the freight system and the possibility of using different explanatory variables; little guidance is currently available on which modeling methodology is suitable for a particular case. The non-parametric models that could solve several limitations with traditional FG models are rarely examined in their predictive ability. This paper offers insights into this frequent research question by comparing the performance of various modeling methodologies that can be used for predicting FG: OLS, weighted least squares, robust regression (RR), seemingly unrelated regression (SUR), multiple classification analysis (MCA), and support vector regression (SVR). To this effect, the research carried out in this study uses a freight dataset of 432 establishments across Kerala State, India. The model outputs are validated using resubstitution and cross-validation methods, and the prediction errors are quantified using root-mean-square error and mean absolute error. The validation results show that the non-parametric SVR models are better alternatives in developing state-, regional- and industrial segment-level models. The MCA models are more precise in predicting FG for suburban models. RR models provide a better predictive ability for modeling FG in some industrial segments. Overall comparison and result interpretations suggest that the non-parametric models are superior in relation to predicting FG. At the same time, RR seems to be the only parametric modeling approach that can provide comparable model performance to non-parametric models.