Discrete choice models (DCMs) are used to forecast demand in a variety of engineering, marketing, and policy contexts, and understanding the uncertainty associated with model forecasts is crucial to inform decision-making. This thesis evaluates the suitability of DCMs for forecasting automotive demand. The entire scope of this investigation is too broad to be covered here, but I explore several elements with a focus on three themes: defining how to measure forecast accuracy, comparing model specifications and forecasting methods in terms of prediction accuracy, and comparing the implications of model specifications and forecasting methods on vehicle design. Specifically I address several questions regarding the accuracy and uncertainty of market share predictions resulting from choice of utility function and structural specification, estimation method, and data structure assumptions. I1 compare more than 9,000 models based on those used in peer-reviewed literature and academic and government studies. Firstly, I find that including more model covariates generally improves predictive accuracy, but that the form those covariates take in the utility function is less important. Secondly, better model fit correlates well with better predictive accuracy; however, the models I construct— representative of those in extant literature— exhibit substantial prediction error stemming largely from limited model fit due to unobserved attributes. Lastly, accuracy of predictions in existing markets is neither a necessary nor sufficient condition for use in design. Much of the econometrics literature on vehicle market modeling has presumed that biased coefficients make for bad models. For purely predictive purposes, the drawbacks of potentially mitigating bias using generalized method of moments estimation coupled with instrumental variables outweigh the expected benefits in the experiments conducted in this dissertation. The risk of specifying invalid instruments is high, and my results suggest that the instruments frequently used in the automotive demand literature are likely invalid. Furthermore, biased coefficients are not necessarily bad for maximizing the predictive power of the model. Bias can even aid predictions by implicitly capturing persistent unobserved effects in some circumstances. Including alternative specific constants (ASCs) in DCM utility functions improves model fit but not necessarily forecast accuracy. For frequentist estimated models all tested methods of forecasting ASCs improved share predictions of the whole midsize sedan market over excluding ASC in predictions, but only one method results in improved long term new vehicle, or entrant, forecasts. As seen in a synthetic data study, assuming an incorrect relationship between observed attributes and the ASC for forecasting risks making worse forecasts than would be made by a model that excludes ASCs entirely. Treating the ASCs as model parameters with full distributions of uncertainty via Bayesian estimation is more robust to selection of ASC forecasting method and less reliant on persistent market structures, however it comes at increased computational cost. Additionally, the best long term forecasts are made by the frequentist model that treats ASCs as calibration constants fit to the model post estimation of other parameters.
| Identifer | oai:union.ndltd.org:cmu.edu/oai:repository.cmu.edu:dissertations-1491 |
| Date | 01 December 2014 |
| Creators | Haaf, Christine Grace |
| Publisher | Research Showcase @ CMU |
| Source Sets | Carnegie Mellon University |
| Detected Language | English |
| Type | text |
| Format | application/pdf |
| Source | Dissertations |
Page generated in 0.002 seconds