| description abstract | When design decisions are informed by consumer choice models, uncertainty in choice model predictions creates uncertainty for the designer. We investigate the variation and accuracy of market share predictions by characterizing fit and forecast accuracy of discrete choice models for the US light duty new vehicle market. Specifically, we estimate multinomial logit models for 9000 utility functions representative of a large literature in vehicle choice modeling using sales data for years 2004–2006. Each model predicts shares for the 2007 and 2010 markets, and we compare several quantitative measures of model fit and predictive accuracy. We find that (1) our accuracy measures are concordant: model specifications that perform well on one measure tend to also perform well on other measures for both fit and prediction. (2) Even the best discrete choice models exhibit substantial prediction error, stemming largely from limited model fit due to unobserved attributes. A naأ¯ve “static†model, assuming share for each vehicle design in the forecast year = share in the last available year, outperforms all 9000 attributebased models when predicting the full market one year forward, but attributebased models can predict better for four year forward forecasts or new vehicle designs. (3) Share predictions are sensitive to the presence of utility covariates but less sensitive to covariate form (e.g., miles per gallons versus gallons per mile), and nested and mixed logit specifications do not produce significantly more accurate forecasts. This suggests ambiguity in identifying a unique model form best for design. Furthermore, the models with best predictions do not necessarily have expected coefficient signs, and biased coefficients could misguide design efforts even when overall prediction accuracy for existing markets is maximized. | |
