An empirical study of predicting car type choice in Sweden using cross-validation and feature-selection

In this paper we analyze the prediction problem and focus on building a multinomial logit model (MNL) to predict accurately, the market shares of new cars in the Swedish car fleet in the short-term future. Also, we investigate whether or not different prediction questions lead to different 'best' models specifications. Most of the studies in the field, take an inference-driven approach to select best models to estimate relevant parameters and project the results to the future, whereas we do take a prediction-driven approach. We use feature (variable) selection and cross-validation algorithms to improve predictive performance of models. These methods have been extensively used in other fields such as marketing but are scarce studies employing them in the choice modeling field. Additionally, we introduce four different prediction questions or loss-functions: overall prediction (log-likelihood), brand market share, ethanol (E85)/brand market share, and total share of ethanol cars and the predicted results of these models are compared. The results show that 'best' models prediction depend different prediction questions to answer. Also, they indicate that log-likelihood does not perform accurately when the objective is to predict a sub-section of population such as total share of E85 cars.