Analysis and Development of Methods for Automotive Market Segmentation

In the automotive sector in recent years the use of data has become increasingly important to allow car manufacturers to analyse the market, competing vehicles, their specifications and their prices so that they could use this information to create a product offering that adapts as much as possible to the market. Some of the more significant analyses require vehicle segmentation. Segmentation represents the division of vehicles into commercially homogeneous groups. The market analysis required by car manufacturers are typically conducted using as comparison vehicles from the same segment. This type of analysis is called basket analysis, where the basket would be the set of vehicles considered most similar to the vehicle being compared. Jato has developed its own more general definition of segmentation and based on this subdivision catalogues the models present in the databases. The definition of the segments and the choice of the segment to be assigned to the vehicles, however, are not immediate processes. In order to correctly identify which segments the various models belong to, it is important to identify physical/performance characteristics that allow to associate the segment label more easily. The first part of the thesis was aimed at defining the characteristics that allow the segmentation to be carried out and to develop a tool capable of carrying out an automatic segmentation of the models that can be used as a support for the segmentation phase. The model developed is based on a random forest algorithm that allows to obtain an automatic classifier to support the choice of the segment. A total of 20 classes were predicted. The first model achieves an 84% accuracy on a set of 293 models. The results were limited by the fact that the total dimension of the population (the total model population) is scarce and by the fact that in some cases segmentation may be driven by commercial considerations difficult to capture. The classifier may thus be employed as a support to classification, but not as a standalone classifier. For this reason, it was interesting to identify a subset of models for which the classification is almost certain, on which a standalone classifier could be applied. A threshold on the confidence of the classification was thus set, and automated classification was performed on the high confidence set. The confidence threshold used was 50%, this value was chosen following a heuristic approach using the Italian regional dataset as a training set. On the high confidence subset, the accuracy of the model reaches more than 96%. In order to validate the choice made, the same analyses were also applied to the British, Russian and German markets, obtaining similar results demonstrating the robustness of the model. The second part of the work was aimed at allowing to systematically define the vehicles that are considered similar to a given vehicle in order to be able to define more precise baskets for basket analysis. Two similarity measures were developed to identify similar vehicles at model level and similar vehicles at version level. Given the presence of categorical and continuous features, the proposed solution is based on the Gower distance measurement that allows to manage both types of variables by combining the use of the Dice coefficient for categorical features and the Manhattan distance for continuous features.