Machine Learning in an automotive B2B setting: prediction of dealer's propensity to buy and potential buy-aways detection

The capacity of making reliable predictions and knowing how to exploit them profitably, is fundamental to support the business of any company, whatever the sector in which it operates. In this perspective, Machine Learning algorithms, Data Analysis and Big Data are extremely useful and widely applied tools. This work is carried out in collaboration with CNH Industrial, a global leader in agricultural and construction machinery and services, providing a wide range of replacement parts to thousands of dealers worldwide. The aim of the thesis is to make predictions on dealers propensity to buy a specific part in a certain month, using Classification learning algorithms. A dealer is labelled as positive if he has purchased at least one piece of that part during the month. In the early stages, the work is focused on dealers operating in ten European countries and on the agricultural sector only. The sample used for the training and validation phases is built by collecting dealers orders placed from January 2021 to December 2022, while the testing set contained those registered during the first two months of 2023. Three different models are trained (i.e., Logistic Regression, Random Forest and Extreme Gradient Boosting Classifier) by combining already available features, for example related to the "machine parc" associated with each dealer, and defining new ones. They are fitted on different segmentations of the data sample: each model is firstly assessed separately for every market region, then each subgroup is further partitioned with respect to different categories of parts. The second level of segmentation is only applied to the French market. In any case, the Extreme Gradient Boosting model shows the best predictive performance, especially in terms of Precision, and therefore it is selected for the testing phase. The results of the model are exploited to try to manage another critical issue: the buy-away phenomenon. It refers to a dealer who buys a part usually provided by CNH, from another supplier. The percentage of false positives (the cases for which the model predicts a purchase but the dealer does not actually buy the part) predicted during the testing phase contains both unavoidable model errors and potential buy-aways. The construction of metrics that combine historical trends of purchases (sell in) and sales (sell out) of the dealers, and the comparison with features associated with homogeneous groups of dealers, allow us to identify possible buy-aways. A work of this kind has an interesting potential in the business field. It can be exploited to recognize dealers with a high willingness to buy specific parts, and at the same time it may be helpful for devising marketing strategies (specific promotions or discounts) to submit to dealers who could be supplied by competitors.