Machine Learning for Geospatial Sales Potential Estimation

In this thesis, a framework has been developed to identify the potential of geospatial sales in the agriculture & automotive sector. The models used so far do not take into account the territorial, demographic and economic component, often based solely on the data provided by companies, leading to investments and choices based only on internal data, not taking into account the context and possible competitors. By integrating the company's data with geospatial, economic and statistical data, it was possible to develop an exploratory and subsequently elaborate and predictive pipeline. Following various analyzes of the reference KPIs, we defined and tested different models for predicting earnings and identifying the potential of dealer collections on a provincial basis. Subsequently, an analysis of the most important features was carried out to interpret and identify the most significant KPIs in the influence of profit from sales for each province. By taking advantage of the Quantile Regression it was possible to identify all the possible collection scenarios and analyze how as different economic and territorial factors vary, the results in marketing change drastically, allowing the gaps between current and potential values to be identified and improved. From the outcomes we have noticed various differences between provinces based on their location and their geographical conformation, bringing to light details not visible with the classic models. This framework allows companies to intelligently and specifically allocate the resources available and to adapt market policies in a functional and appropriate way, allowing stability and optimization in maintenance and inventory management.