Costruzione di un modello di rischio assicurativo basato sul comportamento di guida

More and more insurance companies are proposing to motorists to install the "black box" on board the vehicle. Black boxes are a solution that very often brings savings to the customer and provides insurance companies with a lot of data such as GPS coordinates at any given time or any vehicle accident. A Geographic Information System (GIS) is commonly used to extract information from data emitted from black boxes. It is therefore an IT system capable of associating the data with their geographical position on the earth's surface and of processing it to extract information from it. In the third chapter the world of big data will be introduced explaining both what they are, and introducing the used technologies to work with them. One of the characteristics of the GIS system developed by AgileLab (the company in which I carried out the thesis project) is the map matching, a system that combines the registered geographical coordinates with a logical model of the real world: the roads. The first objective of this project was to add a functionality to the map matching system: the calculation of information such as distance and travel time, between the different GPS points emitted by a black box in the same trip. The fourth chapter describes some notions about machine learning, a very useful tool used in the world of big data. One of the main problems of a map matching system mentioned earlier, is related to the quality of the data, in fact if a sequence of GPS points shows anomalies (points inconsistent with the others) the calculated route can be very different from the one actually traveled from a vehicle. AgileLab's GIS already had an anomaly detection system, however it was able to identify only the most evident anomalies. The second goal of this project was therefore to create a more precise and uncompromising GPS point anomaly detection system. In the fifth chapter, the anomalies relating to the GPS points of travel will then be described, the GPS anomaly detection system used up to that moment, the development of the new anomaly detection system and finally a performance comparison between the two systems. The final goal of the project was to build a system capable of returning a representative index of driving behavior. of a user. The important data for the study and validation of this system are those relating to accidents. As will be read in the thesis, the black boxes have a system inside to detect crashes, however they are not always accidents, on the contrary, most of the times they are false positives. There was an obstacle during the development of this model: the customer for whom this system was intended could not provide us with the accident data for privacy reasons. To solve this problem, I made the decision to try developing a system for detecting real incident detection. In the sixth chapter some notions on the analysis of the signals that have been useful for me to develop the accident detection system will be described, three different systems developed with different techniques able to distinguish real accidents from false positives will be presented and the achieved results. Finally, the last chapter will describe the development of the driving behavior system which will provide a useful index for the creation of an insurance risk model.