| dc.description.abstract | Traditionally, car insurance companies have had to rely on very rough estimates of risk posed by customers, based on age, gender, postal code, etc. Newer technologies such as the smart phone, custom made telematics boxes that can be placed inside vehicles, in addition to more powerful integrated computers in the vehicles themselves, that are connected to vehicle sensors and the engine, all make it possible to collect driving data and use this as a basis for a more accurate risk assessment. UBI (usage based insurance) is a disruptive technology which relies on insurance premiums that reflect the risk profile of the driver \citep{handel2014insurance}. In recent years there have been many projects \citep{ambrosa2017big} that have had the aim of finding the best method of gathering driving data (what hardware and sensors to use), and then how to use the data in order to infer the risk level of a particular driver.
We examine data gathered, and the scoring algorithm used, in ``Telematikk'', a pilot project between a technology startup, Telemotix, and the insurance company Tryg. We start with a preliminary EDA (Exploratory Data Analysis), where we get a more general overview of the data. We proceed with creating our own algorithm for processing the data and dividing them up into non-overlapping circles (topographically), that represent all the driving records from within its borders. We name these circles ``road stretches'', and they contain summary information about the records they represent, such as average speed, number of hard brake incidents, etc. We further process the road stretches to create two lists, top 10 worst speeding road stretches and top 10 worst incident road stretches, which can be explored in the map files accompanying this thesis. We then did a more thorough EDA based on these ``worst areas''. We found some weaknesses with our method, such as a single road stretch containing data from different roads that are near to each other. We also uncovered that the speed limit data accompanying the driving data records were incomplete. After reimplementing the Telemotix scoring algorithm, verifying it, and then modifying it, we show how the driver score distribution changes. We were unable, as initially planed, to verify that our algorithm is better, which we planned to use accident statistics to do, because the sample size was too small, an not enough accidents had been reported. | en |
