Research of the Factors That Determine the Accident Rate of Individual Drivers and Prediction of Probability of Getting Into an Accident

In the paper, the problem of revealing the factors influencing the probability to get into an accident is investigated. It’s relevant as in the insurance business it can be applied for setting a car insurance premium as current psychological tests and personal dossier do not allow to cope with moral hazards issue which can be handled via differentiating of insurance premiums for risk-averse and risk-takers depending on recent driving style. The main hypothesis of the paper is that driving style characteristics can be used for determining chances to get into an accident. There is no business implementation yet as insurance companies have no access to such data. So, the novice of the research is the method that is proposed for measuring the riskiness of driving style - from ride tracking data via kinematic sensors installed on a car. Tested machine learning algorithms allow to achieve 99% accuracy of identification of accident on a daily data and up to 50% of accidents happening next month are predicted having only current week’s aggregated tracking data. Main insight from the research is that logistic regression is performing as well as more advanced algorithms, and they blended together give 0.72 AUC and 0.77 F1-score for next month accidents forecasting, and variables like the fact of a crash during the current week, maximal speed and frequency of braking are enough for obtaining such result.