Abstract

The development of big data technologies and in-vehicle devices has contributed to the growth of Usage-Based Insurance (UBI) in recent years. These in-vehicle devices, such as GPS and sensors, collect certain variables that can represent the driving behaviour of policyholders. This collected data, called telematic data, consist of several variables that have strong relationship with likelihood of having an accident. Consequently, one can use telematic data to improve risk assessment and personalize car insurance premiums. In this thesis, a synthetic car insurance dataset emulated from a Canadian-based insurance company is used to investigate the use of telematic data in predicting the likelihood of having an accident. More precisely four machine learning techniques—logistic regression, random forests, gradient boosting trees, and feed-forward neural networks—are employed to predict the risk of having an accident. Actuaries often use white box machine learning methods like logistic regression for risk assessment due to their interpretability. However, these method are unable to detect non-linear relationships between variables accurately. Therefore, more
complex machine learning techniques such as random forests, gradient boosting trees, and feed-forward neural networks are used to achieve more accurate risk assessment for accidents.
In addition, two variable importance assessment methods—Shapley decomposition and marginal performance loss upon feature removal—are employed to provide insights into the
feature contributions in the overall predictive performance of the models.