Yifan Huang, Shengwang Meng

【Abstract】

Usage-based insurance (UBI), given the development of in-vehicle networking and big data technologies, has received a growing amount of attention in recent years from both insurers and policyholders. The UBI product derives certain driving behavior variables from telematics data, which have stronger causal relationships with accidents and thus effectively improve the pricing accuracy of automobile insurance. This paper mainly investigates the use of extensive driving behavior variables in predicting the risk probability and claim frequency of an insured vehicle. More specifically, logistic regression and four machine learning techniques - support vector machines, random forests, XGBoost, and artificial neural networks - are employed as risk probability models, while Poisson regression as claim frequency model. In addition, aiming at the interpretability requirements of insurance pricing, a data augmentation method of variable binning is adopted to discretize continuous variables and construct tariff classes with significant predictive effects. As a result, our pricing framework can simultaneously improve the interpretability and predictive accuracy of the model, and thus provides a novel solution to implement classification ratemaking for UBI products. The empirical results, based on a dataset from a property and casualty insurance company in China, show the selection of significant variables and the estimation of their specific effects on driving risk, verifying the great potential of driving behavior variables in automobile insurance.

【Keywords】

usage-based insurance, telematics data, driving behavior, classification ratemaking, machine learning