Dimension reduction techniques for summarized telematics data

While use of driver telematics data gained popularity in automobile insurance, dealing with high dimensionality of telematics data used in traditional ratemaking models has gained interest from insurers. In this article, a dimension reduction scheme is proposed based on categorical embedding and principal components analysis to handle both the categorical and continuous covariates efficiently and calibrate a more interpretable and reliable predictive model without losing the essential information of the data. According to numerical studies, the proposed data processing scheme produces more stable predicted values and reasonable goodness-of-fit compared to the classical GLMs without proper dimension reduction.