Is XGBoost better than logistic regression to predict claiming in motor insurance with telematics data?

J. Pesantez-Narvaez, M. Alcañiz, M. Guillen

Machine-learning algorithms are widely used for predicting different types of phenomena in all areas, and in particular XGBoost is recognized for its exceptional predictive capacity. To derive the determining factors associated with traffic accidents, dichotomous response models indicating the existence of accident claims versus no claim can be used. We aim to compare the performance of logistic regression and the new XGBoost approach when predicting the existence of accident claims using telematics data. The dataset contains information from an insurance company including individual driving patterns such as total annual distance driven and percent of distance driven in urban areas. Findings show that the logistic regression is a suitable model due to its interpretability and good predictive capacity. XGBoost demands numerous tuning-model procedures to improve the logistic regression model predictive performance and needs an additional effort for interpretability.

Keywords: binary response; predictive model; tree boosting; GLM

GT15-2 Risk Analysis
September 6, 2019  11:20 AM
I3L8. Georgina Blanes building

C. Bolancé Losilla, R. Vernic, R. Alemany Leira

P. Ortega Jiménez, M. Á. Sordo Díaz, A. Suárez Llorens