Is it all about you or your driving? Designing IoT-enabled risk assessments

• 成果类型: Article
• 署名作者: Ho, Yi-Jen (Ian); Liu, Siyuan; Pu, Jingchuan; Zhang, Dian
• 署名单位: Pennsylvania Commonwealth System of Higher Education (PCSHE); Pennsylvania State University; Pennsylvania State University - University Park; State University System of Florida; University of Florida; Shenzhen University
• 刊物名称: PRODUCTION AND OPERATIONS MANAGEMENT
• ISSN/ISSBN: 1059-1478
• DOI: 10.1111/poms.13816
• 发表日期: 2022
• 页码: 4205-4222
• 关键词: driving risk; Internet of Things; in-vehicle camera; on-board diagnostics; usage-based insurance

摘要

Technological applications disrupt the way to assess risks in the auto-insurance business. Contrasted with the common practice based on static demographics, usage-based insurance predicts risks using driving data collected from Internet-of-things-enabled telematics. This study proposes a novel solution leveraging the synergy between big data and hierarchical modeling. We specifically consider two aspects of mobility, namely, trait and trajectory, monitored by global positioning system (GPS), on-board diagnostics, and in-vehicle cameras in real time. Traits here refer to drivers' distinctive driving behaviors (styles), whereas trajectories consist of the vehicle motion sequences and the contextual factors on trips. We operationalize semantic features of the two to assess risks at both trip and driver levels. Using fine-granular driving data and crash reports, we find that behavioral traits play a significant role in predicting crashes, given individual heterogeneity and temporal dynamics. In a series of empirical validations, the proposed solution outperforms the current practice and alternative predictive models considered by prior literature. We show that the mobility-based models are superior to the demographic-based ones. Moreover, our model achieves the comparable performance of neural networks, improving the recall of class-weighted logistic regression, nested support vector machine, and cost-sensitive random forests by 44.23%, 29.18%, and 24.59%, respectively. Last, our approach is robust, data independent, and computationally efficient for skewed and small samples. This study provides several managerial implications and a blueprint for the auto-insurance industry to operationalize IoT-enabled risk assessments in the era of 5G communication.