Assessing Driving Risk Through Unsupervised Detection of Anomalies in Telematics Time Series Data

TL;DR

This paper presents a novel unsupervised framework using continuous-time hidden Markov models to detect anomalies in telematics time series data, effectively identifying risky driving behaviors without relying on predefined thresholds or covariates.

Contribution

The introduction of a flexible, unsupervised anomaly detection method based solely on raw telematics data using CTHMM, improving risk assessment and behavioral analysis in driving data.

Findings

Higher anomaly levels correlate with increased accident risk.

Classification models achieve ROC-AUC up to 0.86 for trip-level risk detection.

Significant behavioral differences are observed between drivers with and without claims.

Abstract

Vehicle telematics provides granular data for dynamic driving risk assessment, but current methods often rely on aggregated metrics (e.g., harsh braking counts) and do not fully exploit the rich time-series structure of telematics data. In this paper, we introduce a flexible framework using continuous-time hidden Markov model (CTHMM) to model and analyze trip-level telematics data. Unlike existing methods, the CTHMM models raw time-series data without predefined thresholds on harsh driving events or assumptions about accident probabilities. Moreover, our analysis is based solely on telematics data, requiring no traditional covariates such as driver or vehicle characteristics. Through unsupervised anomaly detection based on pseudo-residuals, we identify deviations from normal driving patterns -- defined as the prevalent behaviour observed in a driver's history or across the population -- which are linked to accident risk. Validated on both controlled and real-world datasets, the CTHMM effectively detects abnormal driving behaviour and trips with increased accident likelihood. In real data analysis, higher anomaly levels in longitudinal and lateral accelerations consistently correlate with greater accident risk, with classification models using this information achieving ROC-AUC values as high as 0.86 for trip-level analysis and 0.78 for distinguishing drivers with claims. Furthermore, the methodology reveals significant behavioural differences between drivers with and without claims, offering valuable insights for insurance applications, accident analysis, and prevention.