Driving Risk Assessment Using Near-Miss Events Based on Panel Poisson Regression and Panel Negative Binomial Regression

Abstract

This study proposes a method for identifying and evaluating driving risk as a first step towards calculating premiums in the newly emerging context of usage-based insurance. Telematics data gathered by the Internet of Vehicles (IoV) contain a large number of near-miss events which can be regarded as an alternative for modeling claims or accidents for estimating a driving risk score for a particular vehicle and its driver. Poisson regression and negative binomial regression are applied to a summary data set of 182 vehicles with one record per vehicle and to a panel data set of daily vehicle data containing four near-miss events, i.e., counts of excess speed, high speed brake, harsh acceleration or deceleration and additional driving behavior parameters that do not result in accidents. Negative binomial regression (AICoverspeed = 997.0, BICoverspeed = 1022.7) is seen to perform better than Poisson regression (AICoverspeed = 7051.8, BICoverspeed = 7074.3). Vehicles are separately classified to five driving risk levels with a driving risk score computed from individual effects of the corresponding panel model. This study provides a research basis for actuarial insurance premium calculations, even if no accident information is available, and enables a precise supervision of dangerous driving behaviors based on driving risk scores.

Interpolation of Quantile Regression to Estimate Driver’s Risk of Traffic Accident Based on Excess Speed

“…They also found that the relationship between total distance driven and total distance driven above the legal speed limit is relevant to produce a reference chart and a better fit of the percentiles. Sun et al (2020Sun et al ( , 2021 adjusted ordinary least squares and binary logistic regressions to calculate a driving risk score for different drivers using internet of vehicles (IoV) data. Usage-based insurance is a new methodology based on IoV that is being used to customize insurance prices.…”

“…Pay-how-you-drive (PHYD) also considers available information on driving style. In PHYD, the price is based on relevant information about driving patterns and may increase when dangerous indicators arise Sun et al 2021). Another insurance pricing method that is still under development is known as manage-how-you-drive (MHYD).…”

Bivariate Poisson Credibility Model and Bonus–Malus Scale for Claim and Near-Claim Events

“…Another approach consists in modeling telematics signals as supplementary response variable along claim counts. As part of this approach, Guillen et al (2020) and Sun et al (2021) defined different types of near-miss events (term borrowed to the aviation safety) such as excess speed, harsh acceleration, harsh deceleration or cornering events. They highlighted that significant risk factors (traditional and telematics) depend on the type of near-miss events under consideration.…”

“…Bonus-malus scales are effectively modeled as Markov Chains and relativities are generally computed from the corresponding stationary distribution. In this paper, we propose a novel scale, integrating both claim and near-claim events in a simple and transparent way and we adapt the approaches proposed by Pitrebois et al (2003) and Tan et al (2015) to derive the relativities.…”

Sensor Data Analytics: Challenges and Methods for Data-Intensive Applications

“…Sensor placement optimisation is the topic of another work [ 18 ], using Gaussian priors and the Fisher Information Matrix (FIM) to show important properties that can enhance recommendations on the best possible location for a given device set. Sun et al present an interesting application of sensor data analytics to estimate vehicles accident risk [ 19 ]. This is an emerging topic that raises significant interest among insurance companies, taking advantage of the more precise tracking capabilities enabled by built-in sensors installed in vehicles.…”