Fleet Management

Assessing Driver Risk

Post Share

Dr. Ashwin Sabapathy, Director Data Science

Certain driving behaviors have strong correlations with accidents as seen in statistics published over many decades.

While providing excellent insights into contributing factors of accident risk, these statistics, being population based, do not establish any causation between these factors and accidents. NHTSA’s data on fatal accidents, for example, shows that about a fourth of all fatal accidents are related to speeding but cannot conclude that speeding caused the accident. Similarly, 11% of all fatal accidents occurred on wet or snowy roads but this doesn’t allow us to conclude that wet roads alone caused the accidents.

Telematics based data offers a naturalistic study of driving behavior observed at an individual driver level and provides a means to establish the link between specific driving behaviors and accident risk. 

Our team analysed driving behavior for a sample of over several thousand drivers belonging to enterprise fleets in the US over a 3.5 year period (Mar 2014 to Nov 2017). The list of preventable and non-preventable accidents shared by fleets provided crucial data to carry out this analysis and we are very grateful to these fleets. The data was ingested into our cloud based Fleet Safety Intelligence platform equipped with advanced machine learning modeling capabilities.

Our analysis sought to see whether the risk of an accident increases with corresponding increase in braking, speeding and acceleration behaviors as we expect accident risk to increase with higher levels of these behaviors. We developed statistical models (Cox semi-parametric hazard models for those of you familiar with risk models) of time-to-accident with driving behavior parameters as explanatory variables.

Of the behavioral parameters, average Hard Braking events per 100 miles and average minutes of speeding above 80 mph per 100 miles were found to be statistically significant. Exposure risk, as measured by average time in different risk hours, was also significant for average hours driving per day during moderate risk hours (peak hour driving) – driving for an average of 1 hr during peak hours increases the risk of a preventable accident by 2.26 times. We believe that this is a reasonable finding since the fleets considered in the study operate only during the day.

Azuga’s Driver Scores are based on behavioral events weighted by other risk factors like weather, magnitude of event, duration of events, zip code risk, etc. 

We modelled time-to-accident with our scores as predictors and, not surprisingly, found that hard braking and speeding scores were statistically significant predictors. A ten point increase in the Braking Score, reduces accident risk by 46.6% and similarly a ten point increase in the Speeding Score reduces risk by 21.7%. A 10-point increase in the overall Driver Score, which is a weighted composite of individual component scores, corresponds to a 57% decrease in accident risk.

The behavioral models of accident risk we developed might not be presently generalizable for all fleets given that the fleet tracking data used in the models were not randomly selected and could carry some bias. However, we are very excited by our findings that validate our scoring algorithm in terms of risk assessment and are working towards widening the representation of fleets in the models to make it more generalizable. These findings have wider ramifications not just for fleets to identify risky drivers but also for the auto insurance industry in risk pricing, especially as user based insurance begins to find greater traction across the US and our Driver Scores gain wider acceptance as the standard for measuring driving risk.

Please get in touch with our Data Science team (datascience@azuga.com) if you would like your fleet accident data to be analyzed.