The rest of this piece is about reading a driver risk score honestly. Two scores that look the same can mean different things, and two scores that look different can describe drivers who face different roads rather than different drivers.
Why driver risk scores exist
Insurance underwriting historically rested on claims history and demographics. Both describe what has already happened. Neither describes what the driver is doing this month, which is why a driver can drift for two renewal cycles before the loss data admits it.
A driver risk score, built on observed driving, exists to close that gap. Regional risk products (like Nexar's Risk Index) do the same thing at the level of a road or a region: they exist because insurers and autonomous-vehicle companies asked for risk differentiation per region, instead of the traditional static factors that do not change over years. The driver-level and region-level products are cousins, and the discipline of reading either one honestly is the same.
What a driver risk score actually measures
Two families of signal dominate.
The first is motion. Hard braking, hard cornering, hard acceleration, and verified collisions. Nexar's Risk Index counts exactly these signals (IMU-based hard braking, cornering and acceleration events, plus verified collisions) at the regional level. The driver-level version of the same signal set is the input to most fleet-side driver risk scores.
The second is behavior, from a driver-facing camera. Distraction, phone use, following distance, drowsiness. Motion signals alone read the vehicle; behavior signals read the driver. The two together are how a driver risk score separates driving-the-car from being-the-driver.
The denominator problem
A raw event count rewards volume. A driver with more observed miles will show up in more events, whether or not they are riskier per mile. Every credible driver risk score normalizes by exposure. On Nexar's Risk Index, a score is the number of driving events in a region, normalized by exposure. The denominators available are miles driven, hours driven, or cameras deployed.
Pick a different denominator and the score changes. That is why the denominator has to be published alongside the score, at both the driver and region level. A driver risk score that hides the denominator is a guess with authority attached.
Time of day
Risk is not evenly distributed across the day. Risk can be calculated separately for daytime and nighttime driving. Treating a combined figure as a single truth about a driver hides the half of the day where the difference usually lives. A driver whose daytime score is excellent and nighttime score is poor is a different case from a driver whose scores are the same in both windows; treating them the same is what sends a coaching program in the wrong direction.
Only we know: how a driver's use pattern is identified without asking
Most driver risk models start with a self-declared use case: personal, commercial, rideshare, delivery. Nexar's exposure data does not need the self-declaration. A rideshare driver is recognizable from the signature of many short local trips, rather than from a registration document.
That method matters for two reasons. First, self-declaration is unreliable; drivers do not always report their actual use. Second, the shape of use changes: a driver who signs up as personal-use and starts running for a delivery platform on weekends is doing something different from what their insurance application says. A driver risk score that reads use pattern from observation, rather than from declaration, does not have to wait for a claim to notice the change.
Where a driver risk score fits
Three uses cover most of the market.
Insurance underwriting and pricing. A score that updates monthly is a different input from one that updates at renewal. Autonomous-vehicle developers, for their part, mostly want the raw exposure data rather than the score itself, because they are comparing autonomous safety against human driving and the ranking is less useful to them than the underlying denominator.
Fleet coaching prioritization. The bottom decile of the score is the coaching queue.
Peer-benchmarking. Drivers who see their own score against a peer group often self-correct. Any peer benchmark inherits the denominator problem: comparing a delivery driver against a commuter using the same score is comparing two different jobs.
Limits: what a driver risk score does not do
Predict a specific collision. Risk is a distribution, not a schedule.
Substitute for absolute risk. Nexar's Risk Index is a relative ranking of US regions against each other, not an absolute risk figure. The same discipline should apply to a driver-level score: a driver ranked above another is riskier in this ordering, not "twice as risky" in an absolute sense.
Read intent. A driver looking away from the road is a detection. Whether they were reading a map or checking their phone is inference.
Stay stable when the observation window changes. A driver risk score built on a short observation window is noisy. Extending the window makes it more stable, but hides recent behavior change. Any credible score publishes the window it was built on.
Frequently asked questions
What is a driver risk score?
A single number estimating how likely a given driver is to be involved in a collision, built from observed motion and behavioral signals rather than demographics.
How is a driver risk score calculated?
By combining event counts (hard braking, hard cornering, hard acceleration, verified collisions) with behavioral signals from a driver-facing camera, normalized by an exposure denominator (miles, hours, or vehicles observed).
Is a driver risk score the same as a driver safety score?
They overlap heavily. A driver risk score usually emphasizes the actuarial question ("how likely is a claim?"), while a driver safety score usually emphasizes the operational question ("who should be coached this week?"). The inputs and mechanics are often the same; the audience differs.
Why does a driver risk score need a denominator?
Because a raw event count rewards volume: more miles usually means more events, whether or not the driver is riskier per mile. Normalizing by miles, hours or vehicles is what turns a count into a comparison.
Does a driver risk score change over time?
Yes. Observed behavior is not static. Any credible score updates on a schedule (monthly is typical) so a driver whose behavior is changing is visible before the loss data admits it.
Can a driver risk score account for the roads a driver drives?
Yes, when the score is calculated in combination with a regional risk score. A driver who covers a persistently high-risk region will legitimately face more events; combining a driver-level score with a region-level risk score is what separates "risky driver" from "risky route."
Empiric Earth writes about the measured world. Numbers cited above trace to primary Nexar product documentation.