From Calendars to Sensors: The Unfinished Science of Driving Exposure

February 2026  ·  Long Read

What Insurance Has Always Needed — and Rarely Had

Every insurance premium is, at its core, a fraction. The numerator is loss — claims paid, costs incurred, the actuarial expectation of harm. The denominator is exposure — the unit of time, activity, or presence against which that loss is measured. Get the numerator wrong and you misprice risk. Get the denominator wrong and you misprice it just as badly, only more invisibly, because the error is baked into the foundation of every calculation that follows.¹

Auto insurance has been getting the denominator wrong for most of its history. Not out of negligence, and not for lack of trying — but because the data that would have enabled a better answer simply didn't exist.¹ Each successive approximation of exposure in auto insurance tells the same story: the industry reached not for the most theoretically sound unit, but for the most practically available one.²

The Borrowed Convention: Time as the Default

The first auto insurance policies in the United States were issued around the turn of the twentieth century, adapted almost wholesale from the frameworks that insurers had spent decades developing for fire, marine, and general liability coverage.¹˒³˒⁴ The earliest documented personal auto policy in the country is commonly attributed to Gilbert Loomis of Dayton, Ohio, who in 1897 paid the Travelers Insurance Company for a liability policy covering a horseless carriage he had assembled himself in a local machine shop.⁵˒⁶˒⁷

The policy period — typically one year — was not an actuarial finding. It was a convention, inherited from insurance products where it made reasonable sense: a building stands exposed to fire for a year; a ship completes a voyage in a bounded time; a business accumulates liability over a fiscal cycle.¹˒³˒⁴˒⁸ When the automobile arrived, there was no established alternative framework, and so the annual policy followed the car almost automatically, as if by gravitational pull. The premium was set, the term was set, and the driver and insurer agreed, implicitly and somewhat arbitrarily, that twelve calendar months was the relevant unit of account.¹˒⁴

This was not wholly unreasonable in an era when automobile ownership was rare, roads were poor, and speeds were modest.⁴˒⁹ A Model T owner in 1910 drove perhaps a few hundred miles a year, and the distinction between a driver who used his car occasionally and one who used it constantly was not yet an actuarially significant one.⁴˒⁹ But as car ownership spread, highways improved, and Americans began building their entire lives around the automobile, the policy period became an increasingly poor proxy for the thing it was supposed to represent.⁴ Two drivers on identical annual premiums might accumulate vastly different road time — one a retiree making weekly grocery trips, another a salesman covering three states. The policy period held them as actuarially equivalent.

What the industry needed was something that tracked actual exposure to the hazards of driving, rather than merely the passage of calendar time. For decades, it had no practical way to get it.¹⁰

The Odometer Speaks: Mileage Enters the Rating Manual

The shift toward mileage as an exposure variable was not a single regulatory moment but a slow accumulation of actuarial instinct and practical availability.¹⁰˒¹¹ The intuition was straightforward: a car that spends more time in motion is more likely to be involved in a collision than one that sits in the driveway.¹¹ Distance driven is, at minimum, a more direct measure of contact with the hazards of the road than the calendar.¹¹˒¹²

Mileage also had a natural analog in the existing culture of car ownership. From the earliest decades of mass motoring, drivers tracked odometer readings not out of actuarial interest but out of mechanical necessity: oil changes every few thousand miles, timing belt replacements at set mileage thresholds, and similar maintenance schedules.¹³ Tire rotations, transmission flushes, spark plug checks — the entire rhythm of vehicle maintenance was organized around units of distance, not units of time.¹³ This was not universally true across machines: construction equipment and industrial engines, which may sit idle for long periods but run at high load when operating, have historically been maintained and depreciated by engine hours rather than distance.¹³ A bulldozer that moves a hundred yards in an hour accumulates wear very differently than one covering the same distance on a highway. But the personal automobile, designed for sustained road travel, had settled on distance as its native unit — and every dealer service visit reinforced that convention by recording the odometer reading as a matter of routine.¹³

That recorded mileage turned out to be enormously valuable, and not only to mechanics. Over the course of several decades, annual mileage gradually entered actuarial rating frameworks as a primary exposure variable — the first with direct operational meaning beyond the calendar.¹⁰˒¹¹˒¹⁴˒¹⁵˒¹⁶˒¹⁷˒¹⁸ The system that developed around it was crude, imperfect, and widely understood to understate actual driving. But it was a genuine conceptual advance over the policy period alone.¹⁰˒¹¹˒¹⁴˒¹⁵˒¹⁶˒¹⁸

The Political Turn: California Makes Mileage Mandatory

The most consequential regulatory intervention in the history of auto insurance exposure came not from a federal agency or an actuarial working group but from California voters, who in November 1988 passed Proposition 103 by a margin of fifty-one to forty-nine percent.¹⁹ The initiative was primarily aimed at rolling back premium increases and subjecting auto insurers to rate regulation.¹⁹ But buried in its rating factor provisions was a requirement that would quietly reshape the industry's thinking about exposure for decades: Proposition 103 mandated that California insurers rank miles driven annually as one of the three primary factors in determining premiums, alongside the driver's safety record and years of driving experience.¹⁹ All other factors — the demographic proxies that had traditionally dominated personal auto rating — were subordinated to these three.¹⁹

The implications were significant. California, the nation's largest auto insurance market, had legally encoded the actuarial primacy of mileage over demographic variables.¹⁹ The law formalized an argument that had been developing in actuarial literature for decades: that how much a person drives is more causally connected to the probability of a claim than who they are or where they live.¹⁰˒¹¹˒¹⁸˒²⁰ For carriers writing business in the state, the question of how to verify mileage shifted from an internal data quality concern to a matter of regulatory compliance.¹⁹˒²⁰

California's mandate also accelerated a broader policy conversation about pay-as-you-drive insurance that had been developing in academic and environmental circles throughout the 1990s and 2000s.²⁰ Researchers at the Brookings Institution and the Victoria Transport Policy Institute began publishing analyses arguing that per-mile insurance pricing would not only improve actuarial accuracy but reduce total vehicle miles traveled — and with it, congestion, emissions, and accident rates.²⁰ The argument was elegant: if the marginal cost of driving an additional mile included a small increment of insurance cost, drivers would internalize some of the externalities their driving imposed on others.²⁰ This framing gave mileage-based insurance a second life as an environmental and transportation policy instrument, not merely an actuarial one — a framing that would eventually surface in state-level PAYD legislation and federal transportation planning discussions in the 2000s.²⁰˒²¹

The Promise of Resolution

What the odometer improved, it also limited. Mileage differentiated drivers by how much they drove, rather than treating all annual policyholders as equivalent. But it purchased that improvement at the cost of collapsing everything else about driving into a single scalar.¹⁰˒¹¹

Ten thousand miles is ten thousand miles. The odometer does not record whether those miles were accumulated on a rural interstate at two in the afternoon or on an urban arterial at midnight. It does not record whether the driver moved through light traffic or dense congestion, whether the roads were dry or iced, whether the trip was a five-minute school run or a four-hour highway crossing. The odometer produces a cumulative count, and cumulative counts are by their nature indifferent to the texture of what they measure.¹⁰˒¹¹

When sensors became cheap enough to embed in vehicles and smartphones became capable enough to serve as crude driving recorders, insurers found themselves with access to something that earlier exposure metrics could never provide: a multidimensional portrait of the driving environment itself.²²˒²³ A modern telematics dataset contains not just distance or duration but information about where a driver went, when, at what speeds, through what density of intersections, in what weather conditions, and with what kinematic signature — how they accelerated, braked, and changed direction.²²˒²³˒²⁴ The exposure picture that emerges from this data is richer than any single number can replicate. A mile of late-night urban driving near a high-frequency collision corridor is not the same exposure as a mile of midday rural highway travel, and telematics, for the first time, gives actuaries the raw material to begin treating them differently.²²˒²³˒²⁴

The potential is enormous. In theory, telematics could characterize the conditions surrounding a driver at a resolution of seconds — both the environment outside the vehicle and the environment within it.²²˒²⁴ The road itself, the conditions on it, the behavior of other road users, and the capabilities and condition of the vehicle itself are all dimensions of exposure that vary continuously and interact in ways that a single number cannot capture.²²˒²⁴ An accident, after all, can spin out of control in a fraction of a second. The denominator that insurance has always needed is one that can account for that kind of granularity: not merely how far or how long, but the full context of what a driver was exposed to — the conditions they drove through and the vehicle they drove through them in.¹¹˒²²˒²⁴

The Network That Doesn't Exist

But knowing what the ideal denominator looks like and being able to construct it are two different problems, and the gap between them is not primarily technical. It is organizational, economic, and political.²⁴˒²⁵

To characterize driving risk at that resolution — road segment by road segment, second by second — would require an infrastructure of continuous, high-frequency data collection at national scale.²⁴ Traffic flow and density would need to be observed in real time.²⁴ Road hazards, construction zones, and weather conditions would need to be mapped and updated continuously.²⁴ Vehicle interactions, near-misses, and behavioral patterns would need to be aggregated across millions of trips to produce statistically credible risk profiles for any given corridor.²⁴˒²⁵ No single insurer, no single automaker, and no single telematics startup has managed to build that network.²⁴ Not for lack of ambition, but because the problem is one of coordination, not computation.²⁵

The difficulty starts with the fact that the entities best positioned to collect the necessary data all want it for different reasons. Automakers see connected vehicle data as a competitive asset — a way to differentiate brands, retain customers through subscription services, and build ecosystems that make switching costly.²⁵ Insurance carriers, as industry analyses have documented, want the data to price policies more precisely and to address the well-known problem that voluntary telematics programs tend to attract drivers who already believe they are low-risk — a self-selection dynamic that limits the actuarial value of the resulting datasets.²²˒²³ Retailers and commercial real estate developers want aggregate traffic and footfall data to evaluate potential business sites.²⁵ Municipal governments want it for transportation planning.²⁵ Fleet operators want it for logistics optimization.²⁵ Each of these stakeholders has a legitimate use case, and none of them has a natural incentive to share the raw data with the others in the form or at the resolution that would be required to build the kind of comprehensive risk map that telematics, in its ideal form, promises.²⁴˒²⁵

This is not a novel coordination problem — it is the same one that has stalled or slowed every attempt to build shared data infrastructure across competing commercial interests, from electronic health records to credit bureaus to open banking.²˒²⁶ But in the automotive context, the fragmentation is compounded by the sheer heterogeneity of the players involved.²⁴ OEMs operate on product cycles measured in years. Insurers operate on policy cycles measured in months.²⁵ Telematics startups — the companies that were supposed to bridge the gap — operate on venture capital runway measured in quarters.²⁵ And when the capital markets turned hostile, many of those startups simply ran out of time.²⁷˒²⁸

The Insurtech Correction

The most vivid illustration of this fragility came in 2023, when the collapse of Silicon Valley Bank sent shockwaves through the venture-backed technology sector.²⁷ SVB had been the dominant lender and banking partner for technology startups across every vertical, including the insurtech companies that had raised hundreds of millions of dollars to build telematics-driven insurance products.²⁷ When the bank failed in March 2023 — the largest bank failure in the United States since 2008 — the aftershocks rippled through the startup ecosystem with brutal efficiency.²⁷ Companies that had deposited operating capital at SVB scrambled for liquidity.²⁷ Fundraising, already difficult in a rising-rate environment, became dramatically harder as venture capital firms pulled back.²⁷

The insurtech sector had already been cooling before SVB's collapse. Global insurtech funding fell forty-five percent in 2023 to $4.6 billion, its lowest level since 2017.²⁸ But the banking crisis accelerated a reckoning that had been building for years: many of the telematics-first insurance companies that had raised capital during the zero-interest-rate era could not demonstrate a viable path to profitability.²⁸˒²⁹ Metromile, the company that had pioneered consumer pay-per-mile auto insurance in 2012, went public via SPAC in 2021 at a valuation above one billion dollars, only to see its stock collapse before being acquired by Lemonade in 2022 for a fraction of its former valuation.³⁰ Root Insurance, which had built its entire underwriting model around smartphone-based telematics scoring, saw its market capitalization fall more than ninety percent from its IPO high.³¹ The broader pattern, visible in public filings and press coverage, was unmistakable: the market appeared to have concluded that collecting driving data was considerably easier than turning it into a sustainable insurance business.²⁸˒²⁹

The companies that survived the correction, as industry observers noted, tended to be those embedded within larger carriers — Progressive's Snapshot, State Farm's Drive Safe & Save, Allstate's Drivewise — where telematics was a feature of an existing underwriting operation, not the sole basis for a standalone company.²²˒³² The lesson, as several post-mortems noted, was less about the technology itself than about the business model surrounding it: building a telematics data platform appears to require patient capital, regulatory expertise across dozens of state jurisdictions, and a willingness to operate at a loss for years while the data accumulates.²²˒²⁹ Venture capital, with its expectation of rapid growth and near-term exits, was a poor match for that timeline.²⁹

The Regulatory Headwind

Even as the capital markets were sorting the survivors from the casualties, a second and potentially more consequential force was reshaping the landscape. Across the country, the regulatory environment around vehicle data collection, privacy, and usage transparency has been tightening — unevenly, but unmistakably.²¹˒³³˒³⁴˒³⁵˒³⁶˒³⁷˒³⁸˒³⁹˒⁴⁰˒⁴¹˒⁴²˒⁴³˒⁴⁴

The United States has no single federal framework governing connected vehicle data.³³˒³⁴˒⁴¹ What exists instead is a growing and heterogeneous body of state privacy statutes, federal enforcement actions, industry guidelines, and proposed legislation — each with its own definitions, consent requirements, and enforcement mechanisms, and none of them fully compatible with the others.²¹˒³³˒³⁴˒⁴²˒⁴³˒⁴⁴ The result is that any attempt to aggregate telematics data at national scale must contend not with one set of rules but with dozens, and the rules are still being written.³³˒³⁴˒⁴²˒⁴³˒⁴⁴

For the project of building a better denominator, this matters directly. The value of a high-resolution driving risk model increases with the density and breadth of the data that feeds it.²²˒²⁴ Estimating exposure at the resolution of seconds or subseconds — the timescale at which an accident actually unfolds — requires continuous, high-frequency data collection across a vast and geographically diverse road network.²²˒²⁴ But the tightening of data privacy and usage transparency regulation is making it progressively more difficult to build and sustain that kind of infrastructure at the scale necessary to make it actuarially meaningful.²²˒²⁴˒³³˒³⁴˒⁴²˒⁴³ The technology exists. The legal and institutional conditions under which it could be deployed nationwide, at this point, do not.²²˒²⁴˒³³˒³⁴

The Denominator, Still Unresolved

The industry spent a century reaching for what was most available: first the calendar, then the odometer, then the app.¹˒¹¹˒¹⁸˒²² At each step, the new measure improved on the old by capturing some dimension of driving exposure that its predecessor had missed.¹¹˒¹⁸˒²²˒²⁴ And at each step, the improvement revealed how much texture remained uncaptured — how much of the actual risk environment was being collapsed into a number that, however better than its predecessor, still fell short of the thing it was trying to represent.¹¹˒¹⁸˒²⁴˒⁴⁵

Telematics has not closed that gap.¹¹˒²² What it has done is make the gap visible in a way it never was before, by revealing just how much information about the driving environment can now be captured and how little of it, according to published research, has yet been incorporated into standard ratemaking.²²˒²⁴˒⁴⁵ The data is out there — in the vehicles, in the phones, in the infrastructure — but the network to collect it at national scale, the coalition to share it across competing interests, the legal framework to govern it across fifty states, and the capital to sustain the companies that would build it all remain, for now, works in progress.²⁴˒²⁵˒²⁸˒³³

The denominator that auto insurance needs — a measurement of exposure that reflects the actual texture of driving risk, not merely its volume — is no longer a theoretical abstraction.¹¹˒¹⁸˒²⁴˒⁴⁵ It is a known quantity, visible at the edges of what the current data infrastructure can deliver.²²˒²⁴ Whether the industry can assemble the organizational, legal, and financial machinery to reach it is another question entirely.²⁴˒²⁵˒²⁸ History suggests it will get there, eventually, the same way it has gotten to every previous improvement: not in a single leap, but through the slow, uneven accumulation of what becomes practically possible.¹˒³˒⁴˒⁸˒⁴⁵