The Variable That Predicted More Than Risk#
In 2015, the Consumer Federation of America published a study that the insurance industry spent two years attempting to discredit. The research examined liability insurance premiums for identical driver profiles — same age, same driving record, same vehicle, same coverage limits — across paired zip codes within eighteen U.S. cities. In each pair, one zip code was majority-white; its neighbour was majority-Black or Hispanic. In 70% of the paired comparisons, the minority-majority zip code carried a statistically significant premium premium: higher by 19–30%, after controlling for claim frequency and vehicle theft rates. In Detroit, the differential in some pairings exceeded 40%.
The insurance industry's response was technically accurate and substantively evasive: premiums reflect actuarial risk factors, and zip code correlates with claim costs, not with race. This is true as a matter of actuarial mechanics. It is incomplete as an analysis of what zip code actually measures. Zip code does not test vehicle control. It does not audit driving behaviour. It measures, with considerable accuracy, the characteristics of the road infrastructure, traffic enforcement presence, vehicle age in the surrounding fleet, and emergency response times — all factors that determine claim costs, and all factors that correlate with race and income in American urban geography for reasons that have nothing to do with the driving behaviour of the policyholder being rated.
The pricing machine does not discriminate. It measures proxies. The proxies do the discriminating.
The Rating Variables That Proxy for What Cannot Be Priced#
Auto insurance pricing in the United States and most of Europe draws on a hierarchy of rating variables that extend well beyond driving behaviour. The Mobility Premium Burden analysis from the preceding post described the distribution of insurance cost as a share of mobility expenditure. This post examines the mechanism producing that distribution — the rating system that determines who pays what fraction.
$$MPB = \frac{\text{Annual auto insurance cost}}{\text{Annual household total mobility expenditure}} \times 100$$The variables driving the numerator — annual auto insurance cost — are the focus here. Their logic, their proxying relationships, and their interactions produce the MPB gradient that makes mandatory insurance a structurally regressive tax.
The Three Layers of Premium Determination#
Territorial Rating: The Geography of Prior Claims#
Every major auto insurer in the U.S. and Europe uses territorial rating as the primary risk classification variable. The territory — typically a zip code, postal district, or customised grid zone — is assigned a base rate that reflects aggregate claim costs within that territory over a historical period, typically three to five years. An individual driver's clean record, defensive driving certification, and 30 years of claim-free operation cannot fully override the territory's actuarial premium floor. You are, in the insurance pricing model, a statistical representative of your geography.
The actuarial logic is defensible: claim costs genuinely vary by geography, driven by traffic density, vehicle theft rates, fraud prevalence, and medical cost inflation in the local healthcare market. A driver in a high-density urban core files claims more frequently — not because of personal recklessness, but because high traffic density produces more collision exposure per mile driven. The insurer is pricing the environment, not the driver.
The problem is that the environment-as-proxy produces a pricing signal that a low-income driver cannot escape by driving carefully, taking defensive driving courses, or maintaining a spotless record. A resident of a low-income urban zip code with a median household income of $31,000, a clean 10-year driving record, and a three-year-old Honda Civic will pay a base rate set by the aggregate claim costs of all drivers in her territory — including drivers with imperfect records, older high-theft-risk vehicles, and claim fraud patterns she has no connection to. The Consumer Federation found that in Chicago, this driver would pay approximately $890 more annually than an identical driver in a high-income suburb with comparable actual risk exposure. The $890 represents her geography's actuarial charge — and her inability to exit it.
European jurisdictions vary in their regulation of territorial rating. Germany prohibits zip code as a direct rating factor, substituting regional claim statistics through a scoring system that ultimately produces similar geographic differentiation through a less transparent route. The UK's FCA 2022 pricing reform prohibited the "loyalty penalty" — the practice of charging renewing customers more than new customers — but did not materially restrict territorial rating. France's national vehicle registration database allows insurers to use municipality of registration as a rating input. The European Insurance and Occupational Pensions Authority's 2023 report on insurance pricing transparency found that territorial-based premium differentials across EU member states range from 2:1 (Germany) to 5:1 (UK), with the highest differentials concentrated in the most deprived urban territories.
Credit Scoring: The Financial History That Predicts Claims#
Approximately 47 U.S. states permit insurers to use credit-based insurance scores in premium calculation. The credit-insurance score is a proprietary algorithm derived from consumer credit bureau data — payment history, outstanding balances, credit utilisation, account age, and recent credit enquiries — recalibrated to predict insurance claim frequency and severity rather than loan default. The insurance industry argues, on the basis of proprietary actuarial studies, that credit-based insurance scores are among the most predictive rating variables available: a driver with poor credit files more claims, on average, in the aggregate dataset.
The causal pathway connecting credit history to driving behaviour is not established. The correlation exists; its mechanism is contested. One explanation is that financial stress correlates with deferred maintenance (bald tyres, worn brakes), which increases accident risk. Another is that the correlation reflects common demographic variables — age, income, neighbourhood — that credit scores proxy and that also predict claim behaviour through non-driving channels. A third is that the insurers with access to large proprietary datasets have found that credit score predicts claim costs because it is a strong proxy for everything they cannot legally price: income, race, education, and employment stability.
The practical effect is that a 25-year-old first-generation college graduate with student loan debt and a short credit history — an economically responsible person whose financial profile reflects structural disadvantage rather than personal recklessness — pays a credit-risk surcharge on their mandatory auto insurance. California, Massachusetts, and Hawaii prohibit credit-based insurance scoring. In the 47 states that allow it, the Consumer Federation estimates that the credit score surcharge adds an average of approximately $400–1,200 per year to the premiums of drivers in the bottom two credit quintiles — a transfer primarily from lower-income households to insurer revenue, grounded in a correlation whose causal mechanism is not driving behaviour.
Telematics: The Monitoring That Markets as Fairness#
Usage-based insurance programmes — Progressive Snapshot, Allstate Milewise, State Farm Drive Safe & Save — are marketed as the most equitable pricing instrument available: rather than relying on demographic proxies, they measure actual driving behaviour directly. Install a telematics device (or share smartphone GPS data), drive well, pay less. The framing positions UBI as a correction to the proxy-based discrimination of territorial and credit-based rating.
The data economics beneath the marketing are different. Telematics programmes collect continuous behavioural data — speed, acceleration, braking, cornering, time-of-day driving — that flows permanently to the insurer and can be retained beyond the insurance policy term. The discount offered — typically 5–15% for drivers deemed low-risk by the algorithm — is the transfer price for this data provision. Progressive's 10-K filings document that approximately 65% of Snapshot participants receive a discount; 35% receive a surcharge or unchanged premium. The marketing presents only the 65%.
The variables that UBI algorithms flag as risky include late-night driving (between midnight and 4:00 a.m.), hard braking (deceleration above a threshold gforce), and rapid acceleration. Each of these correlates meaningfully with dangerous driving patterns among some drivers. Each also correlates disproportionately with specific employment patterns: shift workers, healthcare workers, food delivery drivers, and security personnel drive at night and on roads with degraded maintenance that require more abrupt braking. The telematics instrument does not distinguish between a driver who accelerates rapidly because they are reckless and a driver who brakes hard because a pothole on a deteriorated road in a low-income municipality required an emergency response. The algorithm sees the g-force. It does not see the road surface.
The Pricing Machine's Output and Its Structural Beneficiary#
The compound effect of territorial rating, credit scoring, and telematics design is a premium structure that systematically places the highest MPB burdens on the households whose driving behaviour, objectively assessed, is not disproportionately risky — but whose geography, financial history, and employment patterns produce risk-proxy profiles that trigger premium surcharges. The pricing machine is not designed to discriminate. It is designed to maximise predictive accuracy within the variables it is permitted to use, producing a premium distribution whose correlation with income and race is an emergent property of the proxy architecture, not an explicit input.
The structural beneficiary of this architecture is the casualty insurance industry, which captures a mandatory revenue stream whose pricing is insulated from the normal competitive pressures of voluntary markets by the same mandate that creates it. Insurance companies recorded combined ratios — the ratio of claims and expenses to premiums collected — below 100% (net profitability) in the personal auto line in 16 of the 20 years to 2022. The mandatory-purchase framework ensures that the demand for the product does not respond to price signals in the way a voluntary product's demand would. Households that cannot afford the premium are not customers who chose a competitor; they are households who are driving uninsured — a population that is, itself, legally and financially penalised for the market failure the mandatory structure created. The final post traces what happens to MPB when the market withdraws entirely from specific geographies — and examines the feedback between climate risk, insurance retreat, and mobility dependence.
