PREDICTIVE

In most of this country you cannot legally drive without insurance, and where the state allows it, the price of that insurance is set in part by a number that does not come from your driving.

It is built from your credit report. Whether you carry a balance. How long your file has been open. What a collections agency filed about you. The insurer runs that history through a model, and the model produces a credit-based insurance score, and the score helps decide what you pay to do a thing the state requires you to do.

In 2003, Section 215 of the FACT Act ordered the Federal Trade Commission and the Federal Reserve Board to study what these scores do to “the availability and affordability of credit and insurance,” measured “by geography, income, ethnicity, race, color, religion, national origin, age, sex, marital status, and creed,” and to determine whether scoring “could result in negative or differential treatment of protected classes under the Equal Credit Opportunity Act.”

It ordered one thing more. Find out whether the same results could be had “through the use of factors with less negative impact.”

Two of the study’s questions, then. What does this do to Black people, and can it be done with less harm. The statute required both agencies to consult the Office of Fair Housing and Equal Opportunity.

The FTC did the work. It published the answer in 2007, and the answer has been sitting in public ever since.

What the government measured

The report is two hundred and forty-two pages. The measurement is one sentence, in the conclusion.

“The FTC’s analysis revealed that the use of scores for consumers whose information was included in the FTC’s database caused the average predicted risk for African Americans and Hispanics to increase by 10% and 4.2%, respectively, while it caused the average predicted risk for non-Hispanic whites and Asians to decrease by 1.6% and 4.9%, respectively.”

Turn on the scoring and Black drivers are assigned a tenth more risk. White drivers are assigned less.

The Commission did not leave the consequence to the reader. The next sentence: “These changes in predicted risk are likely to have an effect on the insurance premiums that these groups on average pay.”

And the distribution underneath it, from the body of the report: “more than one-half of all African Americans have credit scores in the lowest quarter of the overall score distribution.”

More than half. The lowest quarter.

That is the federal government, under a mandate from Congress and on data collected in 2000 and 2001, reporting that a private pricing instrument raises the predicted risk attached to a legally required product for Black drivers and lowers it for white drivers, and that the change is “likely to have an effect on the insurance premiums that these groups on average pay.” It is a measurement, taken by an agency that was told to take it.

The word

Here is what else the same report concluded. Its conclusion opens: “The FTC’s analysis demonstrates that credit-based insurance scores are effective predictors of risk under automobile insurance policies. Using scores is likely to make the price of insurance conform more closely to the risk of loss that consumers pose, resulting, on average, in higher-risk consumers paying higher premiums and lower-risk consumers paying lower premiums.”

And on the proxy question:

“Scores have only a small effect as a ‘proxy’ for membership in racial and ethnic groups in estimating of insurance risk, remaining strong predictors of risk when controls for race, ethnicity and income are included in risk models.”

Read them beside the measurement above, because they are all true and all in the same document. The scoring raises the risk Black drivers are assigned. And the scoring is not merely a mask for race, because it still sorts risk inside every racial group.

Now read the mandate again. The Commission read it as asking whether the scores act as a proxy for race. On the proxy question its answer was a small effect: with controls for race, ethnicity and income in the models, the rise in average predicted risk for African Americans was 8.9% instead of 10%.

The word doing the work is predictive. Not fair. Not justified. Not right. Predictive.

The sentence they printed anyway

The Commission did not hide it. It printed it third in its own conclusion.

“It has not been clearly established why scores are predictive of risk.”

The government did not establish why it works.

It knew that the number correlates. It tested one explanation, that people with low scores spend less on maintaining their cars, and found the result “weakly consistent,” with maintenance “not the primary cause.” The body of the report says the information available “does not allow the agency to draw any broad or definitive explanations” of why.

So predictive is not an explanation. It is a description of a correlation whose cause the federal government said it could not establish.

The models

This was the second of those two questions, and the Commission did the work. It built two race-neutral models, one that controlled directly for race, ethnicity and income and one built only on non-Hispanic whites.

The result, in the Commission’s words: “although the race neutral models that the FTC built accurately predict risk, they do not decrease the differences in credit-based insurance scores among racial and ethnic groups.”

Then it built a model that discounted the variables with large racial differences. That one did narrow the gap: the share of African Americans in the bottom tenth of scores fell from 27% to 14%. The model was “much less predictive of risk,” and it “provides far less effective risk prediction.” The Commission “was not able to develop” a model that did both. It called that result “by no means definitive,” and said it “strongly suggests that there is no readily available approach for doing so.”

Controlling for race did not close the gap. The variables that show large differences across racial and ethnic groups, the report says, are “those relating to payment history (e.g., delinquencies) and public records, and the variable for the share of accounts with high balances relative to credit limits.” The gap is in the credit histories the model reads. This archive has spent four casefiles documenting the country those histories were written in: the claim denied in Greenwood, the corridor cut through Black Bottom, the map that graded a neighborhood red, the word that condemned Mill Creek Valley. All of it lands in a credit file.

Correction

Corrected 2026-10-01. This essay published on July 22, 2026. The errors are named here, in ten entries, rather than quietly repaired.

The proxy question was called the Commission’s own. The essay said “Congress did not ask whether the scores were a mask for race,” that “The proxy question was its own,” and that “That unasked answer is what let the first finding stand”; the docket said “Section 215 does not ask whether scores act as a proxy for race.” The report reads the mandate otherwise: “The mandated study further must evaluate whether scores act as a proxy for membership in racial, ethnic, and other protected classes,” and Section 215(a)(2) orders “a multivariate analysis that controls for prohibited factors under the Equal Credit Opportunity Act and other known risk factors.” The essay’s own closing section already said “The FTC was asked whether the score acts as a proxy for race.” The paragraph beginning “Now read the mandate again” now says the Commission read the mandate as asking whether scores act as a proxy, and gives the size of the proxy effect, and the docket now says Section 215(a) lists four study items and quotes from two of them.

The alternative models were said to have left the gap untouched. The essay said “They took race out of the model. The gap stayed,” and “If you can remove every racial variable and the racial gap survives untouched, then the gap was never in the model.” The sentence it quoted covers only the two race-neutral models, one of which controlled directly for race, ethnicity and income. A third model, which discounted the variables with large racial differences, “did substantially reduce the differences”: 14% of African Americans in the bottom 10% of scores, against 27% in the baseline model, at a large cost in accuracy. The section now gives all three results.

A rise in predicted risk was stated as a rise in what Black drivers are charged. The essay said the instrument “moves the cost of a legally required product up for Black people,” that “The scoring raises what Black drivers are charged,” and, in the subtitle, that Congress asked “whether credit scoring charges Black drivers more.” The report measured average predicted risk and said the changes “are likely to have an effect on the insurance premiums that these groups on average pay.” The body now says predicted risk and quotes the report on premiums; the subtitle now says Congress asked what credit scoring does to Black drivers.

A group average was stated as what happens to you. The essay closed a section on “The neutral score charges you more” and “You drive to work. Your file is clean. You pay more to get there.” The 10% is the change in “the average predicted risk for African Americans.” The baseline model is “a race ‘blind’ model” built from credit-file variables, and the report warns that group-level attributes “do not necessarily apply to each consumer with a particular level of score.” Both lines are cut.

“Nobody can” explain the link. The essay said “It cannot tell you what a missed payment has to do with a car crash. Nobody can,” and that the government “explicitly declined to claim it understood” the cause. The report tested one explanation, vehicle maintenance, and found its results “weakly consistent with the hypothesis that some of the relationship is attributable to consumers with lower scores spending less to maintain their vehicles.” Its words are that the information available “does not allow the agency to draw any broad or definitive explanations.” The passage now says so, and the sentence calling the correlation “the legal justification for charging Black drivers more” is cut.

The report’s lead finding was left out, and the hook denied it. The essay quoted the third sentence of the report’s conclusion but not the first two, which say the analysis “demonstrates that credit-based insurance scores are effective predictors of risk” and that using them “is likely to make the price of insurance conform more closely to the risk of loss.” The hook line and the body said “A prediction that cannot say why it predicts is not a finding.” The report holds the prediction established and its cause unexplained. The first two sentences are now quoted, the body sentence is cut, and the hook line now reads “The government measured the risk the score assigns Black drivers, and printed that it could not say why scores work.”

A number with “nothing to do with your driving.” The opening said the price is set “by a number that has nothing to do with your driving,” and the excerpt “Your credit report sets the price of your car insurance. Not your driving.” The report found the scores “effective predictors of risk under automobile insurance policies,” and the essay’s own body says the price is set “in part.” The opening now says that, where the state allows it, the price is set in part by a number that does not come from your driving, and the excerpt that, where states allow it, the credit report helps set the price.

A notice duty stated wider than the statute. The service step said “An insurer that acts on a consumer report owes you notice of it,” and the box said of the score “They are required to tell you.” Under 15 U.S.C. 1681m(a), the duty arises when an insurer “takes any adverse action” based on a consumer report, which for insurance includes “an increase in any charge for” coverage. The score disclosure in that section covers a “credit score” as defined for lenders, and an insurance score is not clearly within it. Both lines now say what the statute requires.

Claims about how the practice is defended had no source. The essay said “That result is usually read as a defense of scoring, and the industry has read it that way for nineteen years”; “The score is permitted because it predicts, and prediction is treated as its own defense, a thing that needs no further account of itself. A number that forecasts is a number that has earned its keep”; “Solely. The practice needed nothing more than that”; in the subtitle, “Then it printed a word, and the word held”; and in the excerpt, “and let the practice stand.” No source in the file supports them, and the report itself records that “Fifteen states prohibit certain uses of credit history information” and that Hawaii bans its use by statute. Each is cut; the subtitle and the excerpt now end on the report’s unexplained cause.

The sculpture was called marble. The image description called “Man Controlling Trade” “the marble sculpture.” The General Services Administration’s Fine Arts Collection lists its medium as limestone. The description also said the man grips “the bridle”; the photograph shows his hand on the horse’s head. Both are corrected.

Smaller changes made at the same time:

• “ordered the Federal Trade Commission to study” now reads “ordered the Federal Trade Commission and the Federal Reserve Board to study”; Section 215(a) names both, and “required the Commission to consult” now reads “required both agencies to consult.”
• “Two questions, then” now reads “Two of the study’s questions, then”; Section 215(a) lists four. “What does this cost Black people” now reads “What does this do to Black people.”
• “Congress had a suspicion about this. In 2003 it wrote the suspicion into law.” is cut; the paragraph now opens “In 2003, Section 215 of the FACT Act ordered.”
• “What a collections agency filed about you four years ago” loses “four years ago.”
• “The finding is one sentence, in the conclusion” now reads “The measurement is one sentence”; the conclusion’s lead finding is quoted in the next section. The measurement paragraph now says the data were collected in 2000 and 2001, and “taken by the agency” now reads “taken by an agency.” “Read those two findings side by side, because they are both true and they are both in the same document” now reads “Read them beside the measurement above, because they are all true and all in the same document.”
• “Turn on the scoring and Black drivers get riskier by a tenth. White drivers get safer.” now reads “Turn on the scoring and Black drivers are assigned a tenth more risk. White drivers are assigned less.”
• “And the distribution underneath it, also the Commission’s own words” now reads “from the body of the report”; the quotation is from Part VI, not the conclusion.
• “The government does not know why it works.” now reads “The government did not establish why it works,” and “It knows that the number correlates” now reads “It knew that the number correlates.” The phrase “in the report that let the practice continue” is cut.
• “The models that did not help” is now “The models,” “This was Congress’s second question” now reads “This was the second of those two questions,” and the section now quotes the report on which variables differ by race. “The gap is in the country the model is measuring” now reads “The gap is in the credit histories the model reads,” and “documenting that country” now reads “documenting the country those histories were written in.”
• The section “Nineteen years” is cut: “The report came out in July 2007. Nothing in it required anyone to stop.” and “The FTC was asked whether the score acts as a proxy for race. Its answer, in its own words: scores ‘do not act solely as a proxy for these characteristics.’” The proxy answer and its size are given above.
• In the box, “PREDICTIVE shows a word holding a price in place. One hides the number. The other justifies it.” now reads “PREDICTIVE shows a word standing where a reason would go.”
• The service section’s opening said “The number that set your premium is not secret. It is disclosable.” It now says the report behind the number is free to read. “Pull the credit report the score was built from” now reads “Pull your credit reports”; the supplying agency may not be one of the three nationwide ones. “The sentence conceding the mechanism is unexplained,” in the service section and the docket, now reads “the sentence stating that why scores predict risk has not been clearly established.”
• The Word of the Week entry for PREDICTIVE (week of August 17, 2026) said “Congress asked two questions in 2003,” “Nobody knows what it measures,” and that the price “is still set, in part, by that.” It now follows the corrected essay.
• The FTC source record said the alternative models left the score gap unclosed, and that Congress ordered the study “to determine whether credit-based insurance scores act as a proxy for race.” It now gives the discounted model’s result and the Commission’s reading of the mandate.

Sources: Federal Trade Commission, “Credit-Based Insurance Scores: Impacts on Consumers of Automobile Insurance,” Report to Congress, July 2007, including Appendix A (text of Section 215). 15 U.S.C. 1681a, 1681g and 1681m. General Services Administration Fine Arts Collection, “Man Controlling Trade.” Each read on October 1, 2026.