The Factor That Dare Not Speak Its Name

I. It works, or The confession first

Let us begin with the admission my colleagues would rather I saved for the end, on the theory that it finishes the argument: the thing works.

A score built from a person’s credit file predicts what that person will cost a motor insurer, and it goes on predicting after age, driving history and geography have each had their say. The Americans established this a generation ago and, to their lasting credit, in public. Monaghan laid it before the Casualty Actuarial Society in 2000. EPIC Actuaries followed in 2003 with 2.69 million earned car years drawn from all fifty states, fitted a full multivariate model, and found the score among the three most important rating factors for every coverage studied. The Federal Trade Commission, reporting to Congress in 2007 on some 1.4 million policies, found the worst-scored tenth costing nearly twice the best on property damage liability, and 1.7 times after controls. A peer-reviewed study of 175,647 Texas policies, in the North American Actuarial Journal in 2016, put the average loss at $918 in the worst decile against $558 in the best.

So the signal is real, it is incremental, and it is multivariate. The greater part of the raw effect survives the other rating factors, and anybody who tells you the score is merely postcode in a false moustache has not read the papers. One caveat belongs here and not in a footnote. Every one of those studies is American, and every one was run against the rating plans of twenty years ago, before telematics and before geography was cut as fine as it is today. What the score adds to a modern British tariff is a separate question, and I shall come to who could answer it.

I should declare where I stand, since a polemicist with an undeclared interest is only a lobbyist with a vocabulary. I have been a Fellow since 2004. I co-founded a motor broker. I buy credit scores, I price on them, and I decline customers for want of them. I have no wish to see the factor abolished. I wish to know why it has no birth certificate.

II. Nobody knows why, or The honest blank

Now the second admission, which the first is usually deployed to conceal. Nobody knows why it works.

The FTC, having had the data and the economists, wrote that “the information available does not allow the agency to draw any broad or definitive conclusions as to the reason that there is a relationship between scores and risk.” Nearly twenty years on, I know of no improvement upon that sentence. Brockett and Golden, in the Journal of Risk and Insurance in 2007, gathered the candidate explanations from psychology and biology: the man careful with his money is careful at junctions; the sensation-seeker runs up his card and his rev counter alike. Theirs remains the best map available, and it is a map drawn from other people’s literatures, with no test of mechanism in claims data. Others propose financial stress eating at attention, or maintenance deferred until the tyres are bald. Each story is plausible. None, so far as the published record shows, has been identified in the losses of individual policyholders.

In Britain the page is blanker still. I can find no peer-reviewed study, no Institute paper and no regulator’s analysis connecting British credit data to British motor losses. We have taken a result from a country with different bureaux, a different welfare state and hospital bills that land on credit files, and we have assumed that it travels. Perhaps it does. My own evidence, which follows, suggests the market is very sure that it does.

Let me be fair to ignorance, because ignorance is the respectable part of this story. A rating factor needs a loss ratio and has never needed a philosophy. Our trade has always priced what it could measure ahead of what it could explain, and an actuary who refused every correlation lacking a mechanism would have a short tariff and a shorter career. Not knowing why is no disgrace.

It does carry a duty. A factor whose mechanism is understood gives warning when the mechanism changes. A factor without one can stop working, or go on working for a different reason, and send no notice at all. The only protection is surveillance: data, replication and publication, the tedious apparatus by which a profession checks that what was true is still true. So it is worth asking what, in all its years of use, the British profession has published about this one.

Nothing that I can find.

III. A month at the window, or What the market believes

Since nobody would show me the British evidence, I went and looked at the British price.

From 27 August to 25 September 2026 I took every priced quote that reached my firm through Confused.com, some 59,000 customers, and set beside each one what the rest of the market had quoted the same person. The market’s price, throughout, is the cheapest of those quotes. I then fitted the market’s price against a single variable, the Experian score my firm knows by its catalogue code E5S052, holding fixed every other factor I rate on. The result is a curve, and here is its grade before its shape: one aggregator, one month, one firm’s model, smoothed so that it can only fall as the score rises. Anything my rivals rate on that I do not, and that happens to move with the score, will have been credited to the score. It measures prices. It says nothing whatever about losses.

The shape is a cliff with a lawn on top. A customer scoring below 600 is charged 1.40 times the market average; one scoring 1,300 or more is charged 0.91. The bottom therefore pays about one and a half times what the top pays (1.54, for those who check), with car, driver, postcode and record held where they were. Above 1,100 the curve is nearly flat, some three per cent from end to end. All the violence is at the bottom.

One complication, and an instructive one. The market does two things to a poor score: it charges more, and it stays away. Customers at the top of the scale draw quotes from about a hundred and forty insurers; those below 600 draw a hundred and seven on average, and more than a third of them see fewer than a hundred. Thin competition raises prices unaided, so some of my cliff is absence and some is loading. Where most of the panel does quote, the gap from bottom to top narrows to 1.44 times. Across all quotes it widens to 1.79. My central figure sits between the two.

Consider what that figure means. Each insurer that loads for the score arrived at its figure separately, from its own book and, I must presume, from its own British losses. They did not converge on one and a half times for sport. The British evidence on this factor therefore exists — in locked drawers, as many as a hundred and forty of them. What I have measured is the market’s belief. Whether the belief is correct I cannot tell you, because my firm is months old and has a market’s prices and not yet a book’s losses. You, reader, have the losses. You have had them for years.

IV. With and without, or The two ledgers

There are two kinds of credit data, and the difference between them is the centre of this essay.

The first is public: the electoral roll, county court judgments, insolvencies, the trail of searches. It is what the State and the courts have recorded about you. The second is private. It lives in pooled account records that lenders share with one another through the bureaux (Experian’s pool is called CAIS), and it records how you run your accounts month by month: the balances, the limits, the payment that went out late in March. The first is a register of events. The second is a diary of conduct.

The trade has published a clear account of which ledger may price your motor policy. TransUnion, marketing its public-data score to insurers in January 2023, put it in a sentence: “Unlike private credit data, public credit data can be utilised by insurance providers to enhance policy pricing.” LexisNexis says of the private kind that “the use of this data in underwriting and pricing is prohibited.” Experian’s own guidance to insurers, in 2020, gave as a condition of use that “the CAIS data should be used for making a credit decision”, meaning the decision to let a customer pay by instalments. And the rulebook of the pool itself, the Principles of Reciprocity, opens by declaring that data are shared “only for the prevention of over-commitment, bad debt, fraud and money laundering”, in the cause of responsible lending.

The trade has also published the opposite. The bureaux’ joint privacy notice, the CRAIN, says an insurer may use a person’s “financial standing and history” to help decide “how much the insurance premium should be”, and draws no line between the ledgers. Experian’s marketing of its Delphi scorecards to insurers, which promises “personalised quotes at compelling prices”, rests on a case study comparing a “Generation 9 CAIS score” with a “Generation 11 CAIS score”. Two literatures, then, and they cannot both be describing the same market.

Which one is? Here I can offer something better than exegesis, because Experian supplies my firm with two scores on every customer. One, catalogued as E5S051, is built without CAIS. The other, E5S052, is built with it. (I can find no public document describing either, so you have my word and my invoice.) Across the 2.39 million customers Confused.com sent our way between 24 August and 27 September, I have both.

The without-CAIS score is very nearly blind. Sixty-five per cent of the market sits in just two of its bands, 750 to 849, a great undifferentiated middle. The with-CAIS score takes that same middle and scatters it across its whole range: six in every hundred of those respectable people fall below 600, and at the far end a few clear 1,300. Nearly one customer in ten (9.7 per cent of the market) scores 750 or better on the public ledger and under 800 on the private one. These are people the registers find unremarkable and the diary does not.

I know what the market does with those people, because I ran the experiment by accident and with my own rates. My tariff priced on the without-CAIS score. It looked at that tenth of the market, saw respectable citizens, and gave them discounts of 23 to 31 per cent. The market looked at the same people and loaded them at 1.2 to 1.4 times the average. A tariff that finds itself so far beneath the whole market on one identifiable group should assume its rivals can see something it cannot. The something is in the diary. The proposal now on my desk is to join them.

Let me state the inference at its true grade. I have not seen inside anybody else’s rating engine, and I make no claim about how any other firm came by its number; I know only how I came by mine, which was in a lender’s coat, as I have described elsewhere. Nor can I say that the market uses this particular Experian product and not some cousin from another bureau. What the month shows is that the market’s price follows the with-CAIS score down a cliff that the without-CAIS score cannot even see. The public account says without. The tariffs say with. One of them is mistaken, and I do not think it is the tariffs.

Two further things the private ledger does, and both belong on the record. It acquits as well as convicts. Of the 7.3 per cent of the market that the public score puts in its worst band, only half are also below 600 on the private one, and the rest range up to 999. For those people the diary is the better witness, and a market confined to the register would overcharge them. That is the best argument for CAIS in pricing, and I have never seen the trade make it, since making it would require admitting the use.

And then there are the people with no diary at all. One customer in ten has no with-CAIS score, almost always because the bureau could not match them to a file. The public score files nearly all of them in its comfortable middle. The market sends them fewer insurers (eighty-six on average, on a sample of 311 priced customers that is too small to lean on), and my own firm, I must tell you, declines nearly every one. I do not know who they are. New arrivals, the young, the recently moved, the mistyped: the data cannot say, and I know of no published study that would.

V. The case for saying nothing

Let us now do the silence the courtesy of stating its case at its strongest, because it has one.

First, a rating factor’s relativities are a trade secret, and properly so. An insurer that has spent years learning how much a variable is worth owes its competitors no tutorial.

Second, nothing is being concealed in law. The privacy notices say it. Aviva’s states that it uses credit reference data to “carry out risk profiling which allows us to calculate your premium”. Confused.com’s consumer guide, updated in July 2026, says plainly that “a lower score may mean a higher premium”. A fact printed in the notice and repeated on the aggregator is a strange sort of secret.

Third, and this is the argument that is felt and seldom spoken, publicity invites prohibition. A handful of American states, California, Hawaii and Massachusetts among them, have banned credit in motor rating. A ban here would leave the risk exactly where it was and move the cost. On my own figures, half the market scores 1,000 or better and pays at or below the average. Remove the factor and those people pay more so that others may pay less, on behalf of a principle nobody has yet articulated.

Fourth, a correlation without a mechanism cannot be explained to a newspaper. “We charge people with poor credit more, and we cannot tell you why” is a true sentence and an unsurvivable headline. Better to say nothing than to say that.

Fifth, the factor belongs to the employer, and the actuary is not its press office. The Actuaries’ Code asks members to “speak up if they believe, or have reasonable cause to believe, that a course of action is unethical or is unlawful.” I do not believe the use of this factor is either. On that reading the Code asks nothing.

Every one of those points is true. Together they explain the silence perfectly. They do not excuse it.

VI. What it costs, or The answer

Take them in turn.

Confidentiality protects the relativity. It has never protected the existence of a factor, nor the class of data behind it, nor the evidence that the factor works. The Americans kept their rates and published their science: two papers through the Casualty Actuarial Society by 2003, an industry study, a state insurance department, a federal commission. No British insurer need disclose a single coefficient for the Institute to publish a pooled, anonymised answer to the only question that matters, which is what a credit score adds to a modern British tariff, with CAIS and without. That we have not done so is a choice.

The disclosure that exists is a disclosure of the wrong grain. The notice says credit data is used; it does not say which ledger. The aggregator says a lower score may mean a higher premium; it does not say one and a half times. And beside these stands a vendor’s published statement that the private ledger is prohibited in pricing, which no actuary who knows better has troubled to correct. A customer who reads everything the trade has printed will come away believing something the tariffs contradict. Under the Consumer Duty, good faith is “a standard of conduct characterised by honesty, fair and open dealing and acting consistently with the reasonable expectations of retail customers.” What does a reasonable customer expect, having been told the diary is closed?

As for prohibition, the fear is rational and the inference is wrong. America put this factor through a quarter of a century of daylight, with papers and hearings and a report to Congress, and most of the country kept it. When Washington State’s commissioner banned it in 2022, a court overturned him. The factor survived its own publication across most of a continent because it could show its working. A factor that can survive only unexamined has not been defended, and if ours is as good as the American one we lose nothing by proving it. If it is not as good, I should like to know before I pay for it again next month.

Suppose that Parliament, shown the evidence, decided the comfortable half should carry the bottom fifth. That would be a transfer between citizens, and theirs to settle. The profession has no mandate to settle it for them by keeping the numbers out of the room.

The headline problem is real, and silence makes it worse, because the argument is already under way without us. Citizens Advice reported in 2022 that drivers in areas with large minority-ethnic populations were being quoted some £280 a year more. The Financial Conduct Authority answered in December 2025 with a research note on six million policies, concluded that risk accounts for the overwhelming part of the difference, and of the unexplained remainder said that it does not “and probably can’t” know. The word credit does not appear in the note as published on the regulator’s website, so far as I can find. (If the measure of risk was the insurers’ own expected claims cost, then the score was in the room after all, uninvited and unnamed.) A national argument about who pays more for motor insurance is being conducted in terms of postcode and ethnicity, while a variable worth, on my measure, one and a half times sits in the market’s price and in nobody’s evidence. We watch from the gallery, knowing what is missing from the stage, and keep our seats.

Then there is the cost to ourselves, which ought to trouble an actuary more than any headline. We are pricing on a number whose ingredients we have not been shown. Experian told insurers in 2022 that its eleventh generation gave “14% more predictive power in finding ‘good’ customers”, these being “those who don’t make unreasonable future claims or fail to make direct-debit payments”. Read that definition twice. It means the rating factor changed when the vendor changed it. It means the target may be claims, or missed payments, or some alloy of the two, which, for the customer who pays monthly, we then load on top of an APR already priced for missed payments. A factor with no known mechanism, an undisclosed composition and a moving specification is one on which the only possible control is collective scrutiny. We have arranged to have none.

There is a legal cost in waiting, too. The FTC ran the test everybody fears and printed the answer. Scores raised predicted risk by 10 per cent on average for African Americans and 4.2 per cent for Hispanics; about one point of the former was the score standing proxy for race; the score still predicted within each group; and the Commission could not build a model that kept the prediction and closed the gap. Whatever one makes of that result, it is an answer. Nobody, to my knowledge, has published the British equivalent. If a claim comes under section 19 of the Equality Act, the insurer will have to show “a proportionate means of achieving a legitimate aim” with a variable whose recipe it has never read and a study nobody has published.

Which leaves the Code. I concede that it compels nothing; no principle is breached by an actuary who keeps a lawful factor to himself. But the same document says that it “aims to build and promote confidence in the work of actuaries”. Confidence in work that may not be inspected has another name, and the name is faith. The Institute’s own thematic review of actuaries in home and motor pricing, published in 2021, urged more attention to customer fairness and to new data and modelling techniques. I can find no mention of credit data in its published summary. A bought-in variable old enough to have its own report to Congress was not new enough to make the list, or not mentionable enough.

VII. Derek’s integer

Consider how all this looks from the far side of the counter.

Derek, whom regular readers know, has nine years’ no-claims and pays his bills on time. He therefore lives on the lawn at the top of the cliff, above 1,100 with 37 per cent of the market, where the curve is flat and the factor costs him nothing. He has never been told that he benefits from it. Nor has his neighbour three doors down, who scores under 600, been told what it costs her. Each sees one integer, the premium. The other side of the counter sees the database.

She cannot look the number up. The score Experian shows her on her telephone stops at 999; the one in my table runs past 1,300. Whatever the two have in common, they are not the same number. She cannot correct an error in a file she has not been told is being read, and she cannot dispute a loading nobody has told her was applied.

In America she would get a letter. Under the Fair Credit Reporting Act, an insurer that charges a customer more because of a credit report owes that customer a notice saying so and naming the bureau, and in Safeco v. Burr in 2007 the Supreme Court confirmed that the duty can reach the price of a brand-new policy. Here she gets a paragraph in a privacy notice and, if she goes looking, an aggregator’s guide. The Americans banned the factor in some states and told the customer in the rest, and either way they talked about it. We have found a third way, which is to use it everywhere and discuss it nowhere.

VIII. The beer mat

Four things, then, for the pricing actuary on Monday morning. None requires a regulator, a rule change or a single coefficient disclosed.

1. Say which. In every model document, every governance paper and every customer-facing account of price, write whether the credit score is built with CAIS or without. It is one word, and you already know it.
2. Publish the lift. Take it to the Institute: a pooled, anonymised study of what the score adds to a modern British tariff, on British losses, with and without. The Americans managed theirs in 2003.
3. Run the proxy test. The FTC’s method is public, and the FCA has shown that area-level ethnicity can be matched to policies. Run it on your own book and print the answer, whichever way it falls.
4. Minute the vendor. Ask what is in the score, what it was trained to predict, and what changed in the last generation. Record the reply, or the refusal.

And tell the customer when the number has cost her money. America has required it for decades.

I began by saying the factor works, and I end by saying it again, because I believe it and I shall go on paying for it. A true thing can afford to be named, and a confident profession can afford the sentence “we do not know why”. At present we manage neither. We say both to each other over the second drink, where it costs nothing and convinces nobody.

Its name, since somebody ought to print it, is E5S052. The roof, you will notice, has stayed on.