Your credit score can’t tell if you can pay

Pay down your credit cards and your credit score can go down. Pay off your car and it can go down again. Someone with years of savings and no debt can score worse than someone living on a credit card. None of that is a bug. A credit score answers one narrow question about how you behave as a borrower, and we’ve spent more than thirty years treating it like the answer to a much bigger one.

It doesn’t read what you have

A FICO score reads your credit report and nothing else. FICO says so plainly. Your income, salary, employer and job history are not part of the score. Neither is your checking account, your brokerage account, your 401(k) or the equity in your house. None of that is on a credit report, so to the model none of it exists. VantageScore works from the same reports and has the same blind spot.

What’s left gets split five ways. Payment history is about 35 percent. Amounts owed is about 30. Length of credit history 15, new credit 10, credit mix 10. Look for a single line about whether someone can actually pay. There isn’t one. Every slice is about how they behave as a borrower.

So a person with $2 million in cash and no debt and a person with $200 in checking and $9,000 across three cards can get the same score, and the second one can get the better one. The model isn’t confused about this. It was never asked.

You don’t even get a score unless you borrow. FICO’s minimum scoring criteria need at least one account open for six months and at least one account reported in the last six months. Pay cash for everything for a few years and you become unscorable. The CFPB’s 2025 update put people with unscored credit records at 9.8 percent of adults in 2020, 25.4 million people. Some of them are broke. Some of them just don’t borrow. The score can’t tell the difference and doesn’t try.

It can punish you for paying things off

Zero can be worse than almost zero. In FICO 8, the version behind most card-issuer scores, if every revolving account reports a $0 balance you lose points compared to reporting a tiny one. People who track this closely on FICO’s own forums treat a 20 to 30 point drop as normal when every card reports zero, and one person watched an 850 fall 14 points after paying off the last balance. $1 on one card counts as 1 percent utilization, and 1 percent scores better than zero. Experian’s numbers show where the top of the scale actually lives. People with exceptional scores average about 7 percent utilization, not zero.

Paying off your last loan costs points too. FICO says this directly. An installment loan that’s mostly paid down is less risky, in their data, than no open installment loan at all. So paying off your car can score worse than nearly paying it off.

And the two big models don’t agree on what good behavior is. VantageScore’s analytics lead has said that a zero balance doesn’t hurt you under their model. Same person, same behavior, rewarded by one score and docked by the other.

Here’s why a model ends up like this. It’s trained on millions of credit files to predict who goes 90 days late. In that population, people with no activity are a mixed bag. Some are wealthy and don’t need credit. Some just had their cards cut, or stopped using them because things went wrong. Someone carrying a small balance and paying every month is a well understood, profitable, reliable customer. The model learns that pattern correctly. Then it applies a population average to an individual it can’t see.

That’s the core flaw, and it’s a statistical one. A score is a decent answer to “how do people whose credit files look like yours behave?” It’s a bad answer to “is this person good for the money?” And everyone uses it to answer the second question.

It can’t explain itself

When a score moves, you’re supposed to get reasons. Lenders that turn you down owe you the key factors, and score apps show a list of what changed. But that list shows what changed in your file, not what the model did with it. The model is proprietary, the factors are coarse, and the result is that a drop can arrive with an explanation like “balances decreased” or “payment history improved.” Those are both good news. The app is telling you what happened. It can’t tell you why that cost points, because it doesn’t know either.

A measurement that moves and can’t say why isn’t something people should be making decisions around. Millions of people do anyway, because the alternative is not knowing how the number they’re judged by works.

It has no memory, so it can be gamed

Classic FICO looks at a snapshot. Whatever balance your card issuer reported on your statement date is what you owe, as far as the score knows. It doesn’t matter that you pay in full every month. If you charge a vacation and the statement closes before you pay it, your utilization spikes. Pay it a week earlier and the spike never happened.

So there’s a whole hobby around it. Let exactly one card report a small balance and the rest report zero. Time payments to statement dates. Don’t pay off your last loan before you apply for a mortgage. Get added as an authorized user on someone else’s old card. None of these make anyone a better borrower. They make them a better player of the score. A measurement that can move 20 or 30 points in a month without anything real changing is mostly measuring how well you play it.

Newer models do better. FICO 10T uses trended data, so it can see that you pay in full every month instead of guessing from one day. But lenders choose the version, and adoption is slow.

There isn’t one score, there are dozens

When people say “my credit score” they mean whichever number an app showed them. Lenders often aren’t using that number. There are three bureaus, each with slightly different data, and many FICO versions plus VantageScore, plus industry versions for auto loans and cards. For years mortgages ran on “Classic FICO,” models built in the early 2000s. The FHFA only opened Fannie and Freddie loans to VantageScore 4.0 in July 2025.

So the number you watch move isn’t necessarily the one that will be used on you, and the one that will be used on you may come from a model older than the iPhone.

The data underneath is often wrong

All of this sits on top of credit reports, and the reports aren’t clean. The FTC’s congressionally mandated accuracy study found one in five consumers had an error on at least one of their three reports, and 5 percent had an error serious enough that they could pay more for credit. Five percent of American adults is millions of people paying a tax for a data entry mistake they didn’t make.

The complaints say the same thing at a much bigger scale. The CFPB received about 155,000 complaints about credit or consumer reporting in 2019. In 2025 it received 5.8 million, 88 percent of every complaint the Bureau got that year. Most name Equifax, Experian or TransUnion, and most describe information on a report that the person says is wrong. The Bureau says part of the growth is credit repair companies, social media influencers and AI tools flooding the system with duplicate disputes, so the count isn’t a clean measure of errors. That’s its own indictment. A system where the fastest way to fix your score is to dispute everything on your report is not a system anyone trusts.

Medical debt is the clearest example of the system choosing to keep noise in. The CFPB finalized a rule in January 2025 to take medical bills off credit reports, about $49 billion of it across 15 million people. In July 2025 a federal judge vacated it, at the joint request of the CFPB and the industry groups suing it. A billing dispute with a hospital is still allowed to say something about whether you’ll repay a car loan.

And then we used it for everything else

If the score stayed in the lending department, I’d call it a flawed tool doing a narrow job. It didn’t stay there.

Car insurers use credit-based insurance scores to set premiums. Consumer Reports looked at more than 2 billion price quotes in 2015 and found that in most states insurers treated poor credit worse than a drunk driving conviction. In Florida, single adult drivers with a clean record and poor credit paid $1,552 a year more on average than the same drivers with excellent credit and a DUI. Only California, Hawaii and Massachusetts banned the practice at the time. Today’s gap is just as wide. The Zebra’s state-by-state comparison puts the premium for very poor credit, with the same clean record, at 74 percent above exceptional credit in North Carolina and 285 percent above it in Minnesota.

Think about what that means. The person who has never had an accident pays more than the person who drove drunk, because one of them carried a card balance in a bad year. Insurers will say credit predicts claims, and statistically it probably does. That’s the same argument as above. A population pattern, applied to an individual, priced as if it were about them.

Landlords pull credit. Some employers pull credit reports. Utility companies ask for deposits based on it. A number designed to answer one question for one industry has become a general rating of whether you’re a responsible adult, and it was never built to answer that.

What should change

Count assets. The data exists. Bank account data is already used by some lenders and by opt-in products like UltraFICO and Experian Boost. A person with years of expenses in cash is not a default risk, and any model that can’t see that is missing the most predictive fact there is. Make it standard, not a side product you have to know to opt into.

Kill the zero penalty. If a model docks people for paying everything off, it has learned something about the population that is bad advice for the individual. VantageScore already doesn’t do it. FICO can stop.

Use trended data everywhere. One day’s balance is a terrible summary of a year of behavior. Every mortgage, auto loan and card decision should run on a model that sees the trend.

Make the reasons real. If a score drops, the explanation should name what cost the points, not list whatever changed in the file.

Get it out of insurance. If you want to price car insurance, price driving. More states should do what California, Hawaii and Massachusetts did.

Make the bureaus pay for their errors. One in five with an error is not a rounding issue. If an error costs someone money, the bureau that reported it should owe them the difference.

Until then

Let one card report a small balance each month so you’re not at zero. Don’t pay off your last loan right before a mortgage application. Pull all three reports once a year and dispute anything wrong.

Every one of those is a ritual to keep a number happy. None of them makes anyone better with money.