How Do Credit Scores Work? A Number That Predicts One Thing

The major scoring models differ in detail and agree on the broad weighting, and the factors are all about borrowing behaviour rather than about means:

• •Payment history, the largest single component at around a third or more, meaning whether past obligations were paid on time, with recent misses counting far more than old ones
• •Amounts owed, particularly the proportion of available revolving credit in use, where a lower utilisation scores better and using most of a limit signals distress
• •Length of credit history, since a long record gives the model more to work with, which is why closing an old account can lower a score
• •Credit mix, meaning whether a person has handled different kinds of borrowing, which carries a small weight
• •New applications, since several searches in a short period suggest a search for credit that has not yet appeared in the data, though multiple searches for a single mortgage or car loan are usually treated as one
• •What is not included: income, savings, employment, age, race, religion, marital status and, in most systems, current account balances, because using several of these is unlawful and using the others turns out to add little predictive power

The score is computed from a file held by a credit reference agency, which collects reports from lenders about accounts, balances and payment behaviour, together with public records such as court judgments and insolvency, and electoral roll data used to confirm identity and address. The agencies are private companies, the lenders supply the data voluntarily under reciprocal arrangements, and a person's file may differ between agencies because not every lender reports to all of them. The score itself is calculated by a model applied to that file, so a single person has several scores rather than one, and the number a consumer sees on an app is frequently an educational score calculated by a different model from the one a lender will use. The practical consequence is that comparing numbers between services is meaningless, while the direction of movement and the underlying file are not.

The purpose is narrower than its uses. A score is built by taking millions of past borrowers, recording who defaulted, and fitting a model that separates them, so the output is a rank ordering by default risk and nothing more. That is genuinely valuable, because it replaced a system in which lending decisions were made by branch managers using judgement, which was slower, less accurate, and considerably more prone to discriminating by appearance, accent and acquaintance. Automated scoring expanded access to credit substantially and made pricing risk-based, so that a lower-risk borrower pays less. The problem arises from mission creep, since scores are now used for decisions the model was never validated for, including tenancy applications, insurance pricing in some jurisdictions, mobile phone contracts and, in parts of the United States, employment screening, where the evidence that a credit file predicts job performance is essentially absent.

The system has a structural problem at its edges. A person with no borrowing history is not scored as low risk but as unscorable, which is worse, and this affects young adults, recent immigrants and people who have always used cash, a group numbering in the tens of millions across major economies. Being invisible to the model means being refused by lenders who use it, which prevents building the history that would make one visible, a loop that is difficult to break without a starter product such as a secured card. At the other end, a serious default remains on a file for six or seven years in most systems and depresses access throughout, which can outlast the circumstances that caused it by a long way. Attempts to widen the inputs, using rent payments, utility bills and bank transaction data under open banking rules, are intended to address the first problem and raise a version of the second, since more data means more ways to be adversely classified.

A short list of persistent and false beliefs is worth stating plainly. There is no blacklist, and no central register of people refused credit; a rejection is a lender's decision based on its own criteria. Checking your own file does not affect your score, since that is a soft search visible only to you. Carrying a balance and paying interest does not improve a score, and paying in full each month is better on every measure. Being in debt is not required to build a history, though having and using an account responsibly is. Closing unused accounts frequently lowers a score rather than raising it, by cutting available credit and shortening history. And a person's score is individual: there is no household or shared score, and a partner's file affects yours only where accounts are genuinely joint, which is why a financial association created by a joint account persists until it is formally disassociated.

A credit score is a statistical estimate of the probability of serious delinquency, built mainly from payment history, the proportion of available credit in use, the length of the record and recent applications, and excluding income, savings and employment entirely. It is computed from a file held by a private agency, and different models produce different numbers, so only the direction and the underlying file matter. Scoring replaced subjective lending and widened access, its worst failure is that people with no history are unscorable, and the common beliefs about blacklists and carrying balances are false.