What a credit score actually measures

January 14, 2026 · by Dana Reinholt · checked by Marcus Ople

A generic consumer credit score is a number produced by a statistical model trained to answer one question: among the people whose files looked like this one, what share went at least 90 days delinquent on some account within the next 24 months?

That is the entire definition. It is not a measure of wealth, of income, of character, or of whether a lender will approve you. Income does not appear on a consumer credit report at all, so no score built from that report can be reading it.

Why the definition dissolves most of the folklore

“Carrying a small balance helps.” The model has no way to reward you for paying interest; interest paid is not an input. A reported balance of zero and a reported balance of forty dollars both produce a low utilization figure. Only one of them costs you money.
“Checking my own score hurts it.” Your own review is a soft inquiry. It is visible only to you, and the models do not read soft inquiries at all.
“I have no debt, so my score should be perfect.” A file with no open revolving account gives the model very little to work with. Thin files land in the middle of the range, because the model is uncertain rather than impressed.

There is not one score

There are several model families, multiple generations of each, and three national consumer reporting agencies whose files do not perfectly agree. A mortgage underwriter, an auto lender and a card issuer are frequently looking at three different numbers about you on the same afternoon, and all three are legitimate. The figure a free app shows you is often an educational variant that no lender buys.

On score gaps

A twenty or thirty point gap between two scores is usually a model difference and not an error. A hundred-point gap is worth chasing, because it usually means one bureau holds a tradeline or a collection the others do not — which is exactly what a dispute exists to resolve.

What the model can see

Broadly, and in descending order of typical weight: whether you have paid on time, how much of your revolving limits you are using, how long the file has existed, what mix of account types it holds, and how recently you have sought new credit. The precise weights are unpublished, differ by model, and shift with what else is in the file. Anyone giving you an exact percentage breakdown for your file specifically is guessing.

One consequence worth sitting with: two people can do identical things for a year and see different point movements, because the model is scoring the whole file, not the action. This is why we almost never publish a number of points a change is “worth.”

Next: reading a credit report line by line, where all of these inputs actually live.