Inside a Fair Isaac data lab in San Jose, a model engineer studies a chart of repayment patterns from forty million US credit card accounts, looking for the tiny statistical lifts that justify the next score release. That work, repeated quarterly across FICO, VantageScore, and the largest US banks, sits at the heart of every mortgage approval, card application, and small business loan in the country. This piece walks through how US credit scoring algorithms actually compute a number, version by version and weight by weight.
The FICO formula and the five inputs that matter
Every FICO score, from FICO 8 to FICO 10T, is built on five input categories with publicly disclosed weights. Payment history carries roughly thirty-five percent. Amounts owed and utilization carry thirty percent. Length of credit history carries fifteen percent. Credit mix carries ten percent. New credit and recent inquiries carry the final ten percent. The score range is 300 to 850. The formula is proprietary at the parameter level, but the five buckets and their approximate weights are published by Fair Isaac in lender documentation and consumer education material.
Inside the payment history bucket, a single 30-day late payment can drop a high score by 60 to 110 points, while a 90-day late or charge-off can drop it by more. Utilization is calculated per account and across the file, with the algorithm penalizing a single card above 30 percent of its limit even when the overall ratio is low. Length of history reads the age of the oldest account, the average age, and the age of the newest account, which is why closing an old card can hurt the score. Credit mix rewards a US consumer who carries both revolving cards and installment loans, since mixed performance is a stronger predictor than either alone. New credit inquiries decay after twelve months and drop off the file after twenty-four.
FICO 10, released in 2020, kept the five-bucket structure but retuned the weights inside each bucket using more recent default data. FICO 10T, released in 2024, added trended data, which means the algorithm reads twenty-four months of balance and payment patterns rather than only the current snapshot. A US consumer who pays down a balance over six months scores higher under 10T than one who simply happens to have a low balance on the reporting date. The Consumer Financial Protection Bureau research reports, on the CFPB research reports page, track how these version changes shift score distributions across US income and geography segments.
VantageScore 4.0 and the move to machine learning
VantageScore is the joint product of Experian, Equifax, and TransUnion. VantageScore 3.0, released in 2013, was the first widely adopted US score to reach roughly thirty-three million additional US adults that FICO 8 could not score. VantageScore 4.0, released in 2017, introduced machine learning techniques inside the algorithm. The model still produces a single number from 300 to 850, but the function that maps inputs to output is no longer a simple weighted sum. It is a trained model that captures interactions between variables, for example how the combination of high utilization and recent inquiries predicts default more sharply than either signal on its own.
VantageScore 4.0 also weights rent and utility data when the bureau carries it. Rent reporting has expanded sharply, with most large US property managers reporting to at least one bureau. The TechBullion open banking US update tracks how rent and cash flow inputs have been wired into the bureau pipeline.
The shift to machine learning inside a regulated score raises a clear question. Federal Reserve and OCC guidance under SR 11-7 requires that any model used in credit decisions be developmentally validated, independently reviewed, and monitored for performance drift. VantageScore publishes annual performance studies showing the Gini and KS metrics that lenders use to compare scores, and the model documentation runs to several hundred pages. The Federal Reserve modernization work, summarized on the Federal Reserve payments page, references the supervisory expectations that apply to every score in production at a US lender.
Custom scorecards and gradient-boosted trees inside lenders
The bureau score is only one input at most US lenders. Card issuers, auto lenders, and large banks run their own scorecards that combine FICO or VantageScore with internal data, including deposit history, prior product holdings, and channel of application. The classical engine for these scorecards is logistic regression with weight-of-evidence binning, which produces a score that is easy to explain and easy to audit. Many US issuers still run logistic scorecards for adverse action reasoning even when they use a different model to make the decision.
The newer engine is gradient-boosted trees. XGBoost, LightGBM, and CatBoost are the three most common libraries in production at US lenders. A gradient-boosted scorecard typically lifts Gini by three to seven points over a logistic baseline on the same data, and several US issuers have shifted card and personal loan underwriting to that engine since 2020. The model output is then passed through a calibration step that maps probability of default to a score range the business teams understand. Some US lenders calibrate to the FICO range, others to a 0 to 1000 internal range.
Reason code generation sits on top of the model. For each adverse action, the lender must produce four or five specific factors that drove the decision, in the language a US consumer understands. SHAP values and counterfactual analysis are the two common techniques. The lender owns the explanation, not the vendor, which is why several US issuers have walked back the most opaque deep learning techniques in favor of trees and additive models that are easier to explain.
ECOA, Regulation B, and adverse action notices
The Equal Credit Opportunity Act and its implementing rule, Regulation B, set the legal floor for US credit scoring. ECOA prohibits discrimination on the basis of race, color, religion, national origin, sex, marital status, age, or receipt of public assistance. Regulation B requires that any lender denying a credit application send an adverse action notice within thirty days, with the specific reasons for the denial.
The CFPB issued a 2022 circular making clear that the adverse action requirement applies even when the underlying model is a machine learning system that is hard for a human to interpret. The lender cannot use opacity as a defense. That position has shaped US credit scoring algorithm design at every major issuer, and it is the practical reason that reason code generation is now treated as a model component rather than a reporting afterthought. The TechBullion regtech compliance overview tracks how lenders document these controls under examination.
Fair lending testing is the second leg of the framework. US lenders run disparate impact analysis on every score and every model, comparing approval rates and average scores across protected class proxies. Where a disparity is found, the lender must show either a business justification or move to a less discriminatory alternative model. The McKinsey financial services insights page, available at McKinsey financial services insights, documents the operational steps US lenders take to embed fair lending testing into model development.
Model risk under SR 11-7 and what changes next
SR 11-7, issued by the Federal Reserve in 2011 and adopted by the OCC and FDIC, is the rulebook for model risk management at US banks. Every credit scoring model in production at a covered bank must have an inventory entry, a development document, an independent validation, ongoing monitoring, and a defined retirement plan. SR 11-7 does not name FICO or VantageScore, but every US bank that buys those scores treats them as in-scope models and runs the same controls on them as on internal models.
Monitoring under SR 11-7 watches for population stability, score drift, and Gini decay. When discrimination weakens past a defined threshold, the bank retrains, recalibrates, or replaces it. The COVID era forced this work, since pandemic stimulus and forbearance broke historical relationships the scores rely on. Most US lenders have since rebuilt their challenger models on post-2020 data, and FICO 10T and VantageScore 4.0 reflect the same recalibration at the bureau level.
Three changes will reshape US credit scoring algorithms over the next two years. First, Fannie Mae and Freddie Mac are transitioning conforming mortgages to FICO 10T and VantageScore 4.0 under a Federal Housing Finance Agency timeline first set in 2022. Second, the CFPB open banking rule under Section 1033 expands cash flow data inputs across consumer credit, making UltraFICO-style cash flow features mainstream rather than opt-in. Third, the NIST AI Risk Management Framework, although voluntary, is being cited by US bank supervisors as the benchmark for documentation and governance on every model in production. For broader context, the TechBullion AI in financial services explainer tracks how these governance practices spread across US lenders. The next twenty-four months of FHFA implementation and CFPB rule activity will decide whether US credit scoring algorithms become more inclusive and more transparent, or simply more complex to operate inside the same regulatory frame.
