Machine Learning Credit Scoring: Better Predictions, Bigger Fairness Problems

The three-digit FICO score has determined Americans' access to credit for over 30 years. It is blunt, slow to update, and built on a narrow slice of financial behavior. Yet it remains the gatekeeper for mortgages, car loans, and credit cards for over 200 million people. Machine learning models now offer an alternative that is simultaneously more accurate and more troubling — capable of predicting defaults with substantially less error, while potentially embedding discriminatory patterns that no human reviewer can easily detect or explain. The stakes could not be higher: in 2023, U.S. lenders issued $4.2 trillion in consumer credit.

Educational only — not financial advice. Investing involves risk of loss; past performance does not guarantee future results. Consult a licensed financial advisor before making investment decisions.

Key takeaways

• Machine learning credit models can reduce default prediction error by 15%–25% compared to traditional FICO scores, potentially expanding credit access for underserved populations.
• The same models risk encoding proxy discrimination — using variables like zip code or spending pattern that correlate with race, even when race itself is excluded.
• The Equal Credit Opportunity Act requires lenders to provide adverse action reasons, which black-box ML models struggle to satisfy.
• Regulatory momentum is building: the CFPB and OCC have both issued guidance signaling closer scrutiny of algorithmic lending.

Why Traditional Credit Scoring Falls Short

The classic FICO model looks at five factors: payment history, amounts owed, length of credit history, new credit, and credit mix. This is a remarkably thin profile. It ignores rent payments, utility bills, income stability, educational attainment, and thousands of behavioral signals that modern data infrastructure can capture.

About 26 million Americans are "credit invisible" — they have no credit file — and another 19 million have files too thin to score, according to the CFPB. These populations skew toward younger adults, immigrants, and lower-income households. Traditional scoring leaves them locked out of formal credit markets entirely, often pushing them toward payday lenders charging 300%+ APR.

Worse, FICO scores are notoriously slow to update. A missed payment from three years ago weighs heavily even if a borrower has been perfectly responsible since. For someone rebuilding after a medical emergency or job loss, this creates years of artificially constrained access.

What Machine Learning Brings to the Table

Modern ML models — primarily gradient boosting frameworks like XGBoost and LightGBM, along with neural networks — can ingest hundreds or thousands of variables simultaneously and identify non-linear patterns that simple regression models miss. The question is what data to feed them.

Alternative data is the category generating the most activity. This includes cash flow data from bank accounts (showing income regularity and expenditure patterns), rental payment history, utility and telecom payments, and in some cases, app usage or device data. Upstart, the fintech lender that went public in 2020, built its business on ML models incorporating education, employment history, and cash flow alongside traditional credit variables. The company claims its models approve 27% more borrowers than traditional methods while maintaining similar default rates — and charge borrowers an average interest rate 38% lower than competitors. An independent analysis by the CFPB in 2022 broadly validated these claims, noting Upstart's model approved more minority borrowers at lower rates.

ZestFinance (now Zest AI) takes a different approach: its models aim to make existing credit data work harder by finding patterns that FICO misses within the same data. Analyses using Zest AI's models found they could extend credit to 15% more Black and Hispanic applicants than traditional scoring — without increasing default rates — simply by modeling the same data more sophisticatedly.

The Discrimination Problem

More predictive power does not guarantee fairness, and this is where machine learning credit scoring generates legitimate alarm.

The core concern is proxy discrimination. Even when a model never directly inputs race (which would be illegal under the Equal Credit Opportunity Act), it can effectively approximate it through correlated variables. Zip code, for example, correlates strongly with race due to decades of residential segregation. Shopping patterns, phone type, even word choice in loan applications — all can serve as proxies for protected characteristics. A model that is "colorblind" by design can still be discriminatory by outcome.

Academic research has documented this extensively. A 2019 paper by researchers from UC Berkeley found that mortgage lenders, including algorithmic ones, charged Black and Latino borrowers 8% more in fees and interest than similarly qualified white borrowers. A 2021 University of Chicago study found that fintech lenders' algorithms showed similar disparate impact to traditional lenders in some segments despite advertising fairer models.

The Consumer Financial Protection Bureau issued a circular in 2022 explicitly stating that "black-box" algorithms do not exempt lenders from the requirement to provide specific, accurate adverse action reasons when denying credit. This creates a direct tension: many ML models — particularly deep neural networks — resist interpretable explanations by design.

The Interpretability Imperative

Regulators in both the U.S. and EU are forcing the issue. The EU's AI Act classifies credit scoring systems as "high-risk AI," requiring transparency, auditability, and human oversight. In the U.S., the OCC has issued guidance requiring banks to be able to explain and validate any model used in credit decisions.

This has made explainable AI (XAI) techniques — particularly SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) — essential tools in the ML lending stack. These methods assign contribution values to each input variable for each individual decision, enabling something approximating an explanation: "Your application was denied primarily because of high credit utilization (40% impact) and short account history (32% impact)."

Whether these post-hoc explanations satisfy regulators — or whether they genuinely explain the model's internal reasoning — remains contested. Some researchers argue that SHAP explanations for complex ensemble models are themselves approximations that can mislead.

Firms like Zest AI have responded by building compliance directly into their model architecture, using "constrained optimization" that maximizes predictive accuracy subject to explicit fairness constraints — capping the allowed disparity ratio between demographic groups. This is the leading edge of the field, but it involves explicit tradeoffs: enforcing strict demographic parity typically reduces overall predictive accuracy somewhat.

The Path Forward: Fairness-Aware ML

The field is converging on the view that fairness in credit scoring is not a single concept but several competing ones — equal rejection rates across groups, equal error rates, equal predictive accuracy — and that satisfying all simultaneously is mathematically impossible in most real-world scenarios (a result known as the "impossibility theorem" of algorithmic fairness, formalized by researchers Jon Kleinberg, Sendhil Mullainathan, and Manish Raghavan in 2016).

This means lenders and regulators must make explicit policy choices about which fairness criteria to prioritize — a profoundly political question disguised as a technical one. Hiding that choice inside an algorithm does not make it disappear; it just makes it less visible and accountable.

The most promising path forward involves transparency by design: open-source model documentation, regular demographic disparity audits, and clear regulatory standards for what constitutes acceptable disparate impact in algorithmic lending. The CFPB's current rulemaking on AI in lending is expected to address these questions directly.

FAQ

Does machine learning credit scoring help or hurt minority borrowers? The evidence is mixed. Well-designed ML models using alternative data have demonstrably approved more minority borrowers at lower rates than traditional models. But poorly designed or minimally audited models can reinforce or amplify existing disparities through proxy variables. The outcome depends heavily on the data used, the fairness constraints imposed, and the rigor of disparate impact testing.

What is "disparate impact" in lending? Disparate impact occurs when a lending policy or model produces significantly different outcomes for different demographic groups — even without explicit discriminatory intent. Under the Equal Credit Opportunity Act and Fair Housing Act, disparate impact is actionable regardless of intent. The standard threshold is the "80% rule": if the approval rate for a protected group is less than 80% of the highest-approved group's rate, it may constitute illegal disparate impact.

Why can't lenders just remove all demographic proxies? This approach — sometimes called "fairness through blindness" — doesn't work because many variables correlate with protected characteristics. Zip code, income, employment type, and educational attainment all have demographic correlations. Removing all proxies would require removing most of the informative variables in the model, severely reducing predictive accuracy. Fairness-aware ML instead works by detecting and correcting for proxy discrimination while retaining predictive variables.

What alternative data is used in ML credit scoring? Common sources include: bank account cash flow data (rent, bill payments, income patterns), rental payment history from property management platforms, telecom and utility payment records, employment verification data, education history, and in some markets, social media signals (though use of the latter is controversial and legally risky in the U.S.).

Are ML credit models regulated? Yes. In the U.S., all lenders using algorithmic models remain subject to the Equal Credit Opportunity Act, Fair Housing Act, and CFPB oversight. The OCC requires banks to validate and be able to explain credit models. The CFPB's 2022 circular on adverse action reasons specifically addressed AI/ML models, and further rulemaking on algorithmic credit is underway. The EU's AI Act imposes additional requirements for models used in EU credit markets.

Read next:

• AI Loan Approval: When Algorithms Decide Who Gets Credit — and Who Doesn't
• Machine Learning in Insurance Underwriting: Smarter Pricing, Sharper Risks
• Blockchain + Machine Learning: How AI Is Reinventing Financial Compliance and KYC