Few numbers carry as much weight in American financial life as the credit score, a three-digit figure that plays a role in mortgage, car loan, and credit card access and rates. Introduced as the FICO score in 1989 by Fair Isaac Corporation, the credit score is like other facets of modern life being reshaped by AI and machine learning (ML). The credit industry is navigating that shift carefully, balancing rapid advances in technology against regulatory requirements built around a very different kind of model.
Changes are not yet visible to consumers. Their credit score, a number between 300 and 850, looks the same as it always has. But inside the institutions that build and deploy these models, ML is supplementing the traditional methods that have defined credit risk assessment for decades. Better computing infrastructure, an explosion of available data, and a regulatory environment that is slowly becoming more accepting of algorithmic decision-making are converging to transform how creditworthiness is measured.
Credit Scores Are Built to Be Explained
Traditional credit scoring is built on scorecards: interpretable, engineered models that assign weighted values to a defined set of inputs and produce an output that humans can audit step by step. The inputs behind a FICO score are well-established, with payment history carrying the greatest weight at 35%, followed by amounts owed (30%), length of credit history (15%), new credit inquiries (10%), and credit mix (10%). Auditability is important: the Equal Credit Opportunity Act and the Fair Credit Reporting Act both require that consumers who are denied credit receive a specific explanation: not a probability or a model output, but a ranked list of the actual factors that drove the credit decision. Scorecards produce those explanations cleanly.
Machine learning models are capable of processing far more data than a traditional scorecard, identifying patterns that the traditional fixed set of inputs cannot account for. The result is a model that can be highly accurate but harder to explain in plain terms, and that is at the center of the industry’s challenge. The same laws that require scorecards to produce clear adverse-action explanations apply equally to ML models, regardless of their complexity. A lender cannot tell a denied applicant merely that the algorithm said no. The Consumer Financial Protection Bureau (CFPB) has been explicit on this point: model complexity does not relieve a lender of its obligation to explain a credit decision accurately.
The Case for Machine Learning Is About Speed
FICO’s internal research puts some numbers around the potential benefits of ML in credit scoring. In one analysis, a single analyst built an ML model in roughly 40 hours that matched the predictive performance of FICO Score 9, a model that took five analysts a full month to construct. The accuracy improvement was less than 2%, meaning the argument for adoption is more about speed and less about creating a better or more predictive model.
Despite potential efficiency gains, large institutions have been slow to move. More complex models create regulatory uncertainty, for which lenders have little appetite. Instead, what is occurring is a more gradual ripple of adoption: newer modeling approaches appear first in fraud detection, which has fewer regulatory constraints, then in smaller institutions and new market entrants, and eventually reach larger banks as regulatory acceptance builds.
New Data, New Models
Part of ML adoption in credit circles is a problem that traditional scorecards were not designed to handle: the volume and variety of data now available to credit modelers. When the industry talks about ML changing credit scoring, it is referring mostly to the in-house models lenders build themselves, not the vendor scores sold through credit bureaus. Traditional bureau data captures credit accounts, payment history, and public records. Now the scope of data can include rent and utility payments, payroll and income data, and, most significantly, bank account transaction history. Open banking is the reason: with a consumer’s permission, lenders can pull real-time data directly from linked accounts, showing financial behavior a bureau report has never captured.
Two product launches in late 2025 show how quickly this is moving. In November, Experian announced its Credit + Cashflow Score, combining traditional credit data with consumer-permissioned bank account transactions, alternative credit bureau data, and 24 months of trended data into a single 300-850 score. Early internal analysis showed a roughly 40% improvement in predictive accuracy over conventional credit models. Days later, FICO announced a partnership with Plaid to deliver an updated UltraFICO Score that incorporates real-time cash flow data from more than 12,000 connected financial institutions. Both products represent a substantial expansion of what goes into their respective scores.
Research published on arXiv examining credit risk modeling found that advanced tree-based models produce the best predictive results. Gradient boosted trees, a widely used ML approach in credit risk today, build hundreds of simple decision trees in sequence, each one correcting errors of the one before.
ML models can take in more data to make a predictive output, but the gap in explainability still needs to be closed.
Tools That Explain the Tools
The industry has designed tools to bridge the gap between ML complexity and regulatory scrutiny requirements. SHAP (Shapley Additive Explanations) uses a game-theory framework to assign each input variable a specific contribution value for each individual prediction. LIME (Local Interpretable Model-agnostic Explanations) builds a simplified model around a single decision to estimate the reasoning for each specific point. Both are now widely cited as standard tools in financial services model development to generate reason codes required for an adverse action notice.
Whether they are sufficient is an open question. Cristián Bravo, Canada Research Chair in Banking and Insurance Analytics at Western University, in Ontario, said that the latest Basel Committee on Banking Supervision report suggests SHAP, LIME, and a third approach, counterfactual explanations, are options for satisfying explainability requirements. Having a set of tools does not mean the problem is closed. “That it is very much not a settled question,” Bravo said. “On the one hand, we today are asking far more out of explaining credit risk models than twenty years ago, which is probably a good thing. But on the other hand, these asks haven’t been truly tested.”
Bravo said there is a gap that practitioners and regulators tend to overlook: the difference between explanations that satisfy experts, and those that mean something to the consumer who receives them. “If we give a SHAP force plot or LIME plot to a regular consumer, will they consider this enough? I think we have a set of tools that we as experts consider good enough, flaws aside, but this is untested with the consumer.”
Where the Industry Goes
Bravo assessed the industry’s place on the adoption curve: for tree-based ML models, the transition is already mature. “Most institutions are using some type of tree ensemble model for lending,” he said, “and only those with legacy systems or data issues are using older logistic or spline-based regressions.” He sees an open issue with foundation models, the large general-purpose systems on which ChatGPT and similar tools are built.
A few banks are developing their own. Royal Bank of Canada in 2025 announced ATOM, a proprietary foundation model trained on billions of client transactions that’s applied to credit adjudication and other functions. Whether that approach becomes standard is still unresolved. Said Bravo, “We are in a transition phase where we are finding the role of foundation models in credit risk, how they fit within regulation and modern credit risk frameworks, while at the same time evaluating the profit potential they have in the face of much higher development and deployment costs.”
The traditional FICO score is not going away. It remains the dominant standard in U.S. consumer lending. But the mechanics underneath it are shifting, and the pace of that shift is unlikely to slow. The number on the screen will look familiar for some time to come. What produces it is already something different and becoming more so.
Mark Broderick is the principal banking and payments industry analyst at Panoramic Research.
Join the Discussion (0)
Become a Member or Sign In to Post a Comment