At the end of Q2 2026, the New York Federal Reserve reported that loan originations hit a record $211 billion, the highest figure in the Fed’s data history. At the same time, lenders continued to extend credit to an increasing number of borrowers with lower credit scores, even as many of them transitioned into serious delinquency.

The challenge for many lenders was not that the models they have used for the past year were wrong or badly built; it was that underlying market conditions had changed.

Most lenders’ instinct would be to treat these issues as a model-refinement challenge. Retune the model, adjust thresholds, and “fix” the models for the new reality. The deeper problem, however, is that most lender risk models, whether built in-house or acquired through third-party vendors, are tuned to broad signals shared across multiple lenders and reviewed and updated on fixed calendar cycles.

While historically model builders have worried about models “drifting” out of date, the bigger issue is that some of the highest-value data available to most lenders, their own in-house data, is seldom, if ever, mined for anything beyond a handful of variables that an analyst thought to test and validate when the model was built or last updated.

What a national model can’t see in your book

Models based on industry-standard data are designed, by definition, to identify patterns that fit national averages. This can come at the expense of state or market-specific signals. This tradeoff is necessary because national models must apply across a wide range of lending institutions and markets, which would be prohibitively expensive to achieve with high specificity.

The available data itself shows a compelling story. When looking at auto loan delinquency rates by state, there is a broad range of variability, with some states often running well above national averages, while others sit at a third of the same scale.

The gap between lenders that overperform and those that fall behind stems from subtle but important nuances in regional income, employment figures, and borrower behaviors that nationally based models were never designed to account for. A national scoring model tends to smooth over state-by-state or regional nuances to make it more broadly applicable nationwide.

Each year, TransUnion provides a year-end Consumer Credit Forecast that, for 2026, sees the growth in delinquency rates decelerating while delinquency itself remains a challenge, helped to some extent by refinancing (often led by credit unions). Nationally, these trends have value and will likely hold, but broad indicators offer little guidance to lenders with highly regional or unique portfolio characteristics that don’t fit national averages.

The concentration risk no one is accounting for in their pricing

Lenders nationwide have increasingly adopted industry-standard models or have built models that are largely based on national scores. As these practices expand, the exposure to risk across auto lenders begins to converge and creates conditions where lenders increasingly start to react in similar ways to unexpected shifts in market conditions or economic stress.

Bank examiners tend to refer to loans based on such models as being “dependent on the same scorecard or automated decision model,” and they tend to group such loans into risk categories with other loans that the examiners believe will behave in similar ways when exposed to the same pressures.

Rating agencies also consider portfolio-level risk to be subject to similar exposures when lender models share data points and assumptions. The more lenders share these assumptions and data points, the more likely the losses the lenders will experience will also move in parallel.

Lenders must modernize their practices by growing beyond simply relying on national models or models built only on nationally syndicated data. To modernize their lending process, institutions must adopt a strategy that leverages the wealth of information in historical performance data that all lenders possess and that is unique to each lender.

By adopting such a layered approach, lenders can identify more fine-grained signals that are unique to their market, their region, and their portfolio.

The data sitting unused

The challenge for a modern lender is not to “find the right vendor model,” but to effectively integrate industry-standard models and data points with their own unique signals derived from in-house, historical performance data.

Most lenders retain a wealth of data, including application details, payment speeds, dealer-level data and trends, and regional performance data accumulated over years of originations and servicing. For most lenders, this historical data often remains unexamined except for traditional scorecard variables. Lenders tend to underutilize their own historical data because of practicality, not poor strategy.

For a typical analytics team, it’s simply not feasible to produce a large number of hypotheses about which combinations of data sources, tables, and columns are likely to produce valid model inputs.

Artificial intelligence, especially techniques related to pattern recognition and “signal discovery,” allows lenders to automate the hardest parts of this process by automatically discovering connections between data sets, finding and evaluating patterns, and providing analytics teams with signals that are most likely to provide model lift with minimal effort.

By employing these types of AI techniques, lenders can detect portfolio-level signals to augment their industry-based baselines.

It’s important to understand the difference between “decision automation” and “signal discovery.” Decision automation focuses on processing speed, while signal discovery focuses on accuracy. Automating the lending process without improving signal detection risks amplifying model limitations that your analytics team may not even be aware of.

Boosting originations at 10x speed will also increase delinquencies at the same 10x speed unless you can identify blind spots in the model. Speed is most effective when paired with better signal detection, giving lenders the ability to align portfolios more accurately with changing market conditions.

Validation based on a calendar versus conditions

On April 17, regulators issued SR 26-2, which went beyond simply replacing a decades-old framework. With SR 26-2, regulators quietly agreed with an argument that risk teams have been making for years; namely, that the timing of a model review should be driven by the changes to the underlying model and whether those changes were significant enough to warrant a review, not by an arbitrary date on a calendar.

Static model reviews and updates treat models as “one-off” creations that must be maintained to adjust for changing national scores and inputs. Instead, models should be treated as “systems” that must be upgraded and changed on a regular basis as economic and market conditions change and impact the portfolio. A model calibrated on a yearly, or even quarterly basis, but that uses the same set of inputs simply updated will not address the challenges that regulators are pointing to.

The ‘Black Box’ compliance constraint

One significant challenge with modern risk modeling lies in the difference between “predictions” and “explanations.” Predictions are mathematical exercises. Explanations are regulatory. Because of the Equal Credit Opportunity Act (ECOA) and Regulation B, lenders must provide specific, accurate, and compliant reasons for any adverse action taken against borrowers.

Requirements for model transparency do not mean that models must, by definition, be simple. Rather, model transparency means that the decision-making logic should be easy to export and easy to trace and understand. In simpler terms, consumers and examiners should be able to understand the variables that drove (for example) an adverse action notice, based on codes generated by the model that recommended the decision.

The goal for lenders is to move to an architecture that pairs advanced signal discovery with the precision of mathematical predictions.

A regional lender finds what the national model missed

Consider the following hypothetical example of a mid-sized lender. Its national bureau-based model treats the portfolio as broadly representative of the region the lender serves, but only because the model’s assumptions were based on data from many lenders at once.

By leveraging AI-based signal discovery, the lender can analyze its own historical origination and servicing data to spot segment-specific patterns tied to a local dealer network and a precise borrower cohort. This type of very granular, but very regional and narrow signal is too specific and restricted to an individual lender to be picked up by national models, but is important and impactful enough to be material to the profitability of our hypothetical lender.

The signal was always present in their data; the lender simply did not have the staff, the time, or the resources to spot it without testing dozens of hypotheses and variables manually over the course of weeks or months.

The advantage is already on the balance sheet

To maximize the use of your existing infrastructure and to deliver true strategic value, lending leadership should apply a three-point evaluation process when investing in AI technology and evaluating their risk models:

—Scope: Does the tool alleviate friction in processing, or identify blind spots in risk? It’s exceptionally rare that a tool will be able to successfully do both.

—Auditability: Is it easy to export output as standard logic (e.g., SQL/rules), or is the information locked in proprietary dashboards? Regulatory compliance requires that logic be transparent and accessible for independent review.

—Actionability: How “actionable” is the output? Does the model identify specific segments (e.g., dealer-specific delinquency patterns), or does it give broad generic national trends?

This does not mean that AI-driven risk models are unreliable; in fact, far from it. For most lenders, however, it means that it’s imperative to stop using AI to “buy someone else’s answer.” The use of AI should be tailored to your portfolio and the nuances of your region and market demographics and behavior patterns.

The biggest advantage each lender possesses has always been hidden within each lender’s historical data, waiting for a way to analyze all of it, instead of simply the limited subset selected by an analyst based on available staffing. Institutions that can prioritize mining their own historical data will not only recalibrate their risk models faster than their peers; they will also be able to evaluate risks that other lenders using the same shared model are blind to.

With a PhD in statistics and experience as a two-time founder, Bart Blackburn brings deep data science expertise to his role as a staff data scientist and field product manager for dotData. He focuses on applying advanced machine learning to solve complex business challenges and deliver actionable insights for dotData’s clients. His entrepreneurial background includes co-founding Priceflow, an ML-powered auto-pricing company acquired by TrueCar.