AI in Bank Risk Management: Smarter Stress Tests or Dangerous Blind Spots?

AI in bank risk management and stress testing is real but narrower than the headlines suggest. Machine learning helps banks rank borrowers, build challenger models, and mine unstructured data, yet the models that set capital and drive official stress tests still rely mostly on conventional regression. Supervisors in the United States, Europe, and at the Financial Stability Board have said AI can fit inside existing rules, and also that it can amplify old weaknesses: opacity, shared data, concentrated vendors, and models trained on a past that may not repeat.

Educational only, not financial advice. This article explains how banks and regulators use models. It does not recommend any investment, and nothing in it predicts how any bank or market will perform.

Key takeaways

• In the Federal Reserve's 2025 stress test, 22 large banks were projected to lose $549 billion under a hypothetical severe recession, and their aggregate common equity tier 1 (CET1) capital ratio fell from 13.4% to a minimum of 11.6%.
• European Banking Authority (EBA) consultation feedback found machine learning used mainly to rank credit risk, not to set capital. Most banks still use regression for internal credit risk models.
• U.S. model-risk guidance (SR 11-7, April 2011) already covers any quantitative method, including stress-testing models, and it warns about reliance on common assumptions, data, or methods across models.
• The Financial Stability Board (FSB) named four AI-related vulnerabilities in November 2024: third-party dependencies, market correlations, cyber risk, and model risk with data quality and governance.
• In a 2024 Bank of England and Financial Conduct Authority survey, 75% of 118 firms already used AI, and about one third of use cases were third-party implementations.

What does a bank stress test actually test?

A stress test is a controlled disaster simulation. Regulators write a hypothetical bad economy, banks or supervisors project what that economy would do to loan losses, trading results, and income, and the question is whether capital stays above required minimums. It works like a fire drill rather than a weather forecast: nobody expects the fire on Tuesday, but the building should still empty in time.

The Federal Reserve runs the best-known version in the United States under the Dodd-Frank Act. Its 2025 exercise covered 22 large banks over a nine-quarter horizon, from the first quarter of 2025 through the first quarter of 2027. The Fed describes the results as "hypothetical estimates that involve an economic outcome that is more adverse than expected," not a forecast.

2025 severely adverse scenario (Fed)	Value
Peak unemployment rate	10% (up 5.9 points from 4.1% in Q4 2024)
Decline in real GDP	7.8% from Q4 2024 to the Q1 2026 trough
House prices	About 33% below the Q4 2024 level
Commercial real estate prices	30% below the end-2024 level
Equity prices	Down 50% from Q4 2024 through Q4 2025

Source: Federal Reserve, 2025 Stress Test Scenarios, updated February 13, 2025.

Under that scenario the Fed projected aggregate losses of $549 billion over the horizon, of which $472 billion were loan losses, $42 billion trading and counterparty losses, $30 billion other losses, and $4 billion securities losses. Aggregate CET1 capital (the highest-quality capital, mostly common equity and retained earnings) fell by 1.8 percentage points. Loan loss rates ranged widely by product: 16.9% on credit cards, 7.2% on domestic commercial real estate, and 1.9% on domestic first-lien mortgages.

2025 result (22 banks, severely adverse)	Value
Aggregate projected losses	$549 billion
CET1 ratio, actual 2024 Q4	13.4%
CET1 ratio, projected minimum	11.6%
Decline	1.8 percentage points
CET1 ratio, 2027 Q1 recovery	12.8%

Source: Federal Reserve, 2025 Dodd-Frank Act Stress Test Results, released June 27, 2025 and revised September 30, 2025.

Two details matter for the AI question. First, the Fed said it used "largely identical models" to the 2024 test, and its own results page does not describe machine-learning methods. Second, the Fed acknowledged that its models "tend to amplify the effect of recent changes in bank data," which adds year-to-year volatility. That is a reminder that even conventional models carry weaknesses that new methods must be measured against, not just compared to a perfect ideal.

How do traditional credit risk models work?

Most credit risk modeling breaks a loan loss into three pieces. The probability of default (PD) is the chance a borrower stops paying within a year. Loss given default (LGD) is the share of the exposure the bank loses when that happens, after collateral. Exposure at default (EAD) is how much the borrower owes at that moment. Multiply them and you get expected loss.

Here is a worked example with made-up numbers. A bank lends $1,000,000 to a company with a 2% PD and a 40% LGD. Expected loss is 0.02 x 0.40 x $1,000,000, or $8,000. In a stress scenario that doubles or triples the PD to 6% and pushes LGD up to 55% because collateral values fall, the same loan has an expected loss of 0.06 x 0.55 x $1,000,000, or $33,000. Stress testing is largely the business of choosing how those inputs move together across millions of loans.

Banks that use the internal ratings-based (IRB) approach in Europe estimate these parameters with their own models, subject to supervisory approval. The EBA's 2023 follow-up report on machine learning notes that "the majority of financial institutions today use variants of regression analysis" for these internal credit risk models, while decision trees, random forests, and other machine learning techniques are used by fewer institutions.

Where is machine learning actually used?

The best evidence is not marketing claims but supervisor surveys and consultations. They agree on one pattern: machine learning is used heavily around the edges of risk management and cautiously at the center.

Use	What the technique does	What the evidence says
Credit scoring and underwriting	Ranks borrowers using many variables, including unstructured data	FSB (2024): AI credit models may help customers with thin credit files and better predict default
PD risk differentiation in IRB models	Random forests and gradient boosting trees select risk drivers and rank exposures	EBA (2023): the most common ML use in capital models; used mostly for PD
Model validation and challengers	ML models benchmark and challenge the primary model	EBA (2023): reported by some respondents; mainly to improve challenger models
Collateral valuation	Estimates and monitors real estate values	EBA (2023): few respondents reported it, though supervisors suspect wider use
Capital optimization and model risk management	Operations-focused back-office applications	FSB (2024): firms are willing to pilot and deploy
Volatility and liquidity management	ML models for stock-market moves and options books	FSB (2024): some firms have built such models
Stress scenario design and coding	Generative AI proposed for designing scenarios and writing model code	FSB (2024): described as a proposed use case in supervisory research

Sources: FSB, The Financial Stability Implications of Artificial Intelligence, November 14, 2024; EBA, Machine learning for IRB models: follow-up report, August 2023.

Notice what is missing from that table: a confirmed, widespread use of neural networks to directly produce the stress-test numbers regulators rely on. The EBA found that firms use or intend to use machine learning "only in some areas of IRB modelling, primarily for the PD estimation during the risk differentiation phase." That is much narrower than a blanket claim that AI now runs bank stress testing.

Why is machine learning used for ranking but less for quantifying risk?

One respondent in the EBA consultation gave the reason: risk differentiation can be built on shorter time series, whereas risk quantification has to be based on long-run averages that "require extended historical observation periods, which might not yet be available for new data sources which could be used by ML techniques."

An analogy helps. Ranking borrowers is like sorting runners by finishing time from a handful of recent races. Quantifying capital needs is like predicting how many runners will be injured across a full decade of marathons, including rare, brutal races. The first task can work with a short, rich record. The second needs a long record that includes the bad years, and for new data sources that record does not exist yet.

What can ML models do that regression struggles with?

The EBA report lists several practical contributions. Machine learning can select the relevant variables that feed a model, transform new data such as text through text mining, complete missing data, and handle unstructured data. In ranking exposures, it may help most where traditional risk differentiation has performed poorly, particularly for non-linear relationships, because such techniques "can improve the models' discriminatory power."

That last phrase, discriminatory power, is a technical term. It means how well a model separates the borrowers who will default from those who will not, not whether it discriminates against a group of people. The two meanings collide in fair lending, which is why AI Bias and Discrimination: How Algorithms Fail Us matters to credit models, and How Credit Scores Estimate Risk explains the older scoring methods that ML models are measured against.

Can generative AI design better stress scenarios?

The idea is appealing. Historical data contain very few genuine financial crises, so scenario libraries built from the past may underrepresent the tail of the distribution. If a generative model could produce plausible, novel shocks, supervisors could test banks against more than the repeat of 2008.

The FSB's 2024 report lists this as a proposed rather than established use. It says that proposed supervisory use cases include "employing GenAI to model social media interactions within the context of bank runs or in stress testing, both in conceptually designing stress testing scenarios and in operationalising stress testing models in code," citing a 2023 paper by Kazinnik titled "Bank Run, Interrupted," on modeling deposit withdrawals with generative AI. Much of the other supervisory use the report describes is about efficiency, such as analyzing text and drafting inspection reports.

For readers who want the technical background on one generative method, see Generative Adversarial Networks: How GANs Work. The core caution is that a generator learns the statistical shape of what it was trained on. It can recombine and extrapolate, but it cannot know about a mechanism that no data has ever shown. A scenario that is plausible to the model is not necessarily a scenario that is dangerous in the real economy.

What do supervisors worry about?

The concerns fall into four groups, and each has a primary source behind it.

Model risk is already regulated, and it covers AI

In the United States, the Federal Reserve and the Office of the Comptroller of the Currency issued joint supervisory guidance on model risk management on April 4, 2011, known as SR 11-7. It defines a model broadly as "a quantitative method, system, or approach that applies statistical, economic, financial, or mathematical theories, techniques, and assumptions to process input data into quantitative estimates," and it explicitly lists "conducting stress testing" among the uses. Nothing in that definition excludes a neural network.

SR 11-7 describes model risk as "the potential for adverse consequences from decisions based on incorrect or misused model outputs and reports." It identifies two sources: fundamental errors, and use outside the environment a model was built for. It says that "model risk increases with greater model complexity, higher uncertainty about inputs and assumptions, broader use, and larger potential impact." Those four factors all point toward more complex AI models.

The guidance asks banks to practice "effective challenge," meaning "critical analysis by objective, informed parties who can identify model limitations and assumptions and produce appropriate changes." It lists three core elements of validation: evaluation of conceptual soundness, ongoing monitoring, and outcomes analysis including back-testing.

Shared assumptions can make many banks fail the same way

This is the concern most often attributed to AI, but SR 11-7 raised it in 2011. In its discussion of aggregate model risk, it names "reliance on common assumptions, data, or methodologies" as a factor that could "adversely affect several models and their outputs at the same time."

The FSB extends the idea to AI. It says that "the widespread use of common AI models and data sources could lead to increased correlations in trading, lending, and pricing," which "could amplify market stress, exacerbate liquidity crunches, and increase asset price vulnerabilities." Picture a flock of starlings. Each bird follows simple local rules and the flock looks like a single organism. If all banks' models respond to the same signal in the same direction, the herd turns together, and so does its risk.

The FSB frames this as a vulnerability, not an event that has occurred. The report does not claim that a crisis has already been caused by correlated AI models.

Vendors and cloud providers concentrate operational risk

The FSB identifies reliance on "specialised hardware, cloud services, and pre-trained models" as a source of third-party dependencies and notes that the market for these products and services is highly concentrated. The Bank of England and FCA survey, published November 21, 2024, put numbers on it: about one third of all AI use cases among 118 responding firms were third-party implementations, up from 17% in 2022, and the top three providers accounted for 73% of reported cloud providers, 44% of model providers, and 33% of data providers.

The same survey found that 75% of firms were already using AI, up from 58% in 2022, and that the risks expected to grow most over three years were third-party dependencies, model complexity, and embedded or "hidden" models. Four of the top five current risks were data-related: privacy and protection, quality, security, and bias and representativeness.

Complexity makes models harder to explain and validate

The EBA's follow-up report names three main challenges for ML in credit risk models: overfitting, skills, and explainability. Overfitting is when a model performs well on the data it was built from but poorly on new cases, much like a student who memorizes last year's exam. The EBA says banks are expected to compare performance on the development sample with out-of-sample and out-of-time data, and one respondent noted that low-default portfolios often need larger datasets for proper training and testing.

On explainability, the report found that "a higher level of complexity may lead to better model performance but at the cost of lower explainability." Among 14 respondents to the consultation, Shapley values were the most commonly used interpretability tool, cited by 40%, followed by documentation and reporting enhancements (28%) and graphical tools (20%). The sample is small, and the EBA itself said conclusions "should be taken with caution." For more on opening up opaque models, see Mechanistic Interpretability: Reading an AI's Circuits.

Why history-trained models can miss the next crisis

Every model for risk is a summary of the past. If a shock breaks the relationships in the training data, the model's output may be confident and wrong. SR 11-7 makes the point directly: validation is "an important check on model use during periods of benign economic and financial conditions, when estimates of risk and potential loss can become overly optimistic."

Think of steering a ship by looking at the wake. You can read how the ship has been moving with great precision, and still not see the iceberg ahead. Stress testing was designed to address this limitation by imagining conditions outside the sample, but a scenario built from patterns learned on the past inherits some of the same blind spot.

The Fed's remark that its models "tend to amplify the effect of recent changes in bank data" shows that the problem is not unique to AI. A model that updates quickly on recent data can overreact to noise, and a model that updates slowly can miss a real break. Both types of error appear in practice, and the correct tradeoff depends on the use.

Methods for dealing with limited history are discussed in How Time Series Forecasting Works, and the reason small samples produce fragile estimates is covered in Why Sample Size Changes Confidence. The question of how recessions and crashes originate is explained in What Causes Recessions and How Stock Market Crashes Happen.

Which rules apply to AI in bank risk models?

There is no single AI rulebook for bank risk models. Three frameworks matter, and they apply differently.

Framework	Region	What it covers	AI-specific?
SR 11-7 model risk guidance (2011)	United States	All quantitative models, including those used for stress testing	No, but technology-neutral
EBA IRB framework and ML follow-up (2023)	European Union	Internal credit risk models for capital; principle-based ML recommendations	Partly; addresses ML explicitly
EU AI Act, Annex III	European Union	High-risk systems, including AI used to evaluate creditworthiness of natural persons or set their credit score	Yes

Sources: Federal Reserve, SR 11-7 attachment; EBA 2023 follow-up report; EU AI Act, Annex III.

A common claim is that the EU AI Act treats bank credit risk models as high-risk. The text is narrower. Annex III point 5(b) covers "AI systems intended to be used to evaluate the creditworthiness of natural persons or establish their credit score," with an exception for systems used to detect financial fraud. A portfolio stress-testing model or a corporate capital model does not obviously fall under that wording, though a retail scoring model does. The EBA's report said it supports the Act's objectives but that "the text would benefit from further clarifications to ensure that no legal uncertainties remain." That is the EBA's position from 2023, and the legal picture may have evolved since.

The FSB's recommendations are softer: it said authorities may wish to close data gaps in monitoring AI use, assess whether current frameworks adequately address the vulnerabilities, and build supervisory capabilities through cross-border cooperation. It concluded that "existing financial policy frameworks address many of the vulnerabilities," while more work may be needed.

What about runs, speed, and the human factor?

Stress tests treat the economy as a slow-moving system measured in quarters, while a bank run can unfold much faster, which is one reason supervisors have explored modeling runs with new tools. The Federal Reserve's April 28, 2023 review of the failure of Silicon Valley Bank concluded that the bank's board and management did not manage their risks, that supervisors did not fully appreciate its vulnerabilities, and that the bank had 31 unaddressed supervisory warnings at the time of failure, about triple the average of peer banks.

That review is a useful corrective. The central lesson was about governance, follow-through, and supervisory assertiveness, not about the sophistication of a model. Better algorithms do not repair a situation in which known problems go unaddressed. The same principle appears in SR 11-7's emphasis on the "influence" that validators need in order to ensure that "actions are taken to address model issues."

For related reading on how AI is used on the trading side, see Algorithmic Trading and Neural Networks, and for why diversified exposures reduce risk in the first place, see Why Diversification Reduces Risk.

Frequently asked questions

Q: What is a bank stress test? A: It is a simulation of how a bank's capital would hold up under a hypothetical severe recession. In the Fed's 2025 test, 22 large banks were run through a scenario with 10% peak unemployment, a 33% house price fall, and a 50% equity price fall. The Fed says the results are not a forecast.

Q: Do regulators' stress tests use AI? A: The Fed's 2025 results documentation describes its own stress test models, which it said were largely identical to 2024's, and does not describe machine-learning methods. The FSB reports that some supervisors are exploring AI for tasks such as text analysis and, in research, scenario design, but this is mostly at an early stage.

Q: What is SR 11-7? A: SR 11-7 is the model risk management guidance issued jointly by the Federal Reserve and the Office of the Comptroller of the Currency on April 4, 2011. It defines models broadly, asks for independent validation and "effective challenge," and warns that risk rises with model complexity and with reliance on common assumptions or data.

Q: How do banks validate machine learning models? A: Under SR 11-7, validation covers conceptual soundness, ongoing monitoring, and outcomes analysis. The EBA adds that banks should compare performance on development data with out-of-sample and out-of-time data, and many use interpretability tools such as Shapley values. Validators need ML skills, and complex models take more time and testing to challenge.

Q: Does the EU AI Act cover bank risk models? A: Partly. Annex III classes AI used to evaluate the creditworthiness of natural persons or set their credit score as high-risk, with an exception for fraud detection. The text does not name corporate credit or capital models, and the EBA asked for clarification in 2023.

Q: What happens if a bank's risk model is wrong? A: SR 11-7 says model risk can lead to "financial loss, poor business and strategic decision making, or damage to a bank's reputation," and that it cannot be eliminated. It recommends limits on model use, performance monitoring, supplementing models with other analysis, and "informed conservatism" in inputs or design.

The takeaway

The most accurate picture is also the least dramatic. Machine learning is a useful tool at the edges of bank risk management: ranking borrowers, challenging models, mining text, and speeding up analysis. At the center, in the models that set capital and drive official stress tests, regression and supervisory judgment still dominate, partly because capital models need long histories that new data sources lack.

The risks regulators describe are not exotic. They are old model-risk problems, such as shared assumptions, weak validation, and reliance on history, made larger by complexity, speed, and a small number of vendors. SR 11-7 anticipated several of them in 2011, and the FSB restated them for AI in 2024. The sensible reading is that AI changes how quickly a bank can build and update models, and the discipline of independent challenge decides whether that speed helps or hurts.

Sources

• Federal Reserve, 2025 Stress Test Scenarios, 2025 Dodd-Frank Act Stress Test Results, and Introduction.
• Federal Reserve and Office of the Comptroller of the Currency, Supervisory Guidance on Model Risk Management (SR 11-7), April 4, 2011.
• Financial Stability Board, The Financial Stability Implications of Artificial Intelligence, November 14, 2024, and the BIS FSI executive summary, June 26, 2025.
• European Banking Authority, Machine learning for IRB models: follow-up report (EBA/REP/2023/28), August 2023.
• Bank of England and Financial Conduct Authority, Artificial intelligence in UK financial services, 2024, November 21, 2024.
• EU Artificial Intelligence Act, Annex III: high-risk AI systems.
• Federal Reserve Board, Press release on the review of the supervision and regulation of Silicon Valley Bank, April 28, 2023.

---

Read next:

• Can AI Predict the Next Financial Crisis? The Limits of Machine Learning Systemic Risk Models
• Fraud Detection & Anomaly Learning: How AI Catches Financial Crime
• AI Loan Approval: When Algorithms Decide Who Gets Credit and Who Doesn't