Big data, AI, and machine learning enable banks to track volatility in the market and audit revenue patterns. Could some of these technologies have prevented the recent collapses of banks?

Brian T. Horowitz, Contributing Reporter

June 6, 2023

5 Min Read
Bright US dollar sign on night city background. Zoom effect.
Andrey Khokhlov via Alamy Stock

In a volatile economic climate with financial institutions such as Silicon Valley Bank and First Republic Bank failing, experts say emerging technology could help them meet the requirements of regulators and track financial risk.

The Federal Reserve requires banks to perform tests on their investment portfolios against future financial events. This process is called computer-simulated stress testing. By identifying risks, banks can change their decisions around lending standards and operations, according to Michael Heffner, vice president of industry solutions and go-to-market at cloud computing company Appian, which coordinates stress tests for banks.

“You're changing behavior and decisions around everything from lending standards to operational processes around how you control for risk, which is a big deal for banks and capital market firms,” Heffner says.

Stress tests in the United States include the Comprehensive Capital Analysis and Review and the Dodd-Frank Act Stress Test.

Heffner believes that stress tests could have saved some banks.

“I think it could have made a difference,” he says. “Now, there are other things that went into it, like, the speed at which information or disinformation travels in the world of social media. That would not have been controlled, but I do think there could have been a positive impact to some of those banks that were impacted.”

What tech can help banks stress test?

To enhance stress testing and reporting capabilities amid a heightened regulatory climate, banks are turning to technologies such as distributed computing, orchestration engines and generative AI, according to Vikas Agarwal, a partner and financial services risk and regulatory leader at PwC.

Orchestration engines allow financial institutions to automate stress testing and standardize the stress testing process.

“This reduces the required time to conduct an enterprise-wide stress test from months to hours and enables banks to produce a richer set of ‘what if’ analyses for management’s consideration,” Agarwal says.

Meanwhile, banks need help tracking a large number of data models. AI model operationalization (ModelOps) helps organizations manage the AI and decision model life cycle.

“Applying leading ModelOps practices to streamline end-to-end model development, testing, and deployment processes in a highly controlled manner minimizes the risk of failures that can derail periodic stress testing exercises,” Agarwal says.

AI Models Help Banks With Stress Testing

 Generative AI holds the potential for banks to access detailed narratives and insights on the results of stress tests, according to Agarwal. Automated reports could deliver info on data discrepancies and anomalies and changes to model inputs since previous stress tests were run, he says.

Banks use AI because of the data-intensive nature of reporting to regulators, but not generative AI for sharing proprietary data, Heffner says.

“This is not the type of data you want to share with ChatGPT,” Heffner says. “It's the kind of data that is really about the function of your organization.”

Banks can train AI models within their own private network and do not share the data with other financial institutions, according to Heffner.

At the PwC Tech Showcase 2023 in New York City last month, Agarwal demoed an AI governance model in a PwC product called Model Edge and discussed how it could be used for bank stress testing during this economic climate.

Agarwal explained how regulators are increasingly auditing algorithms and data models. Banks have models around credit risk decisions and use AI algorithms related to sanctions, anti-money laundering and fraud, according to Agarwal. Regulators test the documentation of AI models for bias and fair lending.

Model Edge lets financial institutions build AI model starter kits and customize models, Agarwal says.

“If you think about the failures of the banks over the last couple of months, a lot of it was related to their inability to do more sophisticated modeling as it relates to interest rate risk, bond prices, understanding the effect on cash flow forecasting, liquidity risk management, and balance sheets,” Agarwal says. “We're actually developing models for several of our clients right now in the platform with more low-code techniques that can allow them to do this type of modeling and actually document the models.”

A low-code modeling environment offers enhanced reporting to allow banks to visualize and digest their forecasting results more easily, according to Agarwal.

“With a low-code environment, end users can spend more time analyzing model forecasting results and generating strategic insights for the bank’s management and less on preparing these forecasts,” he says.    

Banks are exploring how to document large language models in a way that can satisfy regulators, according to Agarwal.

“They have a lot of requirements for rules-based models, they actually haven't even caught up to machine learning models yet, and then they'll have to figure out what they're going to do with large language models,” Agarwal says.

Banks must document whether they have complied with regulatory requirements and whether models have undergone robust testing, Agarwal says.

“The largest banks in the world have the ability to look at hundreds, if not thousands, of economic factors and understand how they affect their balance sheets,” Agarwal says.

Agarwal notes that midsize banks have not invested in modeling capabilities as much as larger banks. However, Heffner sees more mid-tier banks having to perform stress tests.

How ML Algorithms Help With Audits

In another demo at PwC Tech Showcase 2023, Jon Swanson, managing director at PwC, showed a new product called Predictive Analytics, which uses machine learning (ML) algorithms to provide insight on a company’s revenue patterns over a period of time. The process is called regression analysis, which is a statistical technique that shows how variables relate to each other, Swanson explains.

Although Predictive Analytics could provide some insights to banks around volatility, the tool is focused on auditing, according to Swanson.

“The primary purpose is to provide audit evidence and reduce audit burden, while providing certain insights to management surrounding what may have been some of the key drivers of their revenue during the past several years,” Swanson says.

To create the insights, the PwC application called Predictive Analytics draws on a company’s historical monthly revenue totals for the past few years, Swanson says. It uses three ML regression algorithms to analyze and make sense of the relationships within the data.

The tool can help develop an understanding of the factors driving revenue to prepare teams to share insights with management, Swanson says.

What to Read Next:

InformationWeek's 2023 IT Salary Survey Results (Download)

From a Gargantuan GDPR Fine to More Hot Water for Meta

When a Bank Fails, Cyber Criminals Will Come Knocking

About the Author(s)

Brian T. Horowitz

Contributing Reporter

Brian T. Horowitz is a technology writer and editor based in New York City. He started his career at Computer Shopper in 1996 when the magazine was more than 900 pages per month. Since then, his work has appeared in outlets that include eWEEK, Fast Company, Fierce Healthcare, Forbes, Health Data Management, IEEE Spectrum, Men’s Fitness, PCMag, Scientific American and USA Weekend. Brian is a graduate of Hofstra University. Follow him on Twitter: @bthorowitz.


Never Miss a Beat: Get a snapshot of the issues affecting the IT industry straight to your inbox.

You May Also Like


More Insights