Will generative AI crack the code for bank tech teams?

One of the major difficulties introducing generative artificial intelligence (AI) into the heavily regulated financial sector is ensuring those who use it have suitable expertise to understand its risks.

That requirement means one particular use case could be easier to adopt: assisting coding teams who already have a good grasp of algorithms and their limitations. Risk.net, a sibling publication of WatersTechnology, has spoken to more than 20 sources and determined that at least four global systemically important banks (G-Sibs) have put testing tasks in the hands of their existing software development and technology teams. They play a key role in building and maintaining the infrastructure for trading systems, algorithm development, and risk management.

“We want to gather feedback from experienced coders before determining the optimal and safe way to roll it out to the other parts of the front office,” says a senior model developer at a global bank.

While the pace varies across banks, there are two key areas that most large firms are actively exploring. First, there is a push to apply generative AI to understand and translate legacy or proprietary code. Several projects in this area could go live during 2024. The second focuses on code improvement and generation, which aims to enable modelers and quant professionals to write algorithms and build models more rapidly. Though this is expected to take longer to enter into production due to the greater risk management controls required, most sources agree it can be a reality by 2025. 

“It’s true that using generative AI for coding comes with risks, but the bigger risk lies in not adopting it,” says an operational risk manager at a second global bank. He suggests banks that are hesitant to explore now will lose their competitive edge as the technology evolves to support modeling functions in the next one or two years.

However, the need for speed of innovation doesn’t negate the equal importance of designing a safe coding workflow to validate and deploy the technology. 

It’s true that using generative AI for coding comes with risks, but the bigger risk lies in not adopting it

Operational risk manager at a global bank

“We are spending as much time, if not more, on building the safeguards so that we can have a high level of confidence that generated code is accurate,” says a senior technology innovation executive at a third global bank.

The long and winding code

It’s no secret that many large institutions still rely heavily on legacy code written anything up to half a century ago, which is often poorly documented, vulnerable to security threats, and costly to modify. Moreover, knowledge of these codes can walk out the door as experienced programmers gradually leave the workforce. At least two large US banks had to hire back retired employees over the past year to sustain their systems.

While banks have failed to find a viable long-term solution for years, the advent of generative AI brings a ray of hope for the future. This year, tech companies like Google, Microsoft, and IBM have successively rolled out generative AI-powered tools for code interpretation and translation (see box: In-house or third party?).

Google’s Codey model, for instance, can understand more than 30 programming languages, including Java, C++, and Python. If properly fine-tuned, it can also explain and translate companies’ proprietary programming languages. Meanwhile, IBM recently released its code assistant tool to facilitate faster translation of Cobol to Java, while also actively exploring the possibility of extending the tool to support more languages. While Cobol is not generally used by trading desks, it is present in many financial accounting and reporting systems.

“Cobol is probably 50 years old and runs on IBM mainframe computers,” says Arthur Rabatin, a former front office and risk technology executive at several global banks. “There are fewer and fewer people in the world who can work in Cobol, so translating it into a modern language such as Java will be helpful—not to improve the system but just to make it maintainable again.”

Generative AI has proven its capability to fulfill this role, says Warren Barkley, senior director of product management at Google Vertex AI. It’s up to the banks whether and how they want to leverage it in their own environment.

Sources suggest that for old yet well-defined programming languages like Goldman Sachs’ Slang—the proprietary language that underpins its SecDB pricing and risk system—generative AI’s translation capability can play a valuable role by automating any adaptation process. Goldman declined to comment on this article, but it is understood that the bank has been looking for solutions for several years to move away from Slang, which can otherwise deter young coders who would rather work on mainstream programming languages.

A technology executive at a fourth global bank says Slang was invented 30 years ago, but it is an “extremely smart” language, and it has the advantage of being immutable—so errors cannot be introduced into existing lines of code. In addition, Goldman has updated it over time.

“With such a good foundation of code bases, it is relatively easy to use generative AI to automatically translate them to Java or whatever other modern languages Goldman preferred,” says the technology executive.

Errors kept in translation

That said, not every bank has the same quality code base foundation as Goldman. Most banks, instead, struggle with poorly managed legacy code systems, characterized by code written with flawed logic and left unmaintained for years. Since generative AI tools for code translation are typically designed to perform a syntactic conversion, they fall short in improving the code quality.

“Translation will work, but it will not improve the system as a whole,” says Rabatin. “It does not change the architecture, and it does not fix any issues that might be problematic with the code.”

Translation… does not change the architecture, and it does not fix any issues that might be problematic with the code

Arthur Rabatin, former risk technology executive

Where generative AI could add even more value, he suggests, is in interpreting and explaining older codes, especially if those who wrote them have long since left the building.

“When we examine poorly written legacy code, the most common question is often: ‘What is it even doing?’” says Rabatin.

Once coders understand the legacy code, they can then better decide how to rewrite it. Or they can create a wrapper to isolate it from the modern environment, which is considered a more efficient approach than overhauling the entire system, Rabatin adds.

New model coder

While it may take longer to approve generative AI adoption for writing fresh algorithms, modelers are excited about the potential. Banks are still debating how to venture into this realm, but some quant developers at less heavily regulated hedge funds have already leveraged the coding capabilities of generative AI to refine their high-frequency trading algorithms.  

This process, known as code refactoring, focuses on improving existing code that is not well-optimized for performance due to inefficient coding styles or inappropriate use of variables. The improved code is capable of faster decision-making, explains Arpit Narain, global head of financial solutions at model and AI vendor MathWorks.

For example, code refactoring can reasonably enhance the performance of statistical arbitrage trading algorithms. It does this by constructing more efficient computational methods and simplifying the code to streamline complex mathematical models, such as regression and cointegration (a technique to find correlation between time series). This not only speeds up calculations but also reduces complexity. 

“Tapping into these refinements offers an edge in trading strategy effectiveness for the hedge funds,” Narain says.

You can feed LLMs all your internal CCAR documents, even books on CCAR, and ask them to help you generate code and improve your models

Operational risk manager at a global bank

While regulation makes banks more cautious about adopting the new technology, the sheer quantity of supervisory documentation that bankers need to digest also strengthens the case for applying generative AI to the task. The operational risk manager at the second global bank gives the example of using large language models (LLMs) to integrate the regulatory specifications for stress tests such as the US comprehensive capital analysis and review (CCAR).

“You can feed LLMs all your internal CCAR documents, even books on CCAR, and ask them to help you generate code and improve your models,” says the operational risk manager. “There’s no doubt it will have this capability soon.”

Manageable risks

A senior risk manager at a fifth global bank notes that banks have been much more cautious than hedge funds or fintechs about embracing generative AI for writing trading or risk management code. 

“But in reality, these risks may not be as significant as they imagine, especially if they put humans in the loop,” says the senior risk manager.

Indeed, the operational risk manager at the second global bank suggests treating LLMs similarly to traditional banking tools like Excel, viewing them as aids to employee productivity rather than a replacement. 

“Ultimately, it remains the responsibility of coders to review and ensure the accuracy of each line, with the model risk team taking on the responsibility of validating the final models,” says the operational risk manager.

Several large banks are exploring additional tools that can assist developers in reviewing and validating the code generated by LLMs.

“We always want the developers to be at the center of the control, with the support of different tools,” says the technology innovation executive at the third global bank. “If we can come up with an effective tool to ensure thorough testing of code against the best practices, we will be more comfortable rolling out generative AI coding use cases across the organization.”

However, another big challenge lies in adjusting the developers’ own skill sets. Those who are accustomed to do programming based on specifications now need to employ a chain of reasoning and thought process when working with generative AI to enable more efficient code generation and improvement, says Sankar Virdhagriswaran, head of AI services at EY.

One G-Sib grants access to 50 quant professionals with varying levels of coding skills to assess how effectively the tool can help with modeling across the team. The result shows highly skilled coders find the tool extremely helpful, whereas those with lower skill levels do not.

To address the challenge, sources say it is valuable to implement systematic training across teams on how to prompt LLMs, with highly skilled coders sharing their experience. That said, most banks have not yet reached this stage.

Proper training can also mitigate the risk of complacency, says the senior model developer at the first global bank. His bank has granted 500 developers access to Microsoft Copilot. After developers input a line of code, it will generate three suggested options for the next line. In such instances, if not properly trained, employees “may just pick whatever is being recommended without thoughtful considerations,” he says.

Another important consideration revolves around how banks want third-party vendors to handle their coders’ selection among prompted snippets of code. This is a matter that banks should clarify before signing contracts. 

“Should we allow Microsoft to learn from our developers’ choices?” asks the senior model developer. The benefit is a more refined model, but there are also drawbacks. Most notably, competitor banks would be able to harness the improvements in Microsoft’s own model that have been made by learning from the bank’s internal data.