Not just ChatGPT: banks race to develop own AI tools

IFR 2482 - 06 May 2023 - 12 May 2023
11 min read
Asia
Natasha Rega-Jones

ChatGPT has opened the world's eyes to the power of artificial intelligence, popularising concepts previously confined to science fiction. For tech whizzes in investment banks, the buzz of excitement is confirmation of what they have long known: that AI tools are set to revolutionise the way their firms do business.

Even as banks ban employees from using external programmes like ChatGPT over security concerns, several have accelerated in-house projects focused on "natural language processing" – the AI that underpins these systems. The premise is simple: well-designed AI should allow investment banks to deploy the vast reams of data in their systems more efficiently, whether that's in pricing securities or interacting with clients.

Putting that into practice is far from straightforward, though, with concerns over the accuracy of data being fed into banks' machines remaining a major stumbling block. But while such issues are giving some firms pause, others are forging ahead with integrating AI across different parts of their business at a faster pace.

BNP Paribas says "execution bots" are now completing an average of 10 to 20 foreign exchange and repo trades a day for clients. At Deutsche Bank, executives say "experimental use" of NLP across the bank has increased "by an order of magnitude" in the last 12 months.

"It's very clear that everyone is racing against each other to come up with and deploy new use-cases," said Tim Mason, head of Deutsche's Innovation Network. "However, the question is how fast you can go in a safe and ethical way."

NLP is a form of artificial intelligence that uses algorithms to find patterns in written or spoken language – data typically referred to as "unstructured" due to its qualitative rather than quantitative nature. Banks began ramping up their investment in this form of AI around five years ago and have gradually applied the technology to a wider range of products across their market businesses – including interest rates, FX and equities.

Using NLP to support automatic trading through execution bots isn't yet widespread, but banks are finding numerous ways to deploy NLP more broadly in their sales and trading divisions. That mainly involves collecting and analysing information from internal sources, such as communications with clients and trading-related data, as well as external sources such as news stories.

Parsing this information carefully can allow banks to streamline their sales process and customise the content they offer clients. It can also alert their traders to information that could influence the direction of markets, allowing them to adjust pricing to increase profits or limit potential losses.

"Traders already make informed pricing decisions every day as they're connected to many news systems, but they can't read everything," said Laurent Carlier, a data scientist on BNPP's global markets research and data desk. "We are working on how NLP can give traders more news analysis based on what they trade and their market-making activity."

Revolutionary

BNPP has increased its use of NLP dramatically in recent years in its global markets business. Human salespeople aren't involved in pricing and executing most client trades across BNPP's flow trading business, with 90% of this activity being automated thanks to NLP. All interest rate product pricing requests from clients now run through the bank's NLP system compared with "closer to 0%" five years ago, said Joe Nash, BNPP's digital chief operating officer for FX, local markets and commodities. Most equity, FX, and repo requests also go through the same system, he said.

Processing most of the simpler flow trading business this way frees up salespeople to focus on more complex transactions such as structured products, which typically require more back and forth with the client. Salespeople can also use NLP to scan the troves of information that banks collect from client conversations about areas of interest and the research papers they read. That can help them pitch tailored trade ideas or specific products.

"Everybody within our global markets business now understands what NLP can do. That's a huge difference compared to five years ago where NLP looked more like a fantasy than something that could actually be a helpful tool in reality," said Carlier.

Increasing productivity

Investment banks are also exploring how NLP can increase productivity across other parts of their sprawling operations. That includes summarising the obligations banks face under the various text-heavy documents they deal with daily, such as procurement contracts and business loans. Deutsche is experimenting with NLP's ability to digest and summarise the data it receives within various different document types – including regulatory and legal documents.

"That's an incredibly manual process but NLP is able to streamline that work to a high degree of accuracy and within a much quicker response time," said Gil Perez, chief innovation officer at Deutsche.

Banks and investors are also using AI to unpack the wealth of information companies provide in the burgeoning ESG space. Among other things, that analysis can help investors get a better grip on how climate events such as droughts or floods could impact firms.

"NLP essentially helps portfolio managers to better manage climate change risks as they arise," said Marina Goche, chief executive of Sentifi, a fintech company that focuses on data analytics.

Silence isn't golden

The sheer amount of data now available has undoubtedly helped fuel the surge in NLP adoption across the finance industry. International Data Corporation estimates that the global "datasphere" is set to reach 175 zettabytes of digital storage capacity by 2025, up from the 100 zettabytes now and 30 zettabytes in 2018. (Today's most advanced smartphones have around one terabyte of memory, which is one-billionth the size of a zettabyte.)

The amount of unstructured data being processed and analysed across the world has doubled over the past five years to about 20% of all available data, according to Iain Brown, head of data science for the UK and Ireland at analytics firm SAS.

"In essence, over the past five years, there has been a significant increase in the amount of data available for organisations to leverage," said Brown, who added that around two-thirds of financial institutions are exploring or using NLP.

For now at least, banks remain wary of plugging this data into external programmes like ChatGPT. The potential for leaks of their proprietary information remains a major concern, while many are also critical of the quality of some of the external data fed into these widely available models.

ChatGPT made every executive think "how can I apply this in my business", said Deutsche's Perez. "[But] we must be prudent and thoughtful on how we deploy and use this technology. It takes time and know-how to successfully integrate such models within a highly regulated banking environment."

Trendy tech

Nonetheless, some bankers downplay how much AI is disrupting their trading operations. One fixed-income executive at a major US bank said every firm's technology department is looking at NLP, but machines aren't taking over trading desks just yet.

"[While] it's very trendy to say you use it, I don't think it has made a dramatic change to our business model yet," the US banker said. "But I'm confident it will. These are very powerful tools."

Another European bank has examined digitising 95% of its "request for quotation" interactions with clients in its fixed-income unit but has delayed rolling out the technology. This is mainly due to the 98%–99% accuracy rate of the bank's NLP model falling short of what executives deem acceptable.

"It's always a concern that gets raised during the proof-of-concept stage," said a source at the European bank. "There's a real burden to prove that an NLP application is accurate and will actually make a meaningful impact to the business," adding that his firm's use of NLP has barely moved in recent years.

A particularly thorny issue for banks is how to treat data originating from the outside world. Online forums such as Reddit were crucial to understanding the "meme-stock rally" of 2021, when swarms of retail investors suddenly pushed up the share price of video game retailer GameStop and other companies. More recently, traders have homed in on the role that viral posts on Twitter could have played in exacerbating panic around some banks.

"We're living in an incredibly volatile world and keeping up with digital chatter as it's evolving is an essential part of ensuring you're actively monitoring risk within the assets you're trading or managing," said Sentifi's Goche.

But determining the reliability and relevance of external data is tricky. Relying more heavily on internal sources of information – such as communications with clients or transaction data – is one way to address concerns over accuracy. SAS's Brown said financial institutions typically used 70%–80% internal data for NLP models, though figures can vary.

"Discerning what data is real and what isn't is more challenging today. So trusting data is also more challenging," said Brown.

For Perez, NLP expansion goes hand in hand with the finance industry's growing adoption of cloud computing, which he believes improves banks' capabilities to index and analyse data. Deutsche has partnered with cloud providers Google, Microsoft, and Nvidia to train and experiment with NLP models.

"To be successful, we must partner [with these firms] to learn how to leverage the latest tools, while also educating the providers of such very powerful tools on the strict regulatory requirements of the banking industry," Perez said.

06G1 graph