Wall Street Firm Citadel Securities Courts AI Startups for Trading Edge
Citadel Securities, a prominent high-frequency trading firm, has told startup founders and investors it is looking to license software from artificial intelligence startups working on large-language models or to buy stakes in them, according to two people with knowledge of the conversations.
The discussions highlight how interest in the latest wave of AI advances has spread to Wall Street, where firms have long battled to eke out gains by making price predictions using vast quantities of data. Citadel Securities hopes to use knowledge gained from relationships with AI startups to potentially develop new trading strategies, one of these people said. LLMs, which power ChatGPT and its ilk, can allow developers to build software that analyzes a wide range of data, including corporate financial statements.
To build those startup ties, Citadel Securities hosted a dinner on May 18 at Selby's, a restaurant near Atherton, Calif., in Silicon Valley. Attendees included representatives from startups such as ChatGPT creator OpenAI and venture capital firms such as Insight Partners and General Catalyst, the people said. Those VC firms have led key investments in startups involved with LLMs.
Citadel Securities is looking to invest in AI startups alongside Sequoia Capital, one of these people said. Sequoia and Citadel Securities have a history: Sequoia took a stake in Citadel Securities in 2021, and last year they both invested in cryptocurrency exchange EDX Markets, according to PitchBook. Sequoia Capital has bought shares in LLM startups such as OpenAI and Harvey, which uses LLMs to analyze legal documents.
Spokespeople for Citadel Securities and Sequoia Capital said in a joint statement for this article that they are “not currently working together on AI investments.” The Citadel spokesperson, David Millar, said the firm is “in talks with developers and researchers working in generative AI about potential strategic or commercial partnerships, and is currently identifying opportunities to use AI across the organization.”
Hedge funds and other investment firms have been experimenting with LLMs to automate parts of their operations, including summarizing earnings calls and market research, as well as writing software code.
The hype around LLMs also has caught the eye of Miami-based Citadel Securities, one of the world’s top market makers, a kind of middleman that pays brokers like Robinhood to buy its customers’ stocks and options rather than routing them to public exchanges. Citadel Securities then usually sells those securities at a premium. The firm generated $7.5 billion in trading revenue last year, Bloomberg reported. It is the sister company to Citadel, a large hedge fund. Ken Griffin founded both firms.
Still, Griffin spoke this week about his belief that tech executives and investors had created too much hype regarding the technology. He said the models still had trouble with accuracy, which was a particularly serious problem in fields like finance and law. “Here’s the problem with large-language models: They are built on the past. Everything we do is about the future,” Griffin said, according to CNBC. “We are at the start of the journey of large-language models.”
‘Better Prices’
Peng Zhao, Citadel Securities’ CEO, said publicly in May that his firm was exploring ways to use LLMs, which train on a massive amount of text and other data to learn the nuances of human language. Businesses are turning to them to automate tasks involving writing, spreadsheet building and software coding. The technology “allows us to change how we organize info, query knowledge and communicate with each other,” Zhao said at the Milken Institute Global Conference in Los Angeles.
The firm is still trying to figure out how the technology could improve its trading, he said. That could be tricky due to the tech’s limitations as well as regulatory considerations.
Zhao added that most of Citadel Securities’ data science work involves numerical data that
powerful new language models may have little impact on. But, he added, “to the extent we’re taking in some language inputs already, we can do so more efficiently and in larger quantities that will help us come up with better prices,” he said.
LLMs have some clear potential uses for traders, said Chris White, CEO of BondCliq, a provider of data for fixed-income traders. He said traders of mortgage, corporate and municipal bonds often sift through reams of “unstructured data” about the assets to figure out at what prices to buy and sell, a process that LLMs could simplify by quickly making sense of the data.
Citadel and other “firms that have done really well with systematic trading are constantly focused on what’s the next frontier,” he said.