Balyasny Asset Management is working to build the AI equivalent of a senior analyst by combining multiple different AI tools, the hedge fund’s head of applied AI told Business Insider.
The $21 billion hedge fund has been building its own souped-up version of ChatGPT, which it made available to all employees in June 2023. Balyasny is working to plug BAM ChatGPT, as it’s called, into every internal and third-party dataset at Balyasny. The tool, built using OpenAI’s API and hosted on Microsoft Azure, can currently tap into 10 different sources, including transcripts, sales and sell-side commentaries, and broker research.
The goal is to create a collection of bots that can proactively push relevant information — like breaking news events on relevant companies or flagging differences in company disclosures — to PMs and other business teams.
“We’re very focused on proactive insights. How do we actually bring these capabilities in a way that we’re actually doing the analysis, in some cases in advance of folks even understanding they should be asking the question,” Charlie Flanagan, head of applied AI, told BI. “We’re pretty laser-focused right now to move from junior analysts to senior analysts,” Flanagan said of the agents’ abilities.
Hedge funds are keen to capitalize on the AI wave sweeping up Wall Street. Many firms are using AI for productivity gains. Software engineers at Man Group use it to code, while Two Sigma researchers use machine learning to analyze hundreds of market variables. Bridgewater, meanwhile, is building an AI investor that will make trades with client money.
Using AI to analyze data in new ways and automate tasks for analysts
AI is already changing how Balyasny’s investment teams work by making it easier for analysts to discover nuggets of information within the fund’s hundreds of datasets and research and save time by automating time-intensive tasks.
In one case, the head of research for one of BAM’s investing strategies and his team built a series of AI bots to shave down the time it takes them to put together an in-depth monthly analysis of a recurring market event. The process went from two days to 30 minutes, Flanagan said.
Analyzing regulatory filings, like 10-Ks, for differences in risk or legal disclosures is another area ripe for AI. Today, analysts use specialized software to conduct such analyses, and it’s a tedious and time-consuming process, Flanagan said. Instead, AI could automatically identify the differences in the disclosure of a particular company as soon as a filing is published and then proactively push that information via Slack or email to the analysts who cover that company. The bots could even be primed to answer follow-up questions the analyst may have.
Then, there are many internal teams responsible for generating so-called morning notes, which provide sector updates on what happened overnight, what’s relevant, and what news and other data sources are saying. “We can do that better, faster, quicker for folks when that data becomes available,” Flanagan said.
These agents aren’t designed to replace analysts completely — just about 10% of their workload. Flanagan said he hopes the AI will turn analysts into editors, rather than creators, of these analyses and summaries.
As Balyasny looks to integrate AI into employees’ jobs, Flanagan’s team has grown to 10 people in the last few months. The team includes researchers, data scientists, and engineers from both Wall Street firms like Citadel and Goldman Sachs, as well as Silicon Valley.
A closer look at BAM ChatGPT
Rather than having one general chatbot that can perform many different tasks, Balyasny’s strategy is to build very specific “agents,” which essentially are bots that have very specific tasks or mandates, and combine several of them to get a richer output. Teams are breaking down the work — whether it’s analysis or sector summaries — into several, very specific sub-tasks, building individual solutions for each of those, and then combining it, “which is the key to unlocking that productivity gain,” Flanagan said.
Looking ahead, Flanagan is encouraging analyst and PM teams to build their own “agents” using AI building blocks created by the Applied AI team. Some investment teams have already begun making their own agents and have discovered intel “leading to profitable traits,” Flanagan said, such as comparing third-party broker research comments about company challenges with what company management says about those challenges during earnings presentations.
So far, BAM ChatGPT has about 80% adoption across Balyasny’s roughly 2000-person workforce, Flanagan added.
Balyasny is also working to improve the models’ ability to understand numbers, graphs, and charts, something the current OpenAI GPT model is not advanced enough to compute.
“As you might imagine, we have huge corpuses of broker research and all sorts of other reports that we get in PDF and other forms. A lot of the valuable information there is hidden within table format or visual formats,” Flanagan said.