Balyasny Asset Management
This story is part of a series highlighting top technology projects on Wall Street.
The AI team at Balyasny Asset Management has been vocal about its ambitions to build an AI equivalent of an analyst, and a recently developed tool called Deep Research is getting it one step closer to its goal.
Built by the hedge fund's Applied AI team, Deep Research helps analysts and portfolio managers answer complex questions to research stocks before making a trade. They also use it to gauge the impact of global market events on a portfolio or set of stocks.
In one recent example, a portfolio manager asked Deep Research to find companies whose supply chains are impacted by tariffs. The tool scanned more than 20,000 documents to identify 120 companies with potential exposure, and provide a report with explanations for each company — all in about an hour.
The goal is to automate parts of the research process, "from days and weeks to minutes and hours," Chen Fang, the data and analytics lead on the Applied AI team and the lead executive behind Deep Research, told Business Insider.
The Applied AI team wants to continue to level up what the bots can do. Instead of just summarizing and linking back to raw documents, the team one day hopes to feed portfolio managers trade ideas, Fang said, much like a senior analyst would do for their PM. Balyasny has supercharged its generative AI tools by training the models to have context on words commonly used in finance and markets. According to the firm, its retrieval system surfaces the most helpful document for a query 60% more frequently than OpenAI's models, which are trained on general-purpose data rather than Wall Street-focused info.
"There's a sort of inflection point where these things are on the edge of being possible and it's a bit of a gold rush to try and do it first," Peter Anderson, head of research on the Applied AI team, told BI.
Currently, Deep Research is in beta across roughly 50 investment teams. Those teams experiment with Deep Research by sending research questions to the Applied AI team to process. The goal is to release the tool firmwide by the fourth quarter, with all 170-plus investing teams accessing the tool directly through an intuitive platform without the need for the Applied AI team to act as a middleman.
Imagine an equities analyst wants to find out what sort of guidance Microsoft has given around its expected earnings, and whether it's beaten or underperformed against that guidance over the last five years.
"Conceptually, pretty simple query," said Anderson, who joined the hedge fund in March after nearly four years as a senior research scientist at Google. But the reality behind the scenes?
"Here's 5 million documents and the information's in there. That's the kind of challenge we have," Anderson said.
The tool pulls in info from a database of about 5 million documents, like SEC filings, earnings transcripts, third-party research and market data, and Balyasny's internal analyses and memos. Depending on the complexity of the question, Deep Research utilizes hundreds or thousands of documents.
But the way the models retrieve this info is different than many of the off-the-shelf models, like OpenAI's ChatGPT, which are not adapted to financial jargon, like ticker, hedge, options, and other words that can have completely different meanings outside of a financial context, according to Anderson, who is the lead exec behind the document-retrieval system.
Balyasny researchers trained the open-source model on financial terms and commonly asked research questions. Making these models financially literate has improved the rate of the AI surfacing the most helpful document to 60% more frequently than without this financial training, according to the firm.
The improved performance has enabled Balyasny to perform deeper analysis and now underpins some of the hedge fund's other generative AI tools, including Deep Research.
