J.P. Morgan Promises Its Fundamental Portfolio Managers an AI ‘Co-Pilot,’ Not a Boss

“We teach our analysts what generative AI can do and what it cannot do. When you tell people about those limitations, they recognize that ‘ok, I’m still the decision maker here,’” says J.P. Morgan Asset Management’s Arezu Moghadam.


Illustration by II

During a demo, Arezu Moghadam, global head of data science at J.P. Morgan Asset Management, shared her screen to show exactly how a generative AI tool can break down a complicated, plain-English query from a hypothetical analyst about a small-cap company. Seconds after the questions are submitted, text appears on the screen showing the steps the system is taking: determining the documents it needs, finding them, going immediately to the relevant sections, and identifying the most important issues — in this case, eight bullet points raised after an earnings call — and comparing that to what was discussed in sell side research and whether the reaction was positive or negative. The system then spits out a final commentary about everything that was learned through the analysis, along with the original sources that were mined for the information.

“Analysts can come here and get the first draft of a research note that they can read, edit, and add their own comments to,” said Moghadam, who joined J.P. Morgan in May 2022 from hedge fund Point72.

JPMAM wants analysts and managers “to be content editors, rather than content creators, which is a very time-consuming task for them,” she added.

The tool Moghadam screen shared, which is still being tested and evaluated, is called SpectrumGPT and leverages Open AI’s GPT4 within J.P. Morgan’s cloud environment.

Earlier this year, the asset manager deployed data scientists to work alongside a team of equity portfolio managers and learn enough about what goes into their investment decisions to build artificial intelligence and machine learning tools, including what JPMAM describes as an AI coach, to help them.

Moghadam is adamant that she didn’t want to build tools for managers somewhere off in a closed room. “We said, ‘let’s not make any assumptions about what’s good for fund managers. Let’s just get closer to them.” Getting the portfolio managers and research analysts involved in designing the tools from the beginning also helped mitigate fears — if they had them — about the technology taking their jobs. “If we do it in isolation and we just think this is good for you, they’ll just ignore it and never use it. That’s the risk in AI adoption,” she said.

Moghadam, who started her career at Goldman Sachs and has a Ph.D in computer sciences and math, has a history of working with AI and machine learning — but mostly with quants. The job at JPMorgan gave her the chance to work with fundamental managers, who aren’t using algorithms to find patterns and trends in data to make investments. Instead, they’re thinking about a company’s strategy, the quality of its management, R&D, patent pipelines, and other real-world business activities to identify stocks.

So Moghadam’s data scientists moved in with the portfolio managers and analysts to understand what a good outcome would be (aside from high returns), their data sources and how they used them, their models — if any, and how the managers came up with insights and made decisions.

It took the data scientists two to three months to learn what they needed to develop SpectrumGPT. Now the firm is testing it with stock managers, but plans to do something similar for bonds and multi-asset portfolios. Ultimately, JPMAM plans to develop AI tools for private equity, venture capital, real estate, and other unlisted assets, even though the availability and quality of data that generative AI depend on is not as rich as in stocks and bonds.

“Spectrum can analyze massive amounts of unstructured data, millions of documents, decades of proprietary internal research, and connect the dots so portfolio managers can distill insights and information from this data in a matter of seconds,” Moghadam said.

Critical for JPMAM was minimizing hallucination, a common problem with generative AI where it comes back with irrelevant conclusions. First, it’s limiting the documents the engine can access, called context learning, and preventing the tool from finding answers anywhere on the internet. It’s also using two proprietary methods to reduce hallucination that Moghadan declined to describe.

Similar to other industries, asset managers have a range of views on AI. Coalition Greenwich found that less than half, 41 percent, of almost 100 asset managers surveyed in September believe that AI will be a “game-changer” within the next one to two years. But over three to five years, four out of five managers think AI will be critical to success. Under 20 percent, however, think they’re ready to take advantage of it. One of the barriers is cost, something that large managers like J.P. Morgan can more easily afford. JPMAM also has access to its parent’s 900 data scientists and 600 machine learning experts.

SpectrumGPT is a building block for what JPMAM calls Smart Monitor and PM Money Ball. Smart Monitor helps investors keep on top of research, news, earnings calls, and other events that affect their positions as well as the thousands of stocks in their investable universe.

Moghadam compared Smart Monitor to Spotify, which learns what music people like and offers up a list of songs culled from the millions available. Every time a user gives a song a thumbs up or down, Spotify gets smarter about their tastes. It’s the same with Smart Monitor. “Over time it learns what investors care about and helps to create prioritized and personalized company insights,” she said.

If smart monitor surfaces information, PM Money Ball helps managers make decisions. The application, which Moghadam calls a “co-pilot,” can pull up information about a manager’s past buys and sells, for example, after positive and negative earnings surprises. “This is how they behaved post that guidance; this is how companies in my benchmark performed. Perhaps I could take this other action to monetize that event better. It’s a feedback loop for the PM,” she said.

The asset manager believes generative AI is mostly a tool to help with repetitive and mundane work — at least as this point.

During the demo, Moghadam said despite how useful it looks, AI is just looking for patterns.

“We teach [our analysts] what generative AI can do and what it cannot do. When you tell people about those limitations, they recognize that ‘ok, I’m still the decision maker here, I’m still the accountable person here.”