- AI Street
- Posts
- How Quants Are Evolving with LLMs
How Quants Are Evolving with LLMs
Matt Robinson
April 06, 2025
Hey, it's Matt. Welcome to AI Street Markets, where I highlight AI investing tools. Was this email forwarded to you? Sign up below:
QUANTS
Quant Investing Meets Generative AI
When will human traders face off with “thinking” machines?
Made with ChatGPT
I've spent the last few Markets editions diving into specific tools (if you’re new or missed a few, links to those are below.)
This week, I wanted to step back and look at the bigger picture: where the technology stands today—and where it might be headed.
What makes the current LLM-driven AI boom different from past waves of AI is its fundamentally new approach to the problem.
LLMs are sometimes dismissed as ‘pattern-matching’ machines. But with enough data, how a prediction is generated matters less than whether it’s useful.
For example, without using physics, an AI model’s weather forecasts are on par and sometimes better than traditional systems used by expert meteorologists, according to a new paper in Nature. Instead of relying on physics equations, it learns patterns directly from weather data.
For finance, currently, most of the AI tools are tailored for fundamental value-oriented investors, the Warren-Buffett types.
But over the last couple of weeks, I’ve covered more on how traditional quantitative hedge funds, which rely on rules-based algorithms, are using probabilistic LLMs. For example:
Man Group’s AlphaGPT is using LLMs to help researchers generate new trading strategies.
Permutable AI is training models by having expert traders provide feedback—effectively teaching AI to reason like a human.
These are very different worlds. Traditional AI is fixed. The output is the same each time. LLMs give you similar but different answers each time.
So, it’s an interesting combination to me.
To get a better sense of where this is all going, I reached out to Bokai Cao, co-author of the paper From Deep Learning to LLMs: A Survey of AI in Quantitative Investment.
The paper, published last month on arXiv, tracks how quant investing has evolved from traditional rule-based models to deep learning to now LLMs. The research makes clear that most of the industry is still in the early stages of adapting this technology for investing.
From Deep Learning to LLMs
Cao told me that LLMs are already “excellent” when it comes to parsing unstructured text—earnings calls, regulatory filings, research notes, and other messy inputs that financial analysts deal with daily. That tracks with what we’ve seen across the space.
The next generation of quantitative investing is automated, interpretable, and knowledge-driven, practicing the philosophy of “end-to-end going all-in on AI” and “AI creates AI” by incorporating state-of-the-art automated technologies. LLMs have brought a glimmer of hope in this direction, but there is still a long way to go.
It’s become routine for firms to run LLMs over transcripts and reports to extract sentiment, surface red flags, or isolate specific disclosures.
Where things get tricky is when you ask these models to reason with numbers.
“Numerical reasoning is not a strong suit,” Cao said.
This makes sense. LLMs are probabilistic. You don’t want different answers for the same calculations.
But these models are improving, now able to write deterministic code to solve calculation problems when asked. So what was a problem a year ago is mostly solved.
Right now, LLM-based agents are showing up most often as copilots in quant workflows—helping with code, research documentation, and idea generation.
“For well-defined tasks like factor mining,” Bokai told me, “we’re already seeing meaningful integration in institutional settings.”
Over the next three to five years, Bokai sees AI evolving into three directions:
Automation: “We are moving from manual, labor-intensive modeling and deployment workflows to an end-to-end paradigm where AI creates AI—making the process more economical, efficient, and scalable.”
Explainability: “The industry is shifting away from black-box deep learning models—often criticized for overfitting and lack of transparency—towards models with stronger interpretability, especially in terms of explaining trading returns and risk.”
Knowledge-driven AI: “We’ll see a transition from purely data-driven modeling to knowledge-driven AI, where financial domain knowledge and logical reasoning are more deeply integrated into model design.”
But for more complex problems like portfolio optimization, he thinks it’ll take another 3–5 years before agents can outperform existing institutional systems.
Before LLMs, the idea that a robot would be able to “think” about how to trade on Wall Street was science fiction.
Now, it could happen by the end of the decade.
ICYMI
Check out the last few editions on using AI for investment analysis,* creating customized news feeds and tracking earnings call mentions:
*Not investment advice
PROGRAMMING NOTE
AI Street Markets is adopting a twice monthly publication timeline, so expect new editions on the first and third Sunday of the month.
If you have an idea or are interested in a platform for an edition, reach out: [email protected].
See you April 20.
How did you like today's newsletter? |
Reply