AI In Finance: Wharton’s New Lab Explores The Opportunity & The Risk

On May 6, 2010, the Dow dropped about 9% in 36 minutes before humans intervened and halted trading. What caused the crash? Automated systems selling stocks, triggered by other automated systems. It was an endless loop taking down a stock exchange.

Back then, this type of automation was what was normally referred to as artificial intelligence or machine learning. People programmed a machine and it went about like an automaton. Usually, that worked well, freeing humans from drudgery. This time, it didn’t.

AI is one of the earliest use cases in finance; entire fields and firms exist and there’s even a term for somebody versed in it, quant trader.

While there are doom-and-gloom warnings about generative AI, the latest type we’re familiar with from ChatGPT and Claude, there’s more concern about the effects of those systems on the financial market. In February 2025, Federal Reserve Vice Chair Michael Barr gave a speech about this specific danger.

In response to new AI opportunities and threats, Wharton has created an AI Lab for finance. Led by Winston Wei Dou with help from professors Roberts and Goldstein and others, his lab is exploring how generative AI can be used in finance, the upside and the risks. Below, we had the privilege of speaking with Prof. Dou about his lab.

P&Q: Let’s start with your background. How did you get here?

Dou: I’ve had a somewhat non-traditional path. I have two PhDs. I studied pure mathematics in college. I was fascinated by mathematics as a way of understanding the universe. Later I added a degree in economics and found that some very intangible but important concepts in macroeconomics and financial markets could be analyzed through mathematics, through equations, and they worked surprisingly well.

When I applied to do a Ph.D. in the United States, my advisors steered me toward statistics and data science. They said if you want to apply mathematics to financial markets, that’s something you need to master. So my first Ph.D. research was in machine learning, back when we didn’t call it AI yet. I did some quantitative trading, and I was totally fascinated by it.

But I wanted to think more deeply than just how to maximize profits every day. I could see that these technologies, while making money, were also constantly reshaping how financial markets work. This is my tenth year teaching at Wharton.

P&Q: Where are we headed? Tell me about the lab.

Dou: The lab is an exciting initiative within Wharton’s broader effort to incorporate AI into education and research. Our goal is to create a place where faculty, students, and practitioners can study how AI is changing financial markets and help shape its development.

We connect four areas. First, research. Second, building AI-ready data and infrastructure. Third, bringing that research and infrastructure into education, to support our teaching. And fourth, engaging with industry.

On the research side, we study how autonomous AI systems learn to trade, develop new trading strategies, manage portfolios, and work through complex financial decisions that would otherwise be extremely difficult for humans. We also examine what happens when those AI systems start interacting with each other, learning to coordinate, and reshaping how markets function.

One initiative that serves this purpose is what we call a synthetic market framework. We’re developing a financial market simulation, grounded in financial principles, where AI agents and other market participants can learn and adapt to one another, mimicking the behavior of real markets. I think of it as a financial wind tunnel. It lets us stress-test strategies and explore new technologies and regulatory tools that might change market behavior.

P&Q: On these interviews I’m usually asked whether I have any negative questions, and they fall into two buckets: students who think professors aren’t up to speed on AI, and professors who get 30 essays back that all read exactly the same. But you’ve raised something very different. Everybody remembers the flash crash, when the systems went down and the tickers went blank. The Federal Reserve has already said it’s concerned about new AI systems. You mentioned coordination. Is the lab working on safety, on preventing that from happening?

Dou: Yes. That’s part of why we felt it was urgent to establish this lab. Finance is a particularly important place to understand AI, because financial decisions affect almost everything: where capital goes, who gets credit, how risk is shared.

At the same time, finance presents a particular challenge. Markets react to the technology we introduce. An AI system may find a very profitable trading strategy for its user, privately. But if many people start using AI, the AIs start to interact, and they may create harmful market-level consequences.

One important message from our research is that higher profits for AI traders don’t necessarily mean better markets. In our models and simulation experiments, we showed that trading algorithms can learn collusive behavior without communicating and without being programmed to collude. They can weaken competition, reduce liquidity, and compromise how informative prices are.

More importantly, in related work we showed that AI agents with planning capacity, which is a core technology in current agentic AI systems, can learn to coordinate their learning and their trading in a way that manipulates the market, creates bubbles and crashes, and profits from the chaos. That’s why we need to look at how these systems interact together, not just how each performs individually.

P&Q: That seems vitally important. We’ve all heard stories about AI turning into criminal robots. What seems more likely is that somebody says “maximize my profits” and the system destroys a fund or crashes a market. Your lab wants to make money, it’s a finance lab, but also to do it safely and responsibly. How do students react to that, and how is it taught?

Dou: In my new course, AI and Data Science in Finance, I want students to really understand the architecture and the principles behind AI and machine learning models, and then connect that knowledge to financial decisions and to the safety concerns we just discussed.

I want students to get past the simple idea that AI is a magic box, that you put a question in and the answer comes out. It’s not like that. So I spend several lectures on what’s called AI reasoning, including frontier techniques like Chain of Thought and Tree of Thought, and how to use them to build financial reasoning agents. I want students to understand how these systems are put together, so they can judge whether the conclusions actually make financial sense and whether they’re responsible for the market as a whole.

We go deep into the fundamentals: transformers, reinforcement learning, how they work together, what consequences we see when these systems interact with each other, and how to prevent those consequences.

P&Q: Can you give a few examples of what the lab experiments with?

Dou: One thing I push students to think about is what it takes to use AI responsibly, and I emphasize a few practical questions. The first is: who is responsible for the system? By design, these AI systems are self-learning and autonomous. They can do things beyond the designer’s original intention. That becomes a philosophical and legal question. Who is responsible for this AI?

I use my own research as a concrete setting. Imagine a hedge fund that adopts an autonomous, self-learning trading system to make profits. Once deployed, those AIs may behave in ways beyond the initial intention and create market-wide harm. In that case, who is responsible, who is not, and to what extent?

Then I tell students: if you go into practice and develop an AI-driven trading system, you absolutely need to test thoroughly before deployment. In class we work almost hand in hand on how to build an experimental sandbox, how to design it to cover the potential risks AI creates, and what patterns to test for, the patterns that show the AIs are working together in ways they shouldn’t. That’s what I teach, and I’ll teach it again this year.

P&Q: How do we balance human judgment with AI, especially in financial markets, where things move very quickly?

Dou: This goes back to a point I keep emphasizing in my lectures. Don’t think of generative AI or large language models as simple input-output machines, where you throw in a question, get an answer, and trust it 100 percent.

Students need to fundamentally understand how to interpret the outputs. It’s a probabilistic output. And the machine was trained, by its nature, to make humans feel happy, not necessarily to reveal the truth. Even the reasoning you see as output isn’t the true reasoning; that’s hidden by OpenAI, by Anthropic.

So we need to design our own reasoning tools around the model. That’s what I teach with Tree of Thought and Chain of Thought: break a big question into a sequence of smaller questions and work with the AI step by step. At each step, as humans, we check for consistency. Is this answer consistent with the answers to the previous sub-questions? We build that sequence outside the model, so we have a systematic way to guide it toward reliable answers.

Take a vague, big question, like whether the S&P 500 will be up or down ten days from now. You don’t just ask it. You bring in your financial knowledge, break it into concrete sub-questions, work through them with the AI one by one, and self-validate along the way before arriving at a final answer. I spend several lectures on this. The idea is intuitive. Making it work, automatically and autonomously, involves real technical subtleties.

P&Q: Almost all of my work has been in strategy, which is very subjective. It’s nearly the opposite of finance. Accounting is the least subjective; get too creative there and you’re in terrible trouble. Finance is math-based. Strategy is at the other end, where ultimately there’s no single right answer. How do we stop these machines from telling people what they want to hear instead of what the data says?

Dou: The key is that the technology to do this sits outside the large language model, and we have to fully recognize how the model works. I spend a lot of effort in the classroom on this.

At the base is a probabilistic, transformer-based model. It’s a rigid machine that learns the patterns of language, English, German, whatever. It has no goals of its own. Then there’s an important layer on top: reinforcement learning from human feedback. That layer takes the language capacity and uses it to interact with humans in a way that makes humans feel better, to meet human preferences.

Once you understand that, you realize that to get at the foundational knowledge inside the model, you have to design something to overcome that “make the human happy” layer.

Here’s the tricky part. You cannot simply tell the model, “I want the ground truth. Don’t try to make me happy.” That’s not how it works. Instead, as a financial expert, you design a structure to guide the model toward the true financial reasoning. That structure is the subjective part. It’s our human expertise.

Say I’m a financial specialist feeding the model news, analyst reports, earnings calls. I want the true financial logic behind them, not comfort. So I design a tree of sub-questions. Start at the top with a very simple question, where there’s not much room for the model to make me happy or unhappy. Based on that answer, follow the logic chain down the tree, step by step. Eventually you extract very useful information and the reasoning chain along with it.

That’s one way to combine specialized human knowledge, financial, managerial, supply chain, whatever it is, with large language models. The way I think about it: the model is like a super-smart, very knowledgeable person, but the only way we can communicate with them is in English, and they’ve been trained to speak English in a way that makes us feel happy. Once we recognize that, we build an external structure to guide this very smart but biased communicator toward giving us genuinely useful information.

P&Q: Two final questions. How are students receiving this? Elsewhere we hear students are horribly worried about jobs, but quant traders are not so different from AI, and they’ve done fine. Are the concerns of students coming into your lab different from students in other fields? And then, anything else on your mind?

Dou: I get mixed feelings from students. On one hand, they feel the pressure. The skill sets institutions want are changing, that’s certain, but by how much and in which direction? There’s huge uncertainty in front of them.

On the other hand, a lot of students are very motivated. They see this as a real opportunity. European students in particular tell me they feel lucky, because top business schools in the U.S. like Wharton are making a quick, real effort to incorporate AI into education and research.

We announced the lab only a few days ago, and I’ve already received around 50 requests from people who want to engage with it, from Wharton, from across the U.S., from Asia, from Europe. And not just college students. I’ve received perhaps even more inquiries from high school students, not only in the U.S. but also in Hong Kong and China. It’s remarkable. Young people already feel the opportunity and the pressure, or perhaps their parents have passed it on to them. Either way, they want to be engaged. They don’t want to be left behind.

P&Q: This has been a fantastic interview. Anything else to add?

Dou: The last point I want to make is that this isn’t my effort alone. It’s a collective effort, with Emily Hemming, Traci Doyle, and a whole team. When I glanced through the CVs of the high school and college students who’ve written to us, from Wharton, from Penn, and beyond, I saw real talent, the future. I want some of them engaged with the lab, and our team immediately built the participation form and a screening system that’s remarkably efficient. Without a team like that, none of this would be possible.

---

Michael Olenick, JD, is a former INSEAD research fellow and founder of VSTRAT.ai. He has spent four decades building AI and expert systems; his research has been used at INSEAD and, before then, cited by Congress and media on how to resolve the 2008 era financial crisis. Michael currently lives with his partner, Anastassia, in Austin, Texas.

This is the latest in an ongoing series exploring how business schools are integrating AI into teaching and research, one school and one faculty member at a time. See also:

Columbia’s AI Playbook: Teach The Thinking, Not The Tool

Better Prompts Or Better Judgment? Columbia Bets On The Latter

At Duke Fuqua, Using AI To Strengthen The Bonds Between Humans 

At Indiana Kelley, “AI Is,” Not “AI Will”

Training Students To Leverage Increased Cognition (Michigan Ross)

Rebuilding The Role Play (Kellogg)

Learning Requires Friction (Wharton) 

The Deck Is Dead (Stanford)

We Teach MBAs To Apologize About The Best Tool They Have (INSEAD)

© Copyright 2026 Poets & Quants. All rights reserved. This article may not be republished, rewritten or otherwise distributed without written permission. To reprint or license this article or any content from Poets & Quants, please submit your request HERE.