How AI Is Reshaping Quantitative Finance

Itay Goldstein: Welcome, everyone. This is Season 3 of the “Future of Finance” series here at the Wharton School. I am Itay Goldstein, a finance professor and currently the chair of the Finance Department. We devote the third season of the “Future of Finance” series to this very exciting new topic of AI in finance. Today we’re going to dive deep into the topic of how AI is shaping quantitative finance.

Quantitative finance has been one of the fast-growing areas of finance, both in research and in teaching. We just launched the new Bruce Jacobs Master’s in Quantitative Finance program in order to help our students get more familiar with some of these new techniques in quantitative finance and in particular AI in finance. We’re going to try to get an understanding from both the academic perspective and the practitioner perspective on how AI is used in quantitative finance.

For this, we have two perfect guests. First let me introduce my colleague here at the finance department, Nikolai Roussanov. Nick Roussanov is the Moise Safra Professor of Finance, and he has been with us for about 20 years now. He is also the MBA major quantitative finance advisor. Nick, it’s great to have you.

Nikolai Roussanov: Always happy to be here. Thank you for bringing me on.

Goldstein: Together with Nick we are very happy to have Ingrid Tierins, who is the head of the data strategy team in the global investment research division at Goldman Sachs. She is also a part of the Wharton community, as a member of the advisory board for Wharton’s Jacobs Levy Equity Management Center for Quantitative Financial Research. This is where a lot of the activity on quantitative finance here at Wharton is taking place.

Ingrid Tierins: Thank you, and a great pleasure to be here.

Goldstein: Let’s dive right in. I’m going to start with you, Nick, for a little bit of an academic perspective. I just said that you’ve been with us for about 20 years now here at Wharton. You have been heavily involved in quantitative finance research. You have seen a lot of the changes in this area. Can you [provide] an overview of how quantitative finance has changed in the last 20 years and how AI is shaping up to change it further?

Roussanov: Let me start by clarifying what we mean by quantitative finance, and then maybe I’ll try to clarify what we mean by AI, and then try to put these things together. When we say quantitative finance, historically this has meant different things. Originally, following breakthroughs in finance theory in particular, [such as with] option pricing and so on in the 1970s, quantitative finance typically referred to quantitative or mathematical modeling of complicated financial assets. In particular, [it referred to] derivatives or bond pricing, understanding the structure of interest rates, the yield curve, and various derivatives associated with interest rates and so on. Eventually that acquired a name – “financial engineering.” That was understood in many different ways. But for many years the view was that quantitative finance is basically [about] using mathematical models to value complicated securities – securities that have complicated payoffs and need some structures of mathematical models to understand them.

Over time, with the advent of statistical methods and computing power, data on financial assets became used more and more widely. Quantitative finance began to be associated with investing using predictive models or statistical models, trying to predict asset returns using historical variables. A lot of debate [was] around market efficiency, following [the work of economist] Eugene Fama. The idea of market efficiency, and using various statistical tests to analyze predictability of returns took over as the larger [debate, including the] use of the term “quant” in the financial world, in particular on the buy side.

In the investing world, the firms that use quantitative, statistical methods, would not necessarily be pricing complicated securities, but instead would be forming trading strategies and building portfolios to optimize the risk-return trade-off using historical data. I think to this day, when we say quantitative finance, it could be applied to largely, perhaps, this quantitative investing world. Oftentimes, people think of it as systematic trading, where the program that is built on data is effectively producing trading signals, analyzing those trading signals and producing portfolio allocations without manual human interference – “systematic” refers to that.

Now, with the advent of larger data sets, [including] alternative data ranging from credit card transactions to satellite imaging and so on, as well as more powerful statistical tools like machine learning, quantitative investing has grown beyond what would traditionally be described as systematic. Even some of the discretionary portfolio managers will use quantitative tools and data. People call this “quantum mental investing.”

It used to be [that] the fundamental or discretionary managers would be [more] important [than] the quantitative or systematic ones. Now, there is a bit of a merge to some extent between those two, because even the fundamental managers would use data and statistical methods, and machine learning tools and so on, to construct their signals and inform their trading strategy. Of course, the risk management that underpins everything that the financial industry does is also heavily quantitative.

Now, how does AI come into this world? Again, it depends a little bit on what we mean by AI. We used to talk of AI and machine learning in one breath, meaning basically large models with large numbers of parameters, larger than what traditional statistics would consider sensible. There have been some breakthroughs in that area over the last 20 to-30 years. As models have grown, we’ve gone from calling this machine learning to calling this AI.

[That is]. because large language models, which is what we now associate with AI – let’s say GPT, Claude, and so on and so forth – are fundamentally based on this transforming network architecture, which is an outgrowth of machine learning. Basically, [these are] very large, multi-parameter, millions and billions of parameter models that make sense of not just individual observations, but link them together. That’s how they’re able to process text, by looking at words together in a sentence and finding and analyzing connections between them as opposed to just looking at individual words.

How is this being applied to quantitative finance? Well, there are several ways in which AI is revolutionizing quantitative finance. The most basic one is the same in which AI is being used elsewhere, which is just a productivity tool for model builders, for researchers on the quantitative finance side, as well as portfolio managers that makes their workflow a lot more efficient in terms of writing code. We know that AI has revolutionized or maybe upended the software industry, and of course quantitative finance is built on a lot of computer code. To the extent that AI is helping increase productivity in the production of that computer code, it is obviously affecting that field.

In systematic quantitative investing, AI, again, is represented by these very large multi-parameter models that are now used directly to predict returns even better than what was done using earlier methods. Finally, people are using large language models themselves – without necessarily tinkering with the underlying architecture – using large language models themselves to analyze vast quantities of textual data. [They can do this] much better and much faster than human analysts, to produce signals for fundamental portfolio managers, including the discretionary ones.

Goldstein: Okay, very good. Ingrid, from where you see it in the industry, and obviously Goldman Sachs is a very important player in the industry, I’m trying to get a sense of how much of a game changer AI is. Is it just helping analysts do their job faster, or is it doing completely different things?

Tierens: The answer is, of course, you know, everything. I’m going to expand a little bit on what was said before here. We talk about quantitative finance. It almost sounds like there’s quantitative finance and there is non-quantitative finance. If we look at the world right now, all finance by definition is really quantitative. [In] every single investment strategy or process, if you’re not using data, analytics, technology, and models, I don’t really think that’s 2026 anymore.

It’s similar to what we were saying before here. A distinction that is more meaningful to me is, there are different investment approaches. There is systematic and there is primarily discretionary. If you think about systematic, to me, that is a mix of very well-diversified portfolios. The bets are typically more limited. [On the] discretionary side, obviously people have more conviction. Human judgment still plays a larger role. And portfolios tend to be more concentrated.

But what is consistent across the board is, AI is reshaping every single investment approach, whether that’s systematic or whether that’s discretionary. And if I look at what’s going on – and we talk to a lot of clients as well – there are three categories that I would look at. Number one, AI is turning a lot of previously inaccessible information into data. That’s super important to think about. Again, if you roll the clock back a little bit, typically investors were looking at structured data. A very simplistic example: prices, earnings, economic releases. We’re in the midst, right now, of the earnings season. Look at an earnings call. Many years ago, a quant signal could simply be an earnings surprise. There are people who build businesses around earnings surprises as their main quantitative signal.

Roll forward. With the advent of AI, you can be analyzing every single sentence spoken by management companies. It could be across thousands of companies. You can do it going back many, many years. The same is the case for filings, research reports, images, videos – you name it.

From that perspective, AI is really turning an increasing share of human activity into data. It’s interesting to think about it, right? If you think about digitization, we started with transactions, then communications. Now we’re actually – thanks to AI – starting to digitize reasoning itself, right? Every single time, when any of us is interacting with a chatbot, you have AI now capturing how you think about things. It’s probably the first time in human history that the thought process itself is being digitized. Just think about the possibilities from that perspective.

The second thing that’s important is, AI allows people to connect information across silos, right? That’s coming to the scalability thing and the productivity thing you were talking about. In the past, again, time is limited. Human capacity is limited. You would look at predefined variables. You would potentially construct strategies around that. AI can do a much better job at starting to sift through diverse sources of information and see if they can identify additional insight from that. Again, a concrete example. As an analyst, or a PM (portfolio manager), I can, at best, read 20 reports on the stocks that I really care about. Now, [with AI, I can process] 20,000 related documents, et cetera.

And then a third one, which is very important as well. Going back to the quantitative aspect, AI is also democratizing quantitative analysis itself, right? In the past, the people who could do quantitative analysis taking advantage of sophisticated risk management tools, portfolio analysis tools, et cetera, tended to be the people who could code; [they] were really the power users of these tools. Now, you put the interaction with natural language in the mix. People one or two steps removed from that can potentially get way more use out of a lot of that tooling. The applications that may have been built many, many years ago, and the applicability is just going to become much broader.

So, addressing [your] question, does that mean that it’s a productivity play, or is it beyond that? There is definitely productivity. Just think about it. Searching, reading, summarizing, translating, all of that. It’s the complete table stakes right now.

But what is way more exciting is the fact that the scope of what you can analyze and the opportunity set being much broader; it opens a lot of doors, right? Again, going back to the earnings season. In the past, [it was]: “Okay, what did management say about whatever.” You know, some example of spending plans. Now you can go back like, “Hmm. Did the market detect that?” We run analyses over multiple companies going back multiple decades.

The other thing that I’ve realized from talking to analysts as well is, as a researcher, you always have tons of ideas of things that you might want to analyze. We’re all self-selecting if we have limited time and limited tools available. And you’re going to go, “Meh, low probability event. I’m not going to do all of the work.” AI, again, opens the door that you might start to analyze lower probability scenarios that potentially can yield insight that otherwise you wouldn’t have been out capturing.

And so, sitting in a sell side research division where the mandate is – our goal is really to come up with differentiated insights in timely manner. The productivity gets to the timeliness. The scope and the increased opportunity set gets to: can we surface things that are more differentiated and make a difference from that perspective?

Goldstein: You mentioned democratization of financial analysis. It brings up a very interesting question. What is AI going to do to competition in financial markets? I guess one scenario is these tools are now going to be available to everyone. So it really opens up the field. There will be many new players, and it’s not clear who is going to have the comparative advantage. But another scenario is the barriers to entry are going to increase because it’s really difficult to have the capacity to use all these tools. And so it’s going to make it less competitive. Where do you think we’re going to go on that?

Tierens: Nick, do you want to take it first?

Roussanov: Sure, I’ll start, and you can see if you disagree. I think there’s definitely going to be an arms race of sorts emerging as a result of the advent of this new technology. Now, of course, finance or investing– even though it’s not zero sum – has a zero sum aspect to it in the sense that for somebody to outperform, they have to basically take somebody else’s lunch, right? Somebody else has to underperform. And there’s always going to be competition for that alpha, for that outperformance. That competition mostly materialized through really lucrative pay packages to star portfolio managers and analysts, and the hiring of top-notch data scientists, and building bigger and better systems and models.

Now with AI, we will see financial firms, I think, entering into the second or maybe another level of horse racing, just like the AI builders themselves are competing on who has the better model. And we have these vast investments into data center capacity to train bigger and bigger models.

We’ll see this happening among financial firms as well. They will invest more and more in the tokens of those AI models that people use, but also build their own models and their own data centers, which they are doing already. Let’s say XTX is a good example, the British high-frequency trading firm that’s building its own data centers in Finland, which is, of course, a cold place, and good for keeping lots of hot servers.

So there will be a degree to which the arms race is going to make things more and more competitive. It is true that in order to be able to invest in data centers, you need a certain scale. So in that sense, maybe this will lead to further consolidation in this industry, and it will be harder for smaller players to make a difference. But from the standpoint of let’s say retail investors or ultimate users of capital firms, it could still help make markets more efficient, even though that competition will be larger and larger players going forward.

Tierens: I’ll add a little bit to that. Obviously, as we were discussing, data is becoming more accessible. The tooling is becoming democratized. But at the same time, I think the noise-to-information ratio is increasing dramatically as well. And so the edge is still going to be coming from, “Do you really understand what’s going on? When does this fail? Where is the domain expertise connecting with what the tools are producing?”

And then really, how is this related to an investment thesis? Those are all still very much open questions. You can think about it in a simple example. Very topical, too. Suppose some AI tool finds that companies that are discussing AI are outperforming. That might be interesting. But is that causal? Is that a temporary phenomenon? Is that already priced into the market? Does it persist? Et cetera. And so, humans still have an important role to play.

The other thing I would put here as well is – it’s a bit analogous to what has been going on with the asset management industry in and of itself. Think about what happened with the introduction of index funds. Before index funds, any manager who could construct a broadly diversified portfolio could call themselves someone who’s running an active strategy. Index funds completely changed that ballgame. If you roll forward, since the advent of index funds, a whole lot of stuff that was called alpha has become completely beta products. And the benchmark moved higher and higher.

So I draw the analogy here with AI doing the same thing with information. Gathering information now is becoming incredibly cheap. If your value-add was, ‘ was the person who was sitting in the midst of information and could gather it,’ you’re probably going to have to rethink a little bit what you’re doing here.

It also starts posing the question of, what is your own personal alpha? What is the alpha of your business? I’s just a trademark of innovation that as things get entered, obviously, we all need to move forward. Hopefully we keep capturing or creating our own differentiated alpha and we keep staying ahead of the benchmark.

Goldstein: This touches on a very important question that we have with AI, which is, what is AI and what is human? If AI can do all these wonderful things, then what is left for humans to do? Connecting to some of the things you both said, I guess one possibility is that we still need humans to it all together, and be a check on AI and make sure that it all makes sense and that there is an underlying intuition and underlying thesis. Or maybe another possibility is that there are still some pieces of information that AI cannot pick up on its own. And humans are those who are bringing those signals from outside the system. Where do you see this tension between humans and AI?

Tierens: I look at them as being very complementary. Not necessarily competing, and hopefully reinforcing each other. Three things I would bring up – and you already hinted a little bit to some of them. [The first is] the context and framing the problem: you can ask AI tools a lot of questions, and it’s always going to give you an answer. But the human is still deciding [as to] what are the questions that are really worthwhile asking. The imagination of the human of what is worthwhile pursuing, that creativity, I think that’s going to remain.

The second thing is judgment, especially judgment when there is uncertainty. That can go back to earlier models, like the whole quantum history. It’s the same thing. You stick numbers in an optimizer, and even if the differences are tiny, the optimizer is going to take it very literal. So you’ve got to build in uncertainty around that, et cetera.

AI is learning very, very quickly about a whole lot of stuff. But it’s still, to a large extent, based on situations that have occurred in the past with probabilities around it. And so, you roll back the clock. How would AI have dealt with episodes like COVID, geopolitical shocks, reducing [those] to a narrow quant concept like momentum. Momentum works really well until there is a break in the system. Again, same thing here – how will AI deal with that? I think it’s still a little bit of an open question.

Last but not least – responsibility and accountability. At the end of the day, AI can create a ton of possibilities, [but] humans are still going to decide which of these possibilities deserve capital to be committed to them. And then, if the outcome isn’t really what you wanted, I don’t think an AI tool is going to volunteer to take the responsibility for that. That’s an open question in a whole host of other fields as well, forget about finance, where it’s an AI-driven outcome that someone may be pursuing. It goes wrong. Who, ultimately, will take the responsibility for it?

Goldstein: Nick, do you want to add to that?

Roussanov: I agree with everything that Ingrid said. The key point is that ultimately asking the right questions is still, fundamentally, the advantage of the humans. And ultimately, judgment. We’ve now seen that I can have AI write what looks like an academic paper in finance, let alone construct a trading strategy. But the question is, is that a good question to ask? Is this a meaningful strategy to use? And is the paper good? I don’t think we’re there yet. But there are people who are trying.

Ultimately, human judgment will be key. [As for the]. democratization that Ingrid mentioned earlier, it just says that the barriers have been lowered to being able to ask interesting questions that require a lot of work to get to the answers. Some questions that are maybe very interesting, but seem unlikely to be solvable, are going to be asked more and more than they were in the past precisely because ultimately our time as humans is limited. And you’re going to prioritize asking questions to which you’re more likely to be able to find an answer.

This democratization of generally research, but quant finance research in particular, will be happening. It certainly speaks to the complementarity between human intelligence and artificial intelligence.

Now that’s not to say that AI will not, at the same time, act as a substitute for certain types of human intelligence. We already have seen that coding is going to be a lot more commoditized. [But it is] not that [coding skills are] going to be irrelevant. You still need to be able to make sure that whatever the code that AI wrote for you actually makes sense and is doing what you want. And this is going to be a big challenge for a lot of companies going forward. They’re relying on AI to produce their code base.

But we don’t necessarily need as many entry-level workers in software, and maybe not as many entry-level workers, ultimately, even though maybe we’ve not seen that yet, in let’s say investment banking where we need analysts basically sifting through company reports or putting together PowerPoint presentations, things that AI is pretty good at.

On the one hand, that means that this effort will be saved and economized in a way that would allow these smart people to do something else, and it may be more interesting. But it’s also possible that it will reduce the headcounts, and is already reducing headcounts at various entry-level positions. The downside of that, of course, is that to be able to progress to become a portfolio manager, you have to be mentored by somebody. But if there’s nobody to be mentored, who is going to step into those shoes eventually, if the bottom rungs of the ladder are going to be hollowed out by AI? That is going to mean that potentially that ladder itself is going to have to change.

I don’t think we have quite yet figured out how to rebuild the career ladder in quantitative finance as well as in other fields in a way that takes advantage of the democratization of, let’s say, research tools that in the past were only available to maybe select [people], but also preserves these mentoring stages of one’s career. I don’t know if Ingrid has any views on that, by the way.

Goldstein: Yes, I was going to maybe follow up on that and ask. Now with the launch of our new Master’s in Quantitative Finance [course], what are the main skills we should try to get for our new graduates? What is the main thing that they need to have in quantitative finance going into the job market?

Tierens: I’ll answer it a little bit from the practitioner’s perspective. I’m going to keep this fairly simple. One, learn data. There is no AI without data. I think with AI, sometimes people seem to think that that’s going to solve all of the data problems magically. Get some experience getting your hands dirty with data, because then you’ll have a much better understanding of what potential shortcuts AI may or may not be making.

It’s a little bit of a self-serving answer as well. [That is]. because my background and my Ph.D. degree and my career here, if I need to reduce that to one sentence, it’s, “I’ve spent my entire life trying to make sense of data.”

The second one is, obviously, learn AI. That doesn’t mean that people need to know all of the ins and the outs of the large language models. But you do need to understand where they do well, where they fail, how you actually validate the outputs. [That is] because, again, in situations where you ask a question, you get an answer, and the output looks so convincing that people sometimes forget to do their due diligence on this.

And then last but not least – which goes back to, there’s an important role for humans to be played here – you need to learn judgment and communication. The people who can connect the dots and connect the technical pieces with the investment decisions, I think those are really going to be the people who are going to have the legs up. So for any student who’s listening here, I would say learning of finance to ask or start asking the right questions, and then learning of technology to be able to use the tools, and then learning of communication so you can actually persuade people.

Goldstein: Okay, we’re coming close to the end of our time here. So I just want to have each one of you maybe take 30 seconds telling us, where do you think the field is going? One thing about AI is that it is moving very fast. And we are seeing the changes happening in real time very fast. So if we’re trying to sit here and project what’s going to happen in 10 years, how the field is going to look like in 10 years, what do you see? Nick, maybe you can get started.

Roussanov: Who knows what the world will look like in 10 years? But I do think that the field of quantitative finance will be alive and well, and perhaps will be more all-encompassing as a subset of finance. It has been historically somewhat siloed, but as Ingrid said, ultimately, all of finance is quantitative. All of finance is using data. Thinking about the risk-return tradeoff is fundamentally a quantitative task. So I think it is going to be permeating [across] more of the financial industry, and its skills and tools of quantitative finance will be a lot more universally applied, in part because of AI-driven democratization, which Ingrid talked about.

AI will certainly be ubiquitous as a tool in getting answers. Humans will still be central, of course, in asking the questions. As well as, again, as Ingrid said, validating those answers. We know already from our experience with chatbots that AI will always produce an answer. And it will learn what answer pleases you, the person who is asking the question. It certainly aims to please more than anything else. And of course, that’s not the way to produce something that is scientifically accurate.

As far as quantitative finance [is concerned], the traditional tools are still going to be relevant. Ultimately, back testing, let’s say, the output of large language models is impossible because large language models have been trained on everything out there and have learned everything. So there’s no way that we can validate out of sample that their predictions actually work.

There are a lot of challenges remaining with the implementation of AI. But it will be ubiquitous. The field will adapt and will use it, with some limitations. Again, there are privacy concerns. People are worried about their code base, about their proprietary models being learned by the AI tools that they use. So a lot of more of the AI models will be trained internally. A lot of money will be spent on that. It will be central. But I don’t think the field of quantitative finance itself will be gone. It will probably take over the rest of finance, precisely for the reason we discussed.

Goldstein: Ingrid?

Tierens: Ten years is way too long because the field is moving way too fast. But I completely agree, it’s going to be so ubiquitous that I think we’re not even going to talk about quantitative anymore. We’re just going to say it’s investing. So that’s prediction number one. We don’t need the word quantitative anymore.

Number two, I do think the human is going to remain front and center to this. The benchmark is going to keep moving. But I do think the field is going to become, by definition, also more conversational. I mean, less a way of, “Do I have a particular skill set?” But you know, “Can I just use all of the tooling that’s out there?” Again, the issue of information is all out there. But the distinction between information and true insight will become even more pronounced than it already may be the case.