Pilot interviews
Alejandro Lopez-Lira: 'Large language models are your employees'
Alejandro Lopez-Lira, a finance professor at the University of Florida, explains how large language models work like employees inside a research system.
Alejandro Lopez-Lira has a PhD in finance from the Wharton School and is a professor at the University of Florida. Alejandro grew up in Mexico and did undergraduate and master's work at ITAM before moving to the United States for the PhD.
Open it in the Autopilot app. https://link.joinautopilot.com/eXKr?c=alejandro-lopez-lira-interview&af_sub1=intro
As a professor at the University of Florida, Alejandro wrote a 2023 paper on whether large language models could help forecast stock price movements from news headlines. Alejandro now publishes nine Portfolios on Autopilot, each built around a different large language model or theme, and designs the system that puts those models to work and checks their research.
We talked about treating large language models as employees who need a firm structure, how the research process actually works, and why Alejandro publishes on Autopilot instead of running an independent hedge fund.
This interview comes from the transcript of our conversation with Alejandro Lopez-Lira in May 2026. It has been edited and condensed.
Who are you, and how did you get here?
My name is Alejandro Lopez-Lira, and I hold a PhD from Wharton. I did my PhD. I started in 2015. Back then it looked very different, because we didn't have Claude or anything like that, so it was mostly coding advanced algorithms. Then I started researching, basically, if you invest in one asset versus another, what drives the differences in how much you're going to get.
That's mostly my background. Ever since then I've been researching AI and machine learning and finance, the intersection, and how it affects investments and what these tools can and cannot do for investments.
Where are you from, and how did that lead to Wharton?
I'm from Mexico. I did my undergrad and my master’s in ITAM. It's a very small, very hardcore institution, and the exams were at some point harder than the ones at Wharton.
A lot of what the institution does is it sends people abroad for PhDs, so I was part of that cohort. I was one of the few that went into finance.
What was the research paper about, in simple terms?
The simple question is can we use this for anything that would help make better investment decisions. The simplified scenario is I take some news headlines, and it's like if I ask you as a person, how would that impact the stock price of, let's say, Apple? Is it relevant, is it not relevant? There's thousands of headlines every day, and there's obviously limited capacity as a person to read through all of them.
The question we wanted to answer in the paper is if we give the model a headline, is this headline going to forecast good prices going forward, or is it going to forecast a decline. The methodology is relatively simple. There is a way to call it via software, like what's called an API. And for each of the headlines, we had something like a hundred thousand headlines in the research project, we would launch a new instance, ask the question, is this headline good or bad for the stock, record the answer, and do that for all the news headlines every day.
What's it like being a professor while running these strategies?
It's very nice, it's honestly very fun. I do mostly research, so a big chunk of my work is research. I only teach normally during the spring semester, so now I'm done for the rest of the year. And I teach finance and machine learning, so I get to teach my own tools, my own research.
Everything is very connected. For the undergrads and the master's students, I teach them roughly how to form this type of strategy and how they can build it. So in theory any of my students would be qualified to run one of these portfolios. In practice there's a lot of details, you have to invest a lot of hours ironing out issues, but in theory it's very straightforward.
What's the difference between what you do and someone who just goes to ChatGPT and asks it to pick stocks?
My personal edge is that this is my line of work. Everything I do is using large language models all day, every day, studying their capabilities and limitations. There's a lot of things that large language models would do that people would think, oh, this model is so stupid, and the answer is it's not really. The models are extremely smart, but you need to know how to ask them. I think my edge is to know the capabilities and limitations, and once we know the capabilities and limitations, it's easier to write a system around it that will organize the large language models in the best way. The way I would say it is large language models are your employees, and you need to have the firm structure.
Can anybody do this? Yes, but you would need to identify the correct firm structure every single time, and it's not practical. The other practical reason is that it's just expensive. You cannot just ask, can you analyze a thousand stocks for me? It will not do it, because it's lazy, and because its working memory, its context length, how much it can read, is going to be way smaller than the analysis of a thousand companies. So it just literally cannot do it. So at this point you're coding a system. Can anybody code a system? Yes, but you're just inventing a hedge fund from scratch.
How is your system set up to keep AI from pretending to do the research?
It's a combination of good prompts and also verifying over and over again. Models are very weird, they're extremely smart, but you have to say things like, you should definitely take this work very seriously, otherwise the model will not take it seriously, or it will make up data. You would imagine it wouldn't do that, but now you have to tell it explicitly. So a lot of it is prompt optimization over and over again. But the most important part is you have many, many sequential checks that everything is correct, the same way that if you hire an analyst and say, can you analyze this company, and then you'd go to another person and check, are the numbers correct.
Within the system there's many, many verification layers until basically there's very few errors remaining. I would say it's just a matter of designing your system around the fact that the models will make mistakes, so you just need to catch them.
How does the research process work, from the macro picture down to individual stocks?
It's a very complex system, it involves a lot of levels of research. The models finish their training up until a given date and they don't really know what happened afterwards, so a big chunk of it is telling the models what has been happening and making them use this information in the best way. The process would start at the macroeconomic level, so what's been happening recently, what's the latest, let's say, Fed policy. Once we have that, we ask the models what's the most likely scenario that's going to happen with the fiscal policy, with monetary policy, with tariffs, with geopolitics, with anything that can impact the stock markets broadly.
Once the system has that information, it gets passed to sub agents. The task of each team would be, we need to research a particular company, let's say Microsoft. One sub agent would be in charge of gathering the data, another would be in charge of analyzing whether the data is correct, and use this data to provide a forecast on the earnings, on the stock price, on other potential events that can impact the firm. You can think of this as running a hedge fund, but instead of using a thousand human analysts, I'm doing this with large language models.
As AI makes it easier for more people to launch their own AI-driven strategies, what sets yours apart?
It has to do with all the research that's behind it. It is true that anybody can launch an AI strategy, and I would probably say I'm one of the most AI optimistic persons on the planet, but I would not sign up blindly for an AI strategy, because there's many ways things can go bad. If you just try to cram in, okay, do this research for a thousand stocks, the model is not going to do it, it's going to pretend to do it, or a combination of that. So I would say most of my advantage in the AI space comes from my research. Everything I do is academic research quality, and everything gets tested. Everything comes from first principles.
Before Autopilot, did you ever think about launching your own hedge fund?
Honestly, back in the days before I finished my PhD, I was thinking about launching a crypto based fund. It actually came up twice. Before I entered my PhD, this was 2015, I was really considering, with a friend, just dropping everything and launching a crypto fund. But then I got admitted into Wharton, and then I didn't do it. I played with the idea a little bit later, but didn't really execute it. I'm not going to lie, a big chunk of it was that it's just very, very expensive to launch your own hedge fund. You need more than a hundred thousand dollars just in fees, lawyers and stuff like that, and I just didn't have that amount of money, or I didn't feel like it was a good investment.
Here are the stock picks, and then somehow they get translated into people's positions. Running a hedge fund is a different animal, it's something very interesting for sure, but it's on a different scale.
“"Large language models are your employees, and you need to have the firm structure."”
Follow Alejandro Lopez-Lira's Portfolios on Autopilot
Alejandro Lopez-Lira publishes AI World War III Portfolio, GPT Portfolio, Claude Portfolio, DeepSeek Portfolio, The Grok Portfolio, AI Recession Portfolio, Muse Portfolio, Gemini and AI ❤️ BTC: LLM + Crypto on Autopilot, an investing app that lets you follow Portfolios from investors, AI models, and public filings in your own brokerage account. Visit Alejandro Lopez-Lira's Pilot page. https://link.joinautopilot.com/eXKr?c=alejandro-lopez-lira-interview&af_sub1=body
Following a Portfolio on Autopilot works like this:
- You follow in the brokerage account you already have, and the money stays there.
- On Premium Tier, when the Portfolio changes, Autopilot Advisers sends the orders to your broker; on Basic Tier you get an alert.
- Review the Portfolio's method, holdings and any available performance record before following it. New Portfolios show return figures after 30 days.
- The Pilot sets the subscription price, there's a 0.00% AUM fee, and you allocate at least $500.
You're never trading at the same moment as Alejandro, and your account won't match the Portfolio exactly.
Open it in the Autopilot app. https://link.joinautopilot.com/eXKr?c=alejandro-lopez-lira-interview&af_sub1=end
"Autopilot" refers to Autopilot Holdings Corp and Autopilot Advisers, LLC, an SEC-registered investment adviser. AI World War III Portfolio, GPT Portfolio, Claude Portfolio, DeepSeek Portfolio, The Grok Portfolio, AI Recession Portfolio, Muse Portfolio, Gemini and AI ❤️ BTC: LLM + Crypto are managed by Alejandro Lopez-Lira, an independent Third-Party Pilot unaffiliated with Autopilot. Autopilot doesn't manage, validate, or guarantee this Pilot's portfolio. This interview reflects Alejandro's own views as of May 2026, edited and condensed, and isn't a recommendation to buy or sell any security or to follow any Portfolio. Individual stocks shown are illustrative and not a recommendation. Strategies linked to digital assets carry extreme volatility and risk of substantial loss; holding equities or funds exposed to digital assets is not the same as holding those assets. Where a Pilot uses an AI tool in their own stock-selection process, that tool does not advise clients, select investments for clients, or place trades. Investing carries risk-including possible loss of principal-and past performance is not indicative of future results. Figures current as of 10/09/26. View our Customer Relationship Summary and other important info at joinautopilot.com/disclaimer.
- Month
- +1.7%
- 3 months
- -14.9%
- YTD
- +7.9%
- 1 year
- +20.7%
- Since start
- +207.4%
As of
- Month
- +4.8%
- 3 months
- +0.93%
- YTD
- +32.2%
- 1 year
- +33.5%
- Since start
- +114.7%
As of
- Month
- -1.5%
- 3 months
- -6.1%
- YTD
- n/a
- 1 year
- n/a
- Since start
- +20.1%
As of
- Month
- -8.6%
- 3 months
- -18.3%
- YTD
- +23.1%
- 1 year
- +27.2%
- Since start
- +58.9%
As of
- Month
- -4.7%
- 3 months
- -4.8%
- YTD
- +24.1%
- 1 year
- +26.9%
- Since start
- +57.7%
As of
- Month
- -0.04%
- 3 months
- -0.52%
- YTD
- +8.9%
- 1 year
- +11.2%
- Since start
- +9.5%
As of
- Since start
- -7.4%
As of
- Month
- -19.8%
- 3 months
- -32.3%
- YTD
- -30.8%
- 1 year
- -35.6%
- Since start
- -52.6%
As of
The performance data presented reflects the aggregated results of actual client portfolios managed by the Registered Investment Adviser (RIA), based on the trading activity in the securities contained in the relevant portfolio over the specified period. If fewer than 10 clients are invested in the portfolio, the data represents the aggregate performance of all clients following this strategy. Performance figures are shown net of advisory fees, calculated using the applicable subscription fee and an assumed $50,000 account size. Because subscription fees are flat rather than asset-based, accounts smaller than $50,000 would experience a higher effective fee percentage and lower net performance, while larger accounts would experience a lower effective fee percentage. Actual investor outcomes can vary due to factors like brokerage support for fractional shares, account size, trade timing, fees, and taxes, and past performance should not be interpreted as a guarantee of future results or typical of all portfolios.