Trade like a bot

0:05Pushkin. Everyone knows the robots are coming, right? Artificial intelligence has the potential to reshape everything, all the kind of professional services, law, accountancy, all those kinds of things. Is journalism safe? I guess we'll find out. But a new study suggests that AI is better at picking stocks than you. So today on the show, we're asking, is nothing sacred? Should fund managers welcome their new robot overlords? This is Unhedged, the markets and finance podcast from the Financial Times and Pushkin. I am not a robot. Honest. I'm Katie Martin, a markets columnist here at the FT in London.

0:45And I'm joined by fellow non-robot human Robert Armstrong, who writes the Unheged Newsletter in New York. Hello, fellow human. It's a great day not to be a robot. Every day is a good day not to be a robot. But how do you know you're not a robot? Is it because you tick those boxes on the internet that say select the squares with cars? That's the way we know we are human. When you start failing those tests, you have to wonder if you're just a simulation. Yeah. So look, we're going to talk about AI, but I want to throw you a curveball first and say you are on record as saying that politics do not matter for markets.

1:21To which I say, vive la France. I mean, what the heck is going on over there? I mean, this is really a fairly sizable move in French stocks, French bonds. They do not like this snap election that's been called by President Macron. Fairly sizable. What do you mean by that? Well, look, it's only a few short weeks. What is that sound that French people make when they're not impressed? I mean, there's a lot of them. Count me not impressed by the moves in this French stock market. So you're giving a Gaelic shrug, but I guess we'll have to see the results, I think, early July. So we're going to see whether your insouciance is warranted.

2:06But so, OK, you've been writing about this today. It has human-like capabilities at picking stocks. So do they suck as badly as humans or do they do something actually good? So a couple of business school professors at the University of Chicago have taken a bog standard large language model and fed it a bunch of balance sheet and income statements. asked it a bunch of questions about the, and I should say that they were anonymized. So the machine didn't know what year it was or what company they were, it was looking at and so forth. And then they said, what's going on with this balance sheet? What are the big changes?

2:47They sort of prompted it into thinking about these financial statements in a kind of financial analyst-y way. You know, calculate the margins. What is the liquidity ratio? What is net cash or whatever? And then they would ask a very simple question of chat GPT. They would say, are earnings going to go up or down next year, whatever the next year is after the balance sheet they're looking at? By a lot or a little? And how sure are you? And the machine hummed and clicked and made whatever noises computers make and came back with a set of predictions that turned out to be not hugely better, but clearly better than the predictions analysts had made about those same companies in the years of all those balance sheets and income statements.

3:41So did it spit out something that said buy, sell, or hold? No, it just said up a lot, up by a little, down a lot, down by a little. Those were the answers it could give. and it would say, and I'm very sure. Or it would say, this one I'm not so sure on. And then these business professors did something clever, which is they took the earnings expectations that the chat GPT was most certain about and built long, short portfolios. So it did better than just the index, right? They did pretty well. They generated alpha versus the broad stock market. I talked to one of the authors, a guy called Alex Kim, And, you know, he said, look, we're not claiming to have solved stock picking forever.

4:32You know, what we've shown here is just a kind of proof of concept that you can take a standard large language model like ChatGPT, give it some light prompting and some financial information, and it can do a bit better than the median stock picker. And the thing he was most impressed by was it was the narrative component that seemed to make the difference in the predictions. So what really mattered was when they asked ChatGPT, what is going on that is accounting for these changes in the financial statements? In other words, the prompting included asking ChatGPT, what is the economic logic of these changes you see in the financial statements?

5:19And it was only after ChatGPT had spat out answers to that question that it was able to predict with greater than human accuracy the direction that earnings were going to go. So it was like this kind of human-like storytelling thing that made the difference. Yeah, because I mean, you know, this is not an original observation, but it can do some clever stuff.

5:45Yeah, when it draws us, we have six fingers and two heads, but it's good at certain other tasks. And it tells us to put glue on our pizza. But anyway, so it is, I mean, these guys are not shouting from the rooftops, but the point is, it's sort of indicative that there might be something here for financial markets in these large language models. One of the lines from the piece that you put out on this today jumped out at me, which was that in back tests, it outperformed on a capitalization-weighted basis. It outperformed the market by 37 basis points. That's 0.37 percentage points. A month. A month.

6:27A month. Exactly. So that compounds over the year, right, if you're doing that. So it comes out to a fairly meaty bit of outperformance. although I should note as the authors point out this is before trading costs. So you have to assume that gets seriously haircut down when you're actually paying some broker to buy and sell all these stocks. But in addition the amount by which it outperforms the market is bigger when you take an equal weighted market view which kind of implies that it's better at picking out opportunities in smaller stocks than humans. Professor Kim who I spoke to was quite struck by this result.

7:08The natural theory of why this is, is that there's more inefficiency at the smaller end of the spectrum. The market has been less well scrubbed and priced carefully. But again, the authors were like, well, this is something we see in the results, but we don't quite understand why this happens. That's one theory of what is going on there, that the small cap market is less efficient than the large cap market. So the reason that line in your piece really jumped out at me is that, first of all, that is interesting that it seems to work better for small stocks. But also those three words in back tests.

7:45Yes. Give me pause. Right. Because everything seems to work really nicely in theory. You can backtest any kind of model and it says, look, you're going to make eleventy squillion dollars if you do this. And then you try and do it in real life and you get chewed out by either trading costs, as you say, or your your opportunity gets eaten up by the rest of the market going the other way. Absolutely. Or you end up with too many models trying to do the same thing at once and it all just kind of falls over itself. All of these things are true. And there is actually another point that I would make about this.

8:17Right. is there are nerds all over the world who are trying to do stuff exactly like these authors did in this paper. And they are talking to ChatGPT and giving it bits of information and prompting it a certain way and seeing what it can do in the financial markets. I really don't want to impugn the excellent work that these guys did, But if a lot of people are doing these studies and only a few studies are getting published, it might be the millions that didn't work. They just disappear. And the one that happened to work appears. Like all the newsletters that you start writing and you think, no, this is rubbish.

9:00And you put it in the bin. Well, I don't write them. It's my team of monkeys that writes them. I have an instant number of monkeys writing the unhedged newsletter. I just hope one of them comes up with something good on a given day. it's sort of more interesting that it can happen, at least temporarily, and what it tells us about what we are doing when we are picking stocks, that the computer can do it a bit better than we can. It can pick out opportunities where our stupid human biases get in the way of seeing them, I guess. Yes. This is why I said this study is not necessarily terribly surprising, Because we've known for a long time that simple rules or algorithms or earlier computer systems can outperform humans on tasks like this, not because they're intelligent, but because the people are so dumb.

9:56In other words, I don't call myself a stock picker. And part of the reason for that is I know I get wrapped up in these biases and narratives and misconceptions. But also your stock picks are terrible. Yeah, they are terrible. As we discussed on this podcast. Yeah, no, no. I'm a person who can't get out of my own way. And what computers are good at is getting out of their own way. Just look at things dispassionately. Yeah. But so I remember I was on a panel once and somebody on it pointed out that this whole AI revolution, right, it will take jobs away from doctors, right, in diagnosing what is wrong with you.

10:39Like the professionals, the highly paid professionals are going to get chewed up by robots to some degree or other. But they won't get rid of the nurses that are there to give you injections and wipe your nose. So it is quite interesting what sorts of professions are going to get kind of eaten by robots and which ones aren't. And we don't know what the economic impact of that really is yet. Correct. I always think about it in the context of what to tell my kids to do for a living. Right? The nurse's example is perfect. I always think about, you know, what is a job where you're having contact with another human being face to face and interacting with them?

11:22You know, that you are teaching them or giving them physical therapy or psychological therapy or you are. Alternatively, maybe you're interacting with meat world, as we call the non-digital world. You're building stuff or fixing stuff or hammering nails into things or whatever. Yeah. Do fund managers need to welcome their new robot overlords? I mean, one thing that fund managers and big asset management firms say to me about this is that they absolutely are super keen on finding out what the technology can do for them. And what they think that it can do for them is basically replace vast teams of junior analysts who are normally employed to scour through earnings releases and listen to all the earnings calls and just pick out little details and tell the people who push the button to say whether to buy or sell something.

12:16It's that kind of really labor intensive, but also kind of brain intensive work that will probably get stripped away. And then you have to ask yourself, what are the junior positions now in asset management or in law or all of those sorts of things when they are replaced by chat GPT? But I don't know. What do you think? Can you imagine there being a fund out there that's entirely run by AI? Well, we don't have to imagine it because there are funds like this or funds very close to this. And indeed, people have been using computer technology to seek statistical regularities and vast oceans of data to beat the market for a long time.

12:59The most famous example being Jim Simon's Renaissance Technologies, which is the fund that made more money than all the other funds. I am haunted by something that he said. When I think about fundamental stock picking and fundamentally think about markets, he was once quoted as saying the correlations that his computers discover that allow them to make the most money make no apparent sense whatsoever. And he said if they made sense, somebody would have discovered them already. So our thing is like the machine tells us to do some weird ass thing. We just do it. Right. So what I mean to suggest in pointing that out is that technologies like chat GPT will eliminate a lot of the very kind of picky analytical tasks of, you know, forming a binary or semi binary earnings forecast for the next six months or a year or whatever.

14:08But what I would note is that a lot, and you know this just as well as I do, that a lot of the very best investors scrupulously ignore these tiny questions about whether EPS is going to go up 2 % next year or down 2%. They're like, all of that is nonsense. We need to think about the big fundamentals of whether this business is structurally strong, what its competitive position is, and how the big world is changing that is going to put this company in a better position or a weaker position. Right. And it's interesting me to think about, yes, chat GPT is better at the grinding statistical task of making an earnings estimate.

14:49But is does it turn out to be equally good at making the kind of big calls where the best investors make the most money? And maybe it will be for all I know. I'm not ruling it out. What I am saying is that it's a different kind of task. It is a different kind of task. Yeah. I think a more interesting question is, can technologies like this help the small minority of people who are pretty good at it? Like, there does seem to be people, not including myself, who are good investors, right? And can technologies like this add to their game? I, for one, very much look forward to adding to your game in this year's stop-picking competition for the FT, at which you have bombed incredibly badly in the previous years.

15:35That was just one year. I almost won the whole thing the year before that. My monkeys back in my office are working very hard, not only to write the column, but to pick my stocks. The amount of money I'm spending on monkey chow is really expensive, so I better get a win out of this. no excuse you must do better okay we're going to be back in a minute when also you must do better for long short

16:08all righty it is time for long short that part of the show where we go long a thing we love short a thing we hate rob i will just tell you that last week i was short slugs which are eating my garden and are evil. And someone emailed me to say that someone in their family used to cut them in half with a pair of scissors and leave them for the birds. I thought you were going to say it left them porridge. How do you say this in French? Pour and courager les autres? Right? As a message to the other slugs. No. Just straight up bird food. They should have put the dead slug on a spike in your garden.

16:46I thought that was a bit harsh. Yeah, that's a lot. That is a lot. Yeah, but what are you long or short this week? I am long Elon Musk getting rich. Come on, man. I know he's already, I think he's sunk to a miserable number two in the world. But there's been this great kerfuffle about his$56 billion pay package that he negotiated back in 2018. And people didn't quite realize back then that if all went right for him, he would get$56 billion of the company's money for his labors. and now various investors are like um maybe we shouldn't do that but i am invoking the rule of the playground no backsies the company said they would pay him shareholders were okay with it at the time you got to give the man his 56 billion dollars even though this was the dumbest pay package of all time you agreed to it it happens give the man his money sometimes you have to do dumb stuff.

17:45Because you said you would. Fair enough. Fair enough. So I am also long. I did one of those weekends where you actually do things in London that make it worth living in such an expensive city. I went to see Yuja Wang, the Chinese pianist. I don't know if you've heard of her, but she's like a thing from another planet who has beamed down to do things that humans can't do with a piano and she's just amazing. Awesome. Ten encores. Ten encores. Ten. Cool. She really rocked that place. Ten encores. How does that work? She played absolutely astonishing music, and then there was a standing ovation, and then she kept coming back onto the stage and kind of leaning on the piano and kind of drumming her fingers and going, okay, fine, I'll do another one.

18:28And then she would play something else just completely amazing. Were you there until four in the morning? No, but I think it did overrun a fair amount. But if you don't know this woman, her first name is spelled Y-U-J-A. Okay. Google her. She's absolutely astonishing. I'll give it a listen. Little cultural lesson for you there, Armstrong. So, listeners, unless you would like to request an encore of this little chat that I've had with Rob this week, then that is us for today. But tune in on Thursday for more.

19:04Unhedged is produced by Jake Harper and edited by Brian Erstad. Our executive producer is Jacob Goldstein. We had additional help from Topher Forhears. Cheryl Brumley is the FT's global head of audio. Special thanks to Laura Clark, Alastair Mackey, Greta Cohn and Natalie Sadler. FT Premium subscribers can get the Unhedged newsletter for free. A 30-day free trial is available to everyone else. Just go to ft.com slash unhedged offer. I'm Katie Martin. Thanks for listening.

19:45We'll be right back.