Money Stuff: There’s a Robot Warren Buffett ETF

What is Warren Buffett? Like, if you were trying to describe and recreate his essential features, how would you do it? I guess a decent answer is “he is fully described by his DNA, so I would simply get his DNA and clone him in a lab,” but that does not seem quite possible on current technology and anyway let’s try for something more parsimonious. We do not really need to capture his physical appearance and personal foibles and ineffable humanity; we are interested in Warren Buffett because he is good at investing, and we just want to capture that. What we want is some way to reduce the complex humanity of Warren Buffett to some reproducible formula that picks stocks the way he picks stocks.

A lot of research in finance aims to answer questions like this, how to measure and describe someone’s investing approach. There’s a kind of famous paper from Andrea Frazzini, David Kabiller and Lasse Pedersen at AQR, titled “Buffett’s Alpha,” that “decompos[es] Buffett’s performance into its components due to leverage, shares in publicly traded equity, and wholly owned companies,” “reveals that Berkshire Hathaway loads significantly on the BAB and QMJ factors,” [1] and finds that “these factors almost completely explain the performance of Buffett’s public portfolio.” Warren Buffett can be reduced to a set of coefficients on various common investing factors, plus a smallish residue of ineffable humanity. [2]

And that was sort of the state of the art of quantitative finance a decade ago, coefficients on factors. In 2024, though, we have large language models, and Warren Buffett is voluble. He has, over the years, written a lot of shareholder letters describing his investment philosophy and his views about particular stocks, and you can learn a lot from them. People do: The letters are popular, and aspiring investors regularly read Buffett’s back catalogue to learn how to think and invest like him.

They do this not only because his stock picks are good, but also because his prose is good. The letters are clear, entertaining and compelling explanations of how Buffett invests; he does not just give the particular reason why he likes some particular stock, but he explains it in a generalizable way. The letters convey a sense of his mind. They could, perhaps, teach you some general skills; reading his descriptions of the stocks he picks could teach you how to think about other stocks, stocks he’s never even thought about, but that you can think about by copying his habits of mind. You can learn to make Warren Buffett-like decisions in your own investing, in situations that Warren Buffett himself has never faced.

And, again, people do that; there are a lot of investors who fancy themselves the next Warren Buffett. But that takes work — you have to read all the letters and apply them with a certain amount of trial and error [3] — and with modern technology it does seem like there would be a shortcut. The shortcut is:

1. You feed all the letters into ChatGPT.
2. You ask it “recommend me a stock in the style of Warren Buffett.”
3. It picks a stock that Buffett would pick, and writes up the thesis for the investment in the way that Buffett would.
4. You go buy the stock, because Warren Buffett — refracted through a model of his mind based on his writings — recommended it to you.
5. It goes up, because Warren Buffett is good at picking stocks, and his writings provide a good model of his mind, and ChatGPT is good at building that model from those writings.

This thesis feels (1) obvious but also (2) insane? I have been suggesting it for years, and I’ve been embarrassed each time. [4]  Back in 2020, when GPT-2 was the state of the art in large language models, I wrote:

If you trained GPT-2 on a bunch of Money Stuff columns it could probably say some stuff that sounds like Money Stuff. If you trained it on a bunch of Warren Buffett annual letters maybe it would say some stuff that sounds like Warren Buffett? Not just in terms of folksy sex jokes but also in terms of penetrating investment insight? Maybe GPT-2 would digest Buffett’s mind, or rather specifically the parts of Buffett’s mind that are exposed when he writes prose, and it would use that understanding of his mind to write Buffett-like prose recommending Buffett-like investing decisions? 

Or if you run an investment firm and you’ve got a corpus of memos from your analysts recommending investment decisions, why not take the memos that worked out—the ones recommending investments that went up—and feed them into GPT-2? Then have it write you a new memo and see if it’s any good?

I mean it’s still kind of a joke, obviously. The normal approach to artificial intelligence in finance is, you know, the computer looks at stocks that went up and tries to spot patterns, to figure out what the stocks that went up had in common. Doing that through the medium of prose—look at the memos recommending stocks that go up and figure out what they had in common—is a weird form of indirection, a dumb and unnecessary complication. But sort of a charming one? One criticism that you sometimes see of artificial intelligence in finance is that the computer is a black box that picks stocks for reasons its human users can’t understand: The computer’s reasoning process is opaque, and so you can’t be confident that it is picking stocks for good reasons or due to spurious correlations. Making the computer write you an investment memo solves that problem! 

Would it work? On current technology? I have my doubts. (As a dumb empirical test, this morning I asked ChatGPT to channel Buffett and write a memo evaluating Trump Media & Technology Group’s stock and, woof, no. [5] ) Still there is something philosophically satisfying about it. Reducing Buffett to a set of factor loadings feels somehow unnatural; the unexplained residue is where all the magic happens. Reducing Buffett to his words, and modeling his mind based on those words, feels more natural and human; maybe his style is where the magic happens, and maybe the computer can capture that.

Anyway!

Fintech startup Intelligent Alpha is launching a chatbot-powered ETF that promises to harness the brainpower of the investment world’s most illustrious minds — Warren Buffett, Stanley Druckenmiller, David Tepper, and more.

With the not-so-subtle moniker — the Intelligent Livermore ETF – the product is built around investment ideas generated by ChatGPT, Gemini and Claude, dubbed the “investment committee,” that are set to be inspired by the thinkings and doings of the famed money managers. It starts trading on Wednesday.

The firm, with roots in engineering and emerging technologies, will instruct the large language models (LLMs) to emulate the investors’ personalities. The trio of chatbots will spit out 60 to 90 global firms that span a number of sectors, themes and geographies, including health care, renewables and Latin America, to name just a few.

The list of personas targeted by the ETF — besides Buffett, Druckenmiller and Tepper — will include Dan Loeb, Paul Singer and others, though the fund’s holdings may not necessarily reflect the real-life bets by those investors. 

I could not love this more. Here is the press release. Here is the prospectus, which describes the approach:

A human analyst (the “Analyst”) establishes the Intended Strategy for the underlying portfolio. …

A large language model, which is a type of AI algorithm that uses deep learning techniques and massively large data sets to understand, summarize, generate and predict new content, is consulted to identify 4-6 major trading trends inspired by the greatest traders in the world. The Analyst will define the list of famous traders and investors for the AI by reviewing long-term (5 years or greater) track records of famous investors as compared to broad market benchmarks, which will vary based on the nature of the trading trend(s) observed. A major trading trend is a clearly defined and articulated trading view expressed by the famous trader. …

The AI will analyze numerous information sources (such as 13F filings, public statements, and interviews) relating to the identified traders and trends when conducting its overall analysis. …

The Analyst gathers the details, information and Philosophy set forth in Step 1 and translates them into an instruction set to be submitted to three large language models (the “AI Models”) for portfolio creation. Each AI Model is similarly instructed to review the data and instructions to create a portfolio of up to 20-30 stocks each, including weights for each position. The AI Models are instructed to limit the maximum weight of any one holding to 10%. Each AI Model has its own discretion in creating its portfolio, which is reviewed by the Analyst in Step 3. The AI Models represent the Fund’s AI investment committee, whereby the AI Models provide their independently created portfolios based on each AI Model’s independent evaluation of the instructions provided to it.

Et cetera. There is a lot of human intervention; this seems more like “ask ChatGPT what Buffett would do and then noodle on the answer” than “ask ChatGPT to pick the stocks Buffett would pick and then buy them.” 

They don’t really name the investors — I guess they’d get sued? — but the proposed ETFs include the Intelligent Livermore ETF (ticker: LIVR) and the Intelligent Omaha ETF (ticker: AIWB [6] ), which I assume will train on Buffett’s writings. I don’t really know what it would mean to train on Jesse Livermore’s writings. Livermore apparently wrote a book, though I feel like the Jesse Livermore book that everyone reads is the one that he didn’t write. And he published his book in 1940, a few months before he died, so it’s not clear how he would trade the Magnificent 7 or whatever. But I guess the ETF will find out!

Stock investing is so obviously a domain in which it makes sense to apply artificial intelligence:

1. There is a ton of data.
2. Humans have pretty well-known weaknesses in dealing with it — limited memories and processing power, emotions, biases, lunch breaks — and a computer could avoid those weaknesses.
3. Everything happens on computers anyway. You don’t have to, like, avoid pedestrians in a self-driving car; you just generate buy and sell orders that get routed by computer to an exchange’s matching engine.
4. If you get it right you make a ton of money.

And when you put those things together, it is so natural to think “yes, I will build a robot that is better than people, that is more rational and logical and data-driven and calculating than people, that finds signals that people miss, that can do things that no human investor can do.” Your computer takes the raw information of financial markets and turns it into buying signals, free from the biases and weaknesses of human intelligence.

This is not that! This is the other, more soothing approach: “Surely no computer can be better at reading the spirits of the market than a 94-year-old guy in Omaha, or for that matter a charming and roguish day trader who died 80 years ago, so what we need to do is to program a computer to think like those guys.”

What would it mean if it’s true that the best an investing robot could do is imitate those guys? What would it mean if imitating those guys is easier than applying AI directly to market data? “Humans are better than machines at investing, but humans are easy to reduce to machines”? I don’t know, I love it.

What is Steve Cohen? As with the other founders of giant multimanager hedge funds, you could tell a basic story like:

1. He’s a guy who was really, really good at picking stocks that would go up, on a fairly rapid basis.
2. He got so good at it, and raised so much money, that he had to hire other people to pick stocks for him just to manage all that money.
3. Not everyone can make this transition: The skill of picking stocks is not identical to the skill of picking people to pick stocks. Identifying, training, managing and allocating capital to good portfolio managers has some overlaps with identifying good stocks, but the overlaps are not perfect, and many good stock pickers will not also be good managers and allocators at multimanager funds.
4. But Cohen made the transition very successfully (with, fine, some hiccups), and now manages a giant firm, Point72, with a lot of managers, a good track record, and a commitment to thoughtful processes for picking and training managers.

On the one hand, picking and managing portfolio managers does seem to be an even more lucrative, higher-leverage skill than picking stocks, which is why the really top hedge fund managers tend to end up doing that. On the other hand, if you got to that position by being really good at picking stocks, and if you then spend a lot of time thoughtfully evaluating stock pickers, you might reasonably come to the conclusion: “Wait, I’m better than these guys are.” An important constraint on the big multimanager firms is that there are just not that many people who are consistently really good at picking stocks, but the boss is one of them.

And so Cohen for a long time kept his hand in picking stocks too. But now he’s stopping:

While the billionaire hedge fund founder remains Point72 Asset Management’s co-chief investment officer along with Harry Schwefel, he’s no longer investing clients’ capital. Cohen, 68, is instead focused on driving the firm’s growth and mentoring and developing talent, the firm said in an emailed statement.

Cohen has been one of the dominant forces in the industry for more than three decades and rebuilt his hedge fund into one of the world’s biggest after a costly insider-trading scandal. Even as he grew his firm into one with more than 185 trading teams and branched out into other interests, including his 2020 purchase of the New York Mets, he retained a small book that he traded regularly.

“There’s huge value in having Steve as an impactful mentor for our investment professionals,” Point72 spokesperson Tiffany Galvin-Cohen said in the statement, adding that he’s “engaged as ever.” “The only change here is that he will spend less time in front of his computer screens and more time working with investment teams and developing talent at the firm.”

I hope they’ve had a robot follow him around for the last year, observing him closely, and now the robot will take over his book.

A lot of US antitrust law is about preventing two companies from doing a merger if the merger would be bad for competition. If one supermarket chain wants to buy another supermarket chain, the Federal Trade Commission or the Department of Justice will review the merger, and if the FTC decides that it would create a monopoly, it will try to block the deal.

Antitrust law is about other things, too, and if a company grows organically to become a giant monopoly, the FTC and DOJ will take an interest and maybe even try to break it up. But it is generally much harder to break up an existing company than it is to stop a merger that hasn’t happened yet, so a lot of the action is in reviewing and blocking mergers.

And so it is very important for antitrust enforcement that the regulators get a chance to review mergers before they happen. If two companies merged without telling anyone, and then the FTC looked into it and decided the deal was bad for competition, it might sue to break them up, but by the time the case was heard the two companies might be fully integrated with each other and breaking them up would be really disruptive. 

And so there is a rule, the Hart-Scott Rodino Antitrust Improvements Act of 1976, which says that, before two companies merge, they have to file a notice with the FTC saying that they intend to merge, and the FTC gets 30 days to decide what to do about it. And usually the FTC is like “yeah that's fine” and they go ahead and merge, but sometimes the FTC asks for more information, and occasionally it demands changes in the deal or tries to block it altogether. And all of this happens before the merger closes.

Now there are a few technical notes on this. First, the rule does not apply to every merger, because a lot of mergers are too small for the FTC to care about. There is a threshold for HSR notification; it changes each year but it’s currently $119.5 million. If you are buying a $100 million company, go ahead; if you’re buying a $200 million company, you have to tell the FTC first. (Not legal advice!)

Second, the rule does not apply only to mergers. There are lots of other ways to buy control of a business, besides a merger. You can buy assets from a company. Or you can buy stock in an existing company: Buying 51% of the stock of a competitor is more or less as good as acquiring that competitor in a merger, and buying 45% of the stock is pretty close too. So the HSR rules apply to buying stock. If you buy more than $119.5 million worth of stock in a public company, you have to file HSR.

Well, no, that can’t really be true. Much of the time, that’s just not very much stock: Buying $119.5 million of Apple Inc. stock would give you less than 0.004% of its outstanding stock, which is not at all like buying control. BlackRock Inc.’s index funds own more than $119.5 million worth of stock of pretty much every large public company, and nobody could possibly think that is an antitrust problem. Ha, I’m kidding, lots of people think that. But, still, the antitrust rules recognize that, most of the time, when someone buys a $120 million block of public stock, they are not looking to acquire control of a business; they are looking to make a financial investment in a company. There’s no need for antitrust review there. And so you don’t need to make an HSR  filing if (1) you buy less than 10% of the stock of a company and (2) you are doing it “solely for the purpose of investment.”

And so if you’re BlackRock buying 5% of a company as a passive index investor, no filing. If you’re an aggressive acquirer buying 9% of a large competitor to influence their business and maybe as a prelude to a merger, though, you are not doing it “solely for the purpose of investment,” and you have to file HSR.

And then there is an odd gray area, which is: What if you are an activist hedge fund? You might buy 9% of a public company with the intention of influencing its business, changing its management and strategy. But you are not really an antitrust problem, are you? You are not a competitor of the company, or a supplier or customer. You’re just a hedge fund, a financial investor. The FTC should not really be concerned with your activism.

And on the one hand, that’s clearly right: The FTC is not going to block an activist hedge fund’s acquisition of 9% of a public company. [7]  But on the other hand, the FTC gets to make that decision: The rule is that you file for HSR, and the FTC reads your filing and says “yeah obviously this is fine,” and you go ahead and buy the stock. But you do have to make the filing first. 

And activist hedge funds forget to do this, with some regularity, and then they get fined by the FTC. Because this is so far afield from where we started, with “a lot of US antitrust law is about preventing two companies from doing a merger if the merger would be bad for competition.” This has nothing to do with mergers, and nothing to do with competition: This is not one company acquiring a competitor, but rather a hedge fund making an investment in a single company. But the rules are the rules, and they make sense in the broad context of preventing anticompetitive mergers, even though they don’t make sense in this particular context.

Anyway it would be extraordinarily, extraordinarily, extraordinarily funny if (1) GameStop Corp. merged with Wells Fargo & Co. and (2) that created a monopoly in banking and/or videogame retailing, but that is not what is going on here:

Today, the Federal Trade Commission announced that Ryan Cohen, managing partner of RC Ventures, LLC, and Chairman and CEO of GameStop Corp., will pay a $985,320 civil penalty to settle charges that his acquisition of Wells Fargo & Company (Wells Fargo) shares violated the Hart-Scott-Rodino (HSR) Act.

According to the complaint, Cohen, who is also the founder and former CEO of Chewy, Inc., acquired more than 562,000 Wells Fargo voting securities resulting in aggregated holdings of Wells Fargo securities that exceeded HSR filing thresholds. Cohen’s purchase triggered an obligation to file an HSR form with federal antitrust agencies and wait before completing the acquisition. Yet Cohen failed to do so, which violated the HSR Act, according to the complaint.

Here is the complaint. He started buying Wells Fargo stock in 2016, and in February 2018 he emailed the company with some business tips and a suggestion that he join the board of directors. In March 2018, he bought a bit more stock, which tipped him over the relevant threshold of $168.8 million at the time. [8]  He never seems to have gotten to even 1% of the stock, but because he was doing (pretty modest) activism — emailing about joining the board of directors — he did not own it “solely for the purpose of investment,” so he had to file an HSR notification. And he didn’t, perhaps because he understandably didn’t think there was any antitrust problem with him buying some Wells Fargo stock. And there wasn’t, not really, but there was a filing the form problem.

I don’t know how important it was to the fraud at FTX Trading Ltd., the big crypto exchange that collapsed in 2022, that it had audited financial statements. FTX was not a public company, and its financial statements were not released publicly; crypto investors did not trust FTX with billions of dollars of their crypto because they read and relied on its audit reports.

FTX did have audited financial statements, though, and it sent those to its equity investors, venture capitalists and others who invested hundreds of millions of dollars in FTX at as much as a $32 billion valuation. I suppose that VC backing, and that valuation, did make crypto traders feel more secure about trusting FTX with their money (“after all, there’s a $32 billion equity cushion”), so if the equity investors relied on the audits then perhaps the audits were indirectly responsible for the collapse. Did they? Eh, maybe. My impression is that FTX’s investors were much more interested in Sam Bankman-Fried’s vision and personality, and the general crypto gold rush atmosphere, than they were in rigorous due diligence and audited financials. But, sure, FTX got audits — from a firm called Prager Metis CPAs LLC, not one of the Big Four accounting firms — which at least suggests that somebody cared about seeing audited results.

Whether or not the audits mattered, though, it does seem fair to say that they were bad. I have not seen FTX’s audited financial statements, but I have seen the infamous quasi-balance sheet that Bankman-Fried sent around in November 2022, looking to raise more money because there was — in the words of the balance sheet — a “hidden, poorly internally labeled ‘fiat@’ account” with a balance of negative $8 billion. If you raise money at a $32 billion valuation with tidy audited financial statements, and then later you try and fail to raise money at a $0 valuation with an Excel spreadsheet saying “uh we forgot we misplaced $8 billion,” your audit missed something important. 

Yesterday the US Securities and Exchange Commission got around to fining Prager Metis $745,000 for those audits:

According to the SEC’s complaint, from February 2021 to April 2022, Prager issued two audit reports for FTX that falsely misrepresented that the audits complied with Generally Accepted Auditing Standards (GAAS). The SEC alleges that Prager failed to follow GAAS and its own policies and procedures by, among other deficiencies, not adequately assessing whether it had the competency and resources to undertake the audit of FTX. According to the complaint, this quality control failure led to Prager failing to comply with GAAS in multiple aspects of the audit—most significantly by failing to understand the increased risk stemming from the relationship between FTX and Alameda Research LLC, a crypto hedge fund controlled by FTX’s CEO.

Yes I mean it would be weird if Prager did follow all of the appropriate auditing standards and missed that!

We talk all the time about here about the US Securities and Exchange Commission’s crackdown on financial firms whose employees sometimes text about work from their personal cell phones. (Which is all of them.) For instance, I wrote yesterday about some municipal bond advisers who paid the SEC five- or six-digit fines for “describing current municipal market conditions” to clients over text messaging. This strikes me as pretty harmless behavior, but rules are rules I guess, and the SEC will fine firms whose employees text about work.

A reader emailed:

Last week a coworker arrived at the office early, saw rainwater pouring in through the ceiling, and texted the manager something along the lines of OMG we have water pouring through the ceiling.

She has now had to endure several calls from compliance and has to write a report detailing why she sent a text about company business. 

I’m sure this is exactly the kind of situation the investment act of 1933 was enacted to prevent. 

Yes this is how financial industry compliance works now.

Fifty basis points. Hedge Fund Titans Breed a $14 Billion Pack of Startup Cubs. Citadel Securities Shelves Plan to Join Ranks of Fed Dealers. SEC Poised to Vote on Overhaul of Stock Pricing, Exchange Fees. Switzerland’s SIX explores launching crypto exchange. Banks Ready to Forget Pain of Hung Debt Go Big on Buyout Funding. BlackRock, Microsoft Partner on Massive New AI Infrastructure Fund. Looming Insurance Crisis Threatens Taxis and Ubers in New York City. Liverwurst discontinued. Train cancelled after squirrels board and ‘refuse to leave.’

If you'd like to get Money Stuff in handy email form, right in your inbox, please subscribe at this link. Or you can subscribe to Money Stuff and other great Bloomberg newsletters here. Thanks!