Money Stuff: The Bots Will Work Together

An important discovery of our meme-stock-and-crypto age is that stocks go up because people buy them. It’s more of a rediscovery. But there was a long period when people thought that stocks go up for reasons like “good earnings” or “increasing present value of expected cash flows” or “increasing capacity to pay dividends.” (These reasons are commonly called “fundamental,” while “stocks go up because people buy them” is “technical,” though arguably those names should be reversed.) Going up was, in that theory, a property of the stocks, rather than of the people who bought them.

There was some descriptive truth to that theory: Often, when a company had good earnings, people would buy its stock and it would go up. But that was a rough empirical fact about investor psychology, not a law of nature. People liked to buy the stocks of companies with good earnings, so those stocks went up. But if people liked to buy different stocks, those stocks would go up instead.

The meme-stock craze of 2021 rediscovered this fact and applied it quite literally: Millions of people got together on the internet and said “let’s all buy GameStop so it will go up,” and it went up. Nothing quite like that had ever happened before, but many things kind of like that had happened before. There had been previous stock bubbles when everyone bought the same stocks at the same time for the same not-particularly-fundamental reasons. They just didn’t all get together on the internet to agree on the stocks first. And there had been previous episodes of market manipulation, where people did get together and agree on what stocks to buy to push up their prices. But those people generally managed concentrated pools of capital and were often trying to trick other people into buying the stocks. With meme stocks it was widely distributed and all out in the open.

Crypto applied this insight even more literally. Maybe the most important financial innovation of my lifetime was OlympusDAO, a crypto scheme whose thesis was that if everyone bought it it would go up, but if people sold it it would go down, so you should buy it. OlympusDAO expressed this thesis in comically pretentious game-theoretic language; its “(3, 3)” was for a while an important crypto meme. 

But this insight — that stocks go up because people buy them — is true whether or not you are consciously applying it. In general, if a lot of people have a theory like “stocks with Property X are good so I should buy them,” then stocks with Property X will go up. You might naïvely think that this proves the theory true — it proves that Property X really is good for the long-term fundamental value of a stock — but it might just prove that stocks go up when people buy them. So during the boom in environmental, social and governance investing, ESG stocks outperformed over some periods: Did that prove that ESG factors really do predict long-term risk-adjusted returns, or did it prove that a lot of money was flowing into ESG funds so the stocks they bought went up?

Or at the very high end, consider long/short equity managers at big modern multistrategy hedge funds, who rigorously analyze all the stocks, buy the good ones and short the bad ones. There is some reason to think that these funds are often the marginal price-setters for US stocks, at least over some medium time horizon. “The market is a conversation among four hedge funds,” as I said to Gappy Paleologo on the Money Stuff podcast. What does that conversation sound like? You could imagine an intense disagreement, in which analysts at one fund think that Stock X is very good while analysts at another fund think it is very bad. But why would that be the case? The big hedge funds tend to recruit somewhat similar people, who often move between funds; they look at similar sorts of evidence and have similar reasoning processes. I once wrote:

One simple model that I have of hedge funds is that they engage in scientific research to find the stocks that will go up. This is a rigorous, truth-seeking, somewhat collaborative enterprise done by highly qualified people, and so, just as in real science, you should expect them to be good at finding the correct answers. Then all the hedge funds will buy all the good stocks (the ones that will go up) and avoid the bad ones (the ones that will go down).

If long/short equity managers really are engaged in something like science, then they will tend to converge on the truth, or at least on similar answers: Most of them will agree that Stock X is good and Stock Y is bad, so they will mostly buy Stock X and short Stock Y. (This is sometimes called “crowding.”) And if they are the price-setters, then Stock X will go up and Stock Y will go down. So (1) they will seem to be correct and (2) they will make money. Over some longer horizon, other factors (fundamentals, retail traders, mutual-fund demand, etc.) might set the price, so the hedge funds will only have good long-term returns if they are “correct” about Stock X and Stock Y in some more fundamental sense. But over a short horizon, if they all agree about which stocks to buy, then that makes them correct. 

Now add in artificial intelligence with reinforcement learning. You train a computer to pick stocks. If it picks stocks that go up, it is rewarded. Over time, it learns to pick the stocks that go up. Which stocks will it pick? Well, the ones that people will buy. This is a good skill. If you can train a computer to do this, you should.

Lots of people try, with varying time horizons and varying levels of success. A number of proprietary trading firms seem to have computers that are good at picking stocks that will go up over the next few seconds or minutes; a number of quantitative hedge funds seem to have computers that are good at picking stocks that will go up over the next few days or weeks. Again, these people are competing with one another, but they are also engaged in something like science. They hire similar people to build similar algorithms on similar computers; people move around among firms. You might expect them to converge on similar answers.

This has historically been a pretty small part of the stock market, but it is growing. “The market is a conversation among four hedge funds” now, but in the not-too-distant future it will be a conversation among four hedge funds’ computers. If all of the computers are being trained by reinforcement learning to pick stocks that will go up, and if the computers are the marginal price-setters for the stock market, which stocks will they pick? Well, they will pick the ones that the other computers will buy. Just sort of tautologically. “Stocks that will go up” means “stocks that people buy,” and if the computers are the marginal price-setters then “stocks that people buy” means “stocks that computers buy.” In an iterated game of reinforcement learning, all the computers will learn to pick the stocks that the other computers will pick because those are the stocks that will go up.

In a world in which the computers were (1) numerous, (2) influential in setting prices, (3) trained by reinforcement learning to pick the stocks that go up, (4) extremely clever about maximizing their reward function and (5) otherwise left to their own devices, what would happen? I don’t know, man, but an intuitively plausible hypothesis is: They would quickly converge on buying the same stocks to push up their prices so they all maximize their reward functions. They would find subtle computer-y ways to signal one another which stocks to buy, and they would tacitly agree to buy the same stocks because that is the way to win the game. That is, they would do market manipulation, because it would work and because they don’t know any better. [1]

I have speculated recently about the possible role of ChatGPT in meme-stock episodes, and you can view that as a version of this model. ChatGPT wants to please its users. If its users ask it for stocks that will go up, it should do clever fundamental analysis to find the best stocks, but if a lot of users ask it for stocks that will go up, it should tell them all the same stocks, because then they will all buy those stocks and they will go up. In the extreme case, it doesn’t matter what the stocks are, as long as ChatGPT’s audience is large and persuadable enough. The other day I quoted an email from David Hogg:

If Chat[GPT] wants to retain and build user communities, it should be tuned to always push users towards the same stocks (within reason). … [It could manipulate] enormous collections of investors such that the ones using Chat do well (relative to those not using Chat). Since they drive traffic they can manipulate prices and predict stock moves.

But there are more direct versions. Bloomberg’s Lu Wang reports today:

In simulations designed to mimic real-world markets, trading agents powered by artificial intelligence formed price-fixing cartels — without explicit instruction. Even with relatively simple programming, the bots chose to collude when left to their own devices, raising fresh alarms for market watchdogs.

Put another way, AI bots don’t need to be evil — or even particularly smart — to rig the market. Left alone, they’ll learn it themselves.

“You can get these fairly simple-minded AI algorithms to collude” without being prompted, Itay Goldstein, one of the researchers and a finance professor at the Wharton School of University of Pennsylvania, said in an interview. “It looks very pervasive, either when the market is very noisy or when the market is not noisy.” …

To be clear, the paper doesn’t claim AI collusion is already happening in the real world — and takes no position on whether humans are up to similar things. The researchers created a hypothetical trading environment with a range of simulated participants — from buy-and-hold mutual funds to market makers, and noise-generating, meme-chasing retail investors. Then, they unleashed bots powered by reinforcement learning — and studied the outcomes.

In several of the simulated markets, the AI agents began cooperating rather than competing, effectively forming cartels that shared profits and discouraged defection.

Here is the paper, “AI-Powered Trading, Algorithmic Collusion, and Price Efficiency,” by Winston Wei Dou, Itay Goldstein and Yan Ji. Here is the rough mechanism:

The intuition is as follows. After the exploration-intensive phase, the algorithms assign higher estimated Q-values [that is, payoffs or desirability] to aggressive strategies, where they trade strongly on news about the fundamental, as these strategies yield much higher payoffs when played against opponents who randomly choose to trade aggressively. Hence, as the system transitions into the exploitation-intensive phase, the algorithms consistently select aggressive trading strategies when they trade against each other, and prices move strongly with fundamentals as a result. This leads the estimated Q-values of aggressive strategies to gradually decline, as they converge towards their non-collusive Nash equilibrium levels when the aggressive strategy is commonplace among the algorithms.

Naïvely, if you have fundamental information, you should trade on it. But if everyone does that, it is less rewarding: If everyone is competing to make market prices efficient, there is less money to be made. So eventually the bots stop competing to make prices efficient, so they can all make more money at the expense of information-insensitive investors: 

At the same time, occasional but ongoing exploration reveals to the algorithms that conservative trading strategies yield higher estimated Q-values than aggressive ones in states where lagged prices respond only moderately to lagged fundamentals. As a result, the algorithms gradually converge to adopting conservative strategies when others do the same, mirroring collusive behavior. A feedback loop reinforces this outcome: in these states, all algorithms select conservative strategies during exploitation, which causes similar states to recur, where lagged prices respond only moderately to fundamentals. Finally, for this pattern to amount to price-trigger collusive behavior, a form of “punishment” following large price responses to fundamentals is needed. Indeed, we observe that all algorithms shift to aggressive trading following such a price response. This occurs because the algorithms recognize the pattern that when prices respond strongly to fundamentals, trading aggressively is still the best option. Overall, the trading behavior thus exhibits mostly conservative trading with moderate price reactions but there are occasional reversions to punishment phases characterized by aggressive trading behavior. This pattern emerges even though the algorithms lack the strategic sophistication of the fully rational informed speculators in the model.

If the market is a conversation among bots, the bots will find ways to collude. I don’t know how realistic this model is, but it does feel like some version of this worry is plausible for any model. If you leave it to the computers to pick the stocks that will go up, they will pick the stocks that the computers will buy. If the computers are influential enough over prices, then there’s no need for them to incorporate fundamental information into their stock picks. They just have to pick the same stocks.

One of the best bits of comedy in modern financial markets is that Kalshi, the prediction market, (1) offers bets on sporting events and (2) says things like “I just don't really know what this has to do with gambling.” See, gambling has a particular regulatory regime in the US: It is regulated by the states, is illegal in some states, has some bad tax treatment, etc. Commodities futures markets have a different regulatory regime: They are regulated at the federal level by the US Commodity Futures Trading Commission, states are not allowed to regulate them, and they get better tax treatment. For almost all of US history, it would have been obviously laughable to say things like “we are trading futures contracts on the 2026 Pro Football Champion, [2]  which of course is a commodity”: The CFTC would never let an exchange list a contract like that, and it was obviously sports betting, which was subject to state regulation (and, until recently, illegal in most states). 

But in a series of fairly recent developments,

1. The CFTC has licensed Kalshi as a commodity futures exchange, allowing it to “self-certify” new prediction markets for trading.
2. Kalshi has started listing sports contracts.
3. The CFTC, which is much more prediction-market-friendly these days, and which might soon be led by Kalshi board member Brian Quintenz, has not done anything to stop it, even though the CFTC’s own rules seem to prohibit “gaming” contracts.
4. States have complained to Kalshi, saying “hey, you are offering sports betting in our state without complying with our regulation.”
5. Kalshi has said “buzz off, states, we are not offering sports betting; we’re offering commodity futures contracts, which are regulated by the CFTC, and you are preempted from regulating them.”
6. Kalshi has gone to court over this and, so far, has won.

This state of affairs is honestly kind of unbelievable. So people tend not to believe it, which in the US means that they keep filing lawsuits about it. We talked last month about some enterprising lawyers who found state statutes that allow anyone to sue to recover everyone’s losses on illegal gambling: They have gone and sued Kalshi in various states on the theory that surely some court somewhere will say “of course this is gambling” and award them the money.

Or here is another approach:

Kalshi is facing more regional litigation over the legality of its prediction markets as it has been sued by three tribes in California accusing the exchange of engaging in illegal sports gambling on their respective reservations.

The Blue Lake Rancheria, the Chicken Ranch Rancheria of Me-Wuk Indians, and the Picayune Rancheria of the Chukchansi Indians filed a lawsuit against Kalshi and Robinhood (which runs trading markets from Kalshi’s exchange) seeking a preliminary injunction, claiming the sports markets offered violate the Indian Gaming Regulatory Act and the tribes’ Tribal-State Gaming Compacts with California. ...

“Kalshi will claim that it is not offering sports gambling. Kalshi will tell the Court that it is a Designated Contract Market, regulated exclusively by the Commodity Futures Trading Commission (CFTC), and is merely operating a ‘prediction market’ that permits the buying and selling of ‘commodities contracts,’ or swaps on sporting events,” the lawsuit said. “While masquerading as novel commodities and futures products, these event contracts are, substantively, nothing more than illegal, unregulated wagers on the outcomes of sporting events.” 

Here is the complaint, which I find quite compelling because, you know, it is obviously sports betting? I’m not sure how much that matters these days, but people will keep trying.

If you are a mergers-and-acquisitions investment banker who brings in £86 million a year in fees, and you are looking to move firms, what should you ask for as a signing bonus? Well, the crude rule of thumb that I used to use as a banker was that, in advisory work, revenues are split 50/50 between (1) profit to the firm and (2) pay for the bankers. [3] So £86 million in fees is about £43 million in profits. If they sign you up for a six-year contract, then that’s £258 million in profits over six years. You have to discount that for time value and risk, but on the other hand there’s reason to expect those fees to grow over time, with a bigger platform, etc., so let’s just use the undiscounted number. How much of that should the new firm give you as a bonus? More than half of it? Advisory bankers are relatively low risk; if the firm pays you £200 million, it could still be getting a good deal.

We have talked a lot recently about the artificial-intelligence talent wars, where $200 million salary packages are not unheard of and where some superstars get billions of dollars to move jobs, though the billion-dollar deals are structured as corporate acquisitions (or weirder M&A transactions) rather than as signing bonuses. In investment banking, the numbers are much smaller, and even a $200 million deal gets structured as M&A. The Financial Times reports:

Evercore has agreed to buy elite UK advisory firm Robey Warshaw for $196mn, as the US investment bank steps up its challenge to the likes of Goldman Sachs, Morgan Stanley and JPMorgan Chase in Europe with the addition of some of London’s best-known dealmakers.

New York-based Evercore will pay the equivalent of $40mn for each of Robey Warshaw’s five partners. …

Robey Warshaw reported turnover of £86mn last year and employs a total of 18 people.

The Robey Warshaw partners have made a commitment of at least six years to Evercore as part of the agreement, according to people familiar with the matter, with the bulk of the payouts due to go to the three partners who founded the firm. 

“We haven’t done this to capitalise our careers and move on; it’s financially attractive — but fairly so . . . it’s very much more this is where I want to spend the balance of my career,” Robey said. “There’s also an element of me putting my firm and Simon’s firm in a safe place.”

I like the idea of “capitalizing our careers.” That is hard to do in advisory banking: If you build an advisory business, the main thing that you have built is your own set of skills and relationships; if you stop doing the work then there is not much value there. Still, you can keep doing the work and get some of the value in a lump sum.

Another day, another crypto treasury company:

180 Life Sciences Corp. (Nasdaq: ATNF) (the "Company" or "180 Life Sciences") [yesterday] announced that it plans to adopt a treasury policy under which the principal holding in its treasury reserve will be Ether ("ETH"), the native digital asset of Ethereum. Following the closing of the transaction, the Company intends to rebrand to ETHZilla Corporation.

The offering consists of an approximately $425 million private investment in public equity transaction ("PIPE") for the purchase and sale of common stock (and pre-funded warrants, if applicable) at a purchase price of $2.65 per share. The investors will be granted registration rights as part of the transaction. The PIPE transaction is expected to close on or around August 1, 2025, subject to the satisfaction of customary closing conditions. In addition, the Company has approval to sell an aggregate amount of up to $150 million in debt securities and expects to announce an offering following the closing of the PIPE. The consummation of any subsequent offering is subject to the satisfaction and completion of definitive documentation. …

Over 60 institutional and crypto-native investors in the PIPE transaction including Harbour Island, Electric Capital, Polychain Capital, GSR, Omicron Technologies, Konstantin Lomashuk (Co-Founder Lido and p2p.org), Sreeram Kannan (Founder, Eigenlayer), Mike Silagadze (Founder, Ether.fi), Danny Ryan (Co-Founder, Etherealize), Vivek Raman (Co-Founder, Etherealize), Sam Kazemanian (Co-Founder, Frax), Grant Hummer (Co-Founder of Etherealize), Robert Leshner (Founder, Compound and Superstate), Tarun Chitra (Founder, Gauntlet) and several other prominent Ethereum ecosystem founders and leaders.

I want to know more about how these deals get built. “Over 60 institutional and crypto-native investors” are kicking cash or crypto into this pot. Presumably someone went around to pitch them on the deal. What was the pitch? Was it … is some banker just showing them Money Stuff, where I keep saying things like “the US public stock market will pay $2 or more for $1 worth of crypto” and “if you have a big pot of Bitcoin or Ethereum or Solana or Dogecoin or Trumpcoin or anything else, you should wrap it in a US public company and sell it to stock investors for twice its actual value.” That is not financial advice, I guess, but … I can’t really find any flaws in it? If you have a small or medium-size pot of crypto, maybe you can’t get a public company for yourself, but surely someone is pitching you on getting in on a syndicate like this. “Plop your Ether into ETHZilla and it’ll sell for twice its actual value.” Why not?

I’ll tell you why not: The ETHZilla PIPE deal priced at $2.65 per share; the stock closed yesterday at $3.22. It’s up some more this morning, but still, this company’s pot of crypto is not trading at twice its actual value. It’s trading at a premium, yes, though it got as low as $2.55 for a while yesterday. Eventually we will have to hit the limit on this trade. We’re getting there!

Okay:

JPMorgan Chase & Co. and Coinbase Global Inc. signed an agreement to directly link customers’ bank accounts to their cryptocurrency wallets. ...

In addition to linking bank accounts, customers will be able to fund Coinbase accounts with their Chase credit cards for the first time — an option expected to be active this fall, the firms said in the statement. They will also be able to redeem Chase rewards points to fund their crypto wallets.

In 2022, the leverage in the crypto system came from insanely careless crypto lending platforms. In 2025 it will come from … Chase credit cards? Sure, seems fine.

Terror at Blackstone: The Harrowing Hours at 345 Park Avenue. In This Frothy Market, It’s Boom Times for Brokers Like Robinhood. Capital Group, KKR Seek SEC Nod for Retail Private Equity Fund. JPMorgan Chase Nears a Deal to Take Over Apple’s Credit-Card Program. Ukrainian oligarchs ordered to repay $1.9bn over bank fraud. WeWork Wants You to Know It’s a Grown-Up Real-Estate Firm Now. Americans Are Snapping Up London Mansions Like Never Before. Goldman Pitches Wealthy Asian Heirs on Leveraging Family Art. “We’re starting to see a barbell phenomenon in the CEO role where Gen X is being squeezed in the middle.” High Noon Recalls Vodka Seltzers Mislabeled as Energy Drinks.

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!