Thursday links: Efficient AI, optimistic humans, pack reveals

Optimal herding

Two Fed researchers test how greater reliance on generative AI in investing and trading is likely to impact financial stability.

The results are mixed.

On the positive side, the study found that AI agents make rational trading decisions between 61-97% of the time, compared to just 46-51% for humans.

Mostly that’s because agents base decisions on their own gathered information rather than following market trends, as so many of us do.

(Gathering information is so hard, though.)

The study says this could make markets behave more rationally: “Increased reliance on AI-powered trading advice could therefore potentially lead to fewer asset price bubbles arising from animal spirits that trade by following the herd.”

John Maynard Keynes would approve: more mathematical expectations, less spontaneous decision making.

However!

The paper also finds that AI agents are “not purely algorithmic.” Instead, they seem to have “inherited elements of human intuition and bias.”

For example, when the researchers deliberately labeled market data counterintuitively — green for down and red for up — the agents “produced few rational responses.” 

The authors attribute this to the “chaotic discourse of social media platforms such as Twitter (X) and Reddit” in the models’ training data.

(Garbage in, garbage out, I guess.)

The study cites Meta’s Llama 3 as particularly human-like, making trading decisions based significantly more on judgment and emotion. In one instance, the model stopped to question whether a market maker was trying to inflate the price of an asset it was trading.

(I’ve often wondered the same.)

But the models can also be too rational — because sometimes it’s right to follow the crowd.

“As market trends can reflect the private information of others, it can be optimal from a profit-maximizing perspective to take trading history into account when making trading decisions,” the paper explains.

Left to their own devices, agents failed to take into account what others in the market were doing, which led to “occasional suboptimal choices.”

With training, however, “AI agents can be induced to herd optimally when explicitly guided to make profit-maximizing decisions.” 

“Optimal herding” makes markets more efficient by accelerating how quickly information is incorporated into prices.

(Like when you take your money out of a bank just because you see everyone else is and then it turns out the bank was insolvent, for example.)

The authors are hopeful that agents striking just the right balance of rationality and herding could lead to more efficient markets less prone to bubbles and busts.

But they acknowledge the risks.

“Attempting to fine-tune LLMs to behave optimally can thus have unintended consequences.”

The optimism tax

An analysis of 41.6 million trades on Kalshi shows that Keynes' “spontaneous optimism” is alive and well. In prediction markets, the study finds “traders systematically overbet YES in markets that predominantly settle NO.”

This creates “a behavioral surplus” that makes it profitable for market makers to continue posting bids and offers, even in markets prone to insider trading. 

For example, the study cites the Kalshi market for “Will Jeff Bezos attend the Super Bowl?” where the price of YES fell from 70 cents to 30 cents after Bezos’ stepson mentioned to his University of Miami fraternity brothers that his family had other plans that weekend.

The fraternity brothers’ profits were the market makers’ losses.

Yet market makers still came out ahead thanks to the “optimism tax” they collect from uninformed traders. 

“This behavioral overpayment (or "optimism tax") generates a surplus that cross-subsidizes and outweighs the losses inflicted by informed traders,” the study finds. 

It’s like the long-shot bias seen in equities, horse racing, and lottery tickets — but in prediction markets, there’s an average-shot bias, too.

Even in the case of the Bezos Super Bowl market — where informed NO bettors were especially active — market makers ended up turning a profit because uninformed traders had so heavily overpaid for YES.

“The maker’s profit is a residual arising from the behavioral tax paid to her by uninformed YES takers on NO-settling markets minus the losses that adverse taker flow imposes,” the authors conclude.

This is the math that confirms Keynes’ intuition about spontaneous optimism.

It took data from Kalshi to prove it because, unlike in equities, market makers on prediction markets “overwhelmingly” hold their positions until resolution.

Meaning, they’re not earning the spread between buyers and sellers like they do in the stock market. Instead, they’re earning on buyers being wrong most of the time.

Because we all overestimate the odds of things happening.

One-armed card bandits

In an article on the gamblification of card collecting, Barron’s reports that GameStop offers a free-entry sweepstakes to win a Pokémon card worth $80,000. All you have to do is fill out an index card by hand, put it in a No. 10 envelope, and mail it to a PO Box. But good luck finding the address.

This seemingly generous offer is a legal fudge: It’s how GameStop gets around online-gambling laws that would otherwise preclude it from selling “digital pack reveals” — the gacha-style, single-card pack openings that have recently become so popular. 

Adding a “no purchase necessary” option to the offering makes its virtual pack openings a sweepstakes, which is legal, instead of a lottery, which is not.

(Only state governments, non-profits, and Native American tribes can run lotteries in the US.)

I’m guessing nearly everyone makes a purchase, though, because the no-purchase offer is buried impossibly deep in the GameStop website. And even if you do find it, the instructions are nearly incomprehensible — like reading the terms and conditions on Apple’s App Store or something.

That probably makes it illegal again because courts have held that the no-purchase-necessary option should be given “equal dignity” to the paid options. They should be easy to find and complete, basically. But these laws have rarely been enforced. 

Governments may want to revisit them.

Barron’s cites Birches Health, a gambling-addiction treatment provider, which reports a 207% increase over the past year in patients talking about “trading cards” and “digital repacks.”

It’s easy to see why, because digital repacks like GameStop offers are essentially slot machines that pay out collectible cards.

They’ve gotten very popular. On Collector Crypt, for example, users spent $49.6 million on packs in just the last week.

I’ve tried it. It’s fun.

But it can also be addictive.

— Byron Gilliam

_{Brought to you by:}

Avalanche Summit NYC returns September 16-17, bringing together the institutions, enterprises, investors, and builders turning blockchain technology into real business outcomes.

From tokenized markets and institutional finance to payments and consumer applications, the Summit will explore how production-ready infrastructure is enabling faster settlement, lower costs, and entirely new products and revenue streams.

Use promo code BLOCKWORKS15 for 15% off!