Thursday links: Betting bias, ultraforecasters, CLARITY, seed phrases, cyber risks

Prediction markets favor the favorites

Traditional betting markets are known for a long-shot bias: Bettors chasing big payouts tend to overestimate the odds of unlikely upsets.

Nate Silver says prediction markets have the opposite problem: Bettors overestimate the odds of favorites.

On Kalshi this week, Abdul El-Sayed’s odds of winning the Democratic nomination for US Senate were 97% the day before the primary. The odds that he’d win by less than three percentage points were effectively zero. 

El-Sayed did win, but by just 15,000 votes — a margin of less than 1%.

Silver says this should not be much of a surprise because the polling in congressional primaries is often “wildly” inaccurate. The lack of reliable polling was not reflected in prediction markets. 

Silver argues this is a structural issue. “People tend to treat [prediction markets] as oracular when they shouldn’t,” he explains, “perhaps producing some self-reinforcing feedback loops.”

Media outlets label someone a frontrunner, prediction markets price that in, leading to more confident reports by the media, which leads to overconfident odds in the market.

The result is that, in low-information events like a Michigan primary, prediction markets are prone to overconfidence. 

“You might want to invest in the underdogs instead," Silver concludes.

The future of forecasting

Scott Alexander assigns a 70% chance of AIs becoming “ultraforecasters” — models that exceed the accuracy of human superforecasters by between four and 12 percentage points. 

His lower bound is based on his assumption that AIs will improve on human superforecasters by as much as prediction markets once improved on statistical models (which was four percentage points).

Four percentage points may not sound like a lot, but it is. Alexander says it’s equal to half of the benefit of just knowing what you’re betting on.

“The effect of going from a statistical model to a prediction market,” he explains, “is half as large as the effect of knowing which two teams were playing and how good they are!” 

He argues that this is the minimum improvement we can expect from AI because sporting events are optimized to be unpredictable (through rules, salary caps, drafts) and sports betting is therefore the most difficult area to improve on human forecasting.

The upper bound of his range is based on chess, because chess may be the domain where AIs have the greatest advantage over humans.

AI chess engines are now so far ahead of humans that they can beat a grandmaster despite starting three pawns down.

That sounds like a lot. 

Alexander then calculates that if an AI forecaster pulled ahead of human superforecasters by a comparable margin, it would be the equivalent of adding 12 percentage points of accuracy to their predictions.

In practical terms, this means that Alexander expects that what looks like a 50% probability on prediction markets now will be something between a 54% and 62% probability in the near future.

“All of this keeps me excited about AI superforecasters,” he concludes, “even though I don’t expect miracles.”

Betting against CLARITY

Jeff Dorman, CIO at the crypto investment firm Arca, estimates the CLARITY Act has less than a 10% chance of becoming law by the end of the year — far below the 60% chance implied by Kalshi.

If so, Arca's $10 million over-the-counter prediction-market bet from June is looking pretty good.

The bet was facilitated by Galaxy as part of its new over-the-counter betting service for institutions who want to bet bigger than is possible with public prediction markets.

Dorman describes the transaction as a small hedge against Arca’s book of crypto longs, which will presumably suffer if CLARITY fails to pass.

But I’d also describe it as an alternative way for Arca to monetize its many years of paying attention to crypto.

It should be a welcome one: Selling a 10% probability for 60 cents seems like a far better risk-reward than anything that crypto has offered investors over the last few years.

Crypto is unpopular

Semafor reports that Senate Democrats have been circulating previously unreported polling on how their primary voters feel about the crypto industry.

It’s even worse than you probably think.

The poll (now posted online) found that voters in Democratic primaries view crypto less favorably than a who’s who of Democrats’ favorite villains: Big Pharma, data centers, and oil companies. Even Wall Street — crypto's own favorite villain — is viewed more favorably.

(Apparently, the enemy of an enemy isn’t always a friend.)

Only ICE and the NRA are more unpopular than crypto with this group.

When told only that a candidate is supported by the crypto industry, 84% of respondents said they held a negative opinion of them. No further information about the candidate was necessary.

Senate Democrats are circulating the poll now because the crypto industry needs seven Democratic votes to overcome a likely filibuster of the CLARITY Act.

They might want to start with senators that don’t have a primary coming up.

The FBI has some seed phrases

FBI agent Patrick Yaroch has confessed to stealing roughly $1 million in cryptocurrency by memorizing seed phrases he found among intelligence the FBI had collected investigating "Adversarial Nation 1."

How the FBI obtained the seed phrases is sadly not explained.

Yaroch used the seed phrases to access the wallets on his personal devices, transferring the funds into wallets of his own.

Yaroch is not one of the FBI’s crypto experts. He told investigators he researched crypto to learn how to use seed phrases to access the adversarial wallets — and then sent the funds to a Kraken wallet linked to an account in his own name. 

His explanation: He “became frustrated when he could not do more to disrupt this individual’s use of cryptocurrency.”

I get it. How many bad guys are using crypto wallets that the FBI has the seed phrases for???

Yaroch deposited most of the $1 million on the niche DeFi protocol SuiLend. According to the complaint, “Yaroch stated he chose this service simply because he liked that the logo was a water droplet.”

That’s also where the money stayed. He did not transfer or spend any of the crypto, or attempt to conceal it.

Soon consumed by guilt, Yaroch turned himself in to his colleagues at the FBI. He has lost his job and faces charges for the interstate transportation of stolen goods.

(For legal purposes, crypto that’s moved from one wallet to another is considered to have crossed state lines.)

Did it really have to end this way? It wasn’t even the FBI’s money. Just send it back! Pretend nothing happened. I’m confident no one would have noticed.

(Other than the wallet owner in Adversarial Nation 1.)

The trillion-dollar expense line

The authors of a recent paper analyzed transcripts from 404,940 quarterly earnings calls using computational linguistics to estimate the annual cost of cyber risk to corporations — including the cost of additional spending on IT, lost sales due to systems downtime and reputational risk, legal expenses, fines, and insurance.

Their estimate? $1.14 trillion.

Annually.

For perspective, McKinsey estimates that global corporations make an economic profit — the profit earned above a company’s cost of capital — of $1.2 trillion a year.

If the cost of defending against cyber attacks was equal to the world’s economic profit in 2025, what will it be in 2026? More than that, presumably, given the advent of generative AI.

And what about 2027? 2028? Does the estimate go up with every new model release?

If so, it stands to reason that the stock market should go down with every new model release.

Scary.

— Byron Gilliam