Prediction Markets and Worse is Better
If you like reading bad news, it's a great time to be alive. Not that the world is in a terrible state; most of the long-term graphs look pretty good, actually! But bad news has always outsold good, and so there's relentless selection to make news stories sound worse than they really are, ideally in a topical way.1 If you see a story about the upcoming collapse of a government, spiralling inflation, chaos in the Middle East, etc., there's a very easy trick to filter for noise: look at asset prices. Stocks will react one way to "the end of democracy in America" and another way entirely to a histrionic op-ed talking about "the end of democracy in America." Oil prices will shrug off irresponsible speculation about war with Iran, but will move very fast if it's actually going to happen. Armen Alchian figured out one of the ingredients in hydrogen bombs by looking at stock prices.
(This isn't a perfect indicator, of course—it would have made you late to take Covid seriously, for example. But if it's an actively bad one, you can quickly amass a fortune by scrolling through Twitter and reacting to all the news you see by trading futures.)
In one sense, this simplifies the world a bit. You can spend less time following some stories, simply because you know they're not really going to matter. But it's irritatingly imperfect. Yes, the value of the euro will react if Italy seriously attempts to leave the EU. But the euro reacts to plenty of other things, too. And while you can use trading as a yes/no signal on the importance of assorted headlines, you generally can't use asset prices as an indicator of odds, except in special cases.
Since markets are a good way to quickly aggregate information, but a messy way to aggregate specific information, many smart people have proposed prediction markets as a way to get better estimates of the odds of geopolitical events. You could bet on oil as a proxy for the odds of a coup and wars in oil-producing states, but you could also directly bet on the odds themselves.
This is a very elegant solution, and in some ways it's been implemented. There are crypto prediction markets, real-money prediction markets like PredictIt, and, of course, countless places for betting on the outcomes of sporting events. After 9/11, there was a very interesting effort to create a prediction market specifically for betting on terrorist attacks and other political events. The goal was to use prices as a guide to which threats were credible—possibly because informed analysts would make big bets when they were confident, and possibly because people planning the attacks would accidentally leak information or even, hypothetically, trade themselves. It would be very weird indeed to build a market that let terrorists earn a return on their activities, but the assumption was that the market's information content would offset whatever incremental impact it had on behavior.
But, aside from the bad PR (which shut down the Policy Analysis Market soon after it was announced), there's another reason prediction markets are a challenging way to bet on real-world events: they're too good. A market requires someone to take each side of a trade, so for every person betting that a given event won't happen, there's someone betting that it will.
Consider the strategy of a market-maker who trades the "will there be a successful coup in Venezuela in year X?" market. That trader might decide that the base rate for coups in a country like that is 10%, and could quote odds of 5% for people who wanted to bet against a coup, and 15% for people who wanted to bet on it. Under normal circumstances, the price would drift along according to whatever was in the news—down a bit if inflation slows down, up a bit when a popular opposition leader gets arrested, etc. Now, suppose someone starts bidding on the 2022 coup contract, and bidding a lot. They're willing to buy at 15% odds, at 20% odds, and 25% odds. The other contracts don't move.
There are two possibilities:
- The buyer has a random fixation in which they're overconfident.
- The buyer has specific knowledge of an imminent plot, and knows the date on which it will happen.
It's a classic case of adverse selection, and prediction markets are designed so that their biggest value creation happens through adverse selection against market-makers. The more specific the contracts are, the tougher this problem gets; if the market-maker is trading monthly rather than annual coup odds, and sees that there's lots of interest in the March 2022 contract and none whatsoever in the other months', that's an even stronger signal.
Our market-maker is savvy, though, and knows how to recognize a change in market regime. Now the bet is not against a random adversary, but an informed one, and pricing is different: informed people may know something you don't, but they can also be overconfident. Now, the relevant base rate is: what percentage of scheduled coups get either rescheduled or cancelled before they happen? (The best example of this is how Gaddafi took power. The Prize talks about the time a Libyan military officer was awakened in the middle of the night by one of his soldiers, and told the soldier that the coup wasn't scheduled to begin for a few days. They were involved in two different coups.) Say the odds of a cancelled or rescheduled coup are 10%. So our market-maker now requires a bettor to accept 90% odds that the event will happen. Suddenly, a 7x return from betting on a rare event has turned into an 11% return from betting against a rare event. Our well-informed but frustrated trader may, at this point, give up.
Or the trader might look at other assets that would react to a coup, and trade those. Defense contractor stocks would probably go up at least modestly, and they're quite liquid; oil, treasuries, and currencies would be affected, at least a little.
And in those markets, the adverse selection market-makers worry about is generally not that someone has fundamental information that will make their trade better; the market-maker is buying and selling rapidly, and their positions at any moment today have almost nothing to do with what they'll own tomorrow or in a week. Market-makers only have to worry about information asymmetry about what a trader's next trade will be; they don't want to be run over from taking the first slice of a tiny trade. They don't want to be like the funds that bought ViacomCBS at a discount to its last price, just before it lost more than half its value ($, WSJ). A hypothetical well-informed geopolitical trader betting on major markets may have a lot of money in an objective sense, but their trades are basically meaningless at the scale of FX ($6.6tr traded daily), treasurys ($0.5tr) or oil (~$50bn on the CME).2
A persistent puzzle in finance is the question of why people trade as much as they do. Most people don't reevaluate the estimated net present value of a stock on a regular basis, and most of us aren't constantly rethinking our expected future consumption baskets, loss tolerance, and other risk determinants. You can just call frequent trading irrational and leave it at that, but it's a deeply dissatisfying answer. The presence of "noise traders," whose behavior can be modeled as practically random, is a fortunate feature of markets; one way to think of them is that they pay a continuous fee to market-makers, in exchange for which market-makers allow informed traders to sometimes make money at their expense. This ecosystem works fine with stocks, because there are so many reasons to trade them, or with treasuries and FX, because there are so many parties who have to trade them. But it doesn't apply very well to prediction markets, especially very narrow (and thus very useful) ones. Noise traders are the cure for adverse selection, or, more properly, a treatment: the more of them there are, the more lucrative it is to provide liquidity, and thus the more money an informed trader can make.
Long Bets is an example of a prediction market that avoids adverse selection. You can have an information advantage on Long Bets—if someone at OpenAI was betting that an AI would write a platinum-selling album in the next year, I'd be reluctant to take the other side. But the site generally aggregates differences of informed opinion, and over timescales so long that the informational edge from secret knowledge is far smaller than the informational edge from being good at predicting the future. It's a small market, though, and the bets get donated to charity (this bet, for example, will pay out the same amount to the same charity no matter who turns out to be right). But the function of Longbets is generally not to aggregate information in the form of prices; it's to aggregate information in the form of the arguments each side makes; the bets themselves are a fun feature, but they're not fundamental to the service.
Asset prices are an incredibly efficient way to aggregate collective information, but the mechanics of markets depend on lots of traders with varying motivations. The breadth of opinions that can be represented by a single trade is a bug from the perspective of someone who wants to isolate just one aspect of the world, but a feature for anyone whose top priority is to use prices to stay informed. Prices are a noisy signal, but high-bandwidth; prediction markets are much less noisy because they're so specific, but that specificity makes it hard for them to absorb the kind of volume necessary to attract the most informed traders. We want prediction markets so we can measure the odds of low-risk events, but those are exactly the events for which these markets are illiquid and one-sided. Fortunately, liquid markets offer a very convenient alternative to prediction markets: every time prices instantly react to a news event, it creates a retroactive prediction market telling you that whatever just happened was a very big deal. And if you happen to know of an important event, or just have an edge in predicting one, then you'll get enough liquidity to profit from that, while pushing prices in a more informative direction.
A Word From Our Sponsors
Here's a dirty secret: part of equity research consists of being one of the world's best-paid data-entry professionals. It's a pain—and a rite of passage—to build a financial model by painstakingly transcribing information from 10-Qs, 10-Ks, presentations, and transcripts. Or, at least, it was: Daloopa uses machine learning and human validation to automatically parse financial statements and other disclosures, creating a continuously-updated, detailed, and accurate model.
If you've ever fired up Excel at 8pm and realized you'll be doing ctrl-c alt-tab alt-e-es-v until well past midnight, you owe it to yourself to check this out.
De-Dollarization on Paper
Russia's sovereign wealth fund is cutting its dollar holdings from 35% to 0% ($, WSJ). This is one of those stories that will be shared as evidence for the decline of the dollar as a reserve currency, but:
- The immediate effect is just shuffling ownership: "the shift will take place as an internal transaction between the government and the Russian Central Bank. The bank holds around a fifth of its assets in dollars and analysts don’t expect that share to fall quickly."
- Russia has unusually poor relations with the US, and has been hit with sanctions in recent years.
- This doesn't involve de-dollarizing Russian exports, just a subset of state savings.
If a country that is, off and on, ranked as the US's number-one geopolitical foe stops using dollars in a cosmetic way that doesn't affect the real world, and only after years of friction, it's evidence that foreign governments in general are deeply reluctant to even consider alternatives to the dollar. Or, to put it in crypto argot: Actually, this is good for USD.
The Amazon Bundle
Amazon's Prime Day will start with a concert featuring a few popular musicians. They've done this a few times before, and Alibaba has used this technique to great effect (both companies, for example, have had Taylor Swift as the headline act for their biggest annual sale). At one level, it's just an attention-maximizing trick; stars attract a big spotlight, and having some of that spotlight fall on Amazon is obviously helpful for their business. Since the show is viewable on Prime, it's partly a way to push a handful of marginal users from thinking Prime isn't quite worth it to deciding that it is. But mostly, it shows off how effective Amazon's bundle of miscellany is: with a sufficiently diverse set of products and streaming media services, there's nothing that isn't relevant to the bundle, so adding a live show from popular artists is less unnatural than it would be if, say, Bed Bath & Beyond offered a livestreamed music event only to members of their loyalty program. Since Amazon is partly a media company, it can engage in an effective arbitrage: media assets can be monetized through subscriptions, ads, or both, and Amazon is the only advertiser on Amazon-owned media that doesn't have some kind of information asymmetry.
Disclosure: Long Amazon
The Extension Market
I've written before a few times about browser extensions: they're fairly easy to build, but often hard to monetize, so extension owners will sometimes use them for gray-hat VPN products that mimic non-automated traffic. A frequent pattern with Google is that it builds a distribution tool that gives access to a vast market, the tool gets gamed aggressively, and Google switches to shutting down the most aggressive behaviors. This happened with search engine optimization, when Google routinely updated its algorithm to penalize the most scalable ways to improve sites' rankings. It also happened in email, with the "Promotions" tab exiling more commercial emails from users' inboxes. Google is now warning users if Chrome extensions are untrustworthy. Chrome extensions give developers access to a massive market—2.65bn global users. And since so many products are browser-based, extensions give developers access to a surprising amount of users' data. It makes sense for Google to set rules early and focus on whitelists rather than blacklists; as in other domains, the black-hat behavior scales better than white-hat behavior, so spam and malware are the default.
FT profiles Central and Eastern European low-cost carrier Wizz Air ($), which had been growing fast ahead of the pandemic and has survived it so far. One of the dynamics they benefit from is that state airlines used to be heavily subsidized, but were generally subpar operators—if part of the point of an airline is to create jobs, inefficiency is a benefit. They also provide travel, of course, which has some positive externalities. Wizz can drastically undercut them on prices, fill more seats, and get decent economics. And that creates a useful political dynamic for them: if there are cheap flights available already, having a state airline matters less to passengers, so shutting it down only irritates employees. Wizz is in the lucky position that its expansion makes it politically viable to cut off competitors' funding—and when they cut routes in response, Wizz can replace them.
United Airlines has ordered fifteen supersonic planes from Boom Aero, following its decision to order eVTOL aircraft from Archer. While aircraft have been getting better over the last few decades, they haven't been getting that much different; size, fuel efficiency, and safety have improved, but the fundamental products would be recognizable to someone from the 1960s. Part of the reason for this is overlapping capital intensity: planes cost a lot, airports cost a lot, and repairs are much cheaper when fleets are standardized. A few years ago, United embarked on an expansion plan that was controversial with investors (airline investors have nightmares about aggregate capacity growing faster than GDP, and United planned to grow much faster than that). Their plan worked out, pre-Covid. In one sense, the company is reconsidering fundamentals for big carriers, but in another sense, it's paying attention to them: from a first-principles standpoint, it makes sense to buy the most flexible short-range aircraft on the market (eVTOL) and the fastest long-range one (Boom's product isn't, strictly speaking, on the market, but they're making progress on prototypes). Airlines are a surprisingly network effects-driven business: flights beget connecting flights; dense route maps beget loyal customers who make airline mileage operations more profitable. And in a network effects-driven industry, the first company that figures out a new way to expand the network can earn a significant lead.
A general piece of life advice you may never have an reason to care about is: if you insist on committing a crime, be sure to only do one at a time. This is the kind of thing smart protestors will tell each other before a protest—it’s best to make sure that you aren’t carrying anything illicit if you run the risk of getting pickup up by the cops. And I suspect that a sufficiently savvy crime-prevention ML system could identify the experienced drug dealers by their assiduous compliance with traffic laws. In corporate fraud, there are two ways to think about this rule:
- A successful fraud is a boring company that appears to grow faster than its competitors for a long time. This company never attracts attention, so nobody ever bothers to figure out why the numbers look surprisingly good.
- Alternatively, a smart fraudster should do legal-but-morally-dubious things, specifically as a decoy. Given the power of compound interest, fraud eventually draws attention to itself because five- and ten-year charts of companies and their competitors show such sharp divergence. In this model, you want to give investors a plausible explanation for the company's outperformance. (This was something Madoff mastered: he managed to hint that his fund made money by exploiting the order flow from his market-making business. His investors suspected that the numbers were too good to be legitimate, and he gave them a reason to think the illegitimacy was at somebody else’s expense.)
Wirecard, the German payments company that turned out to be partially fraudulent, also made some of its money by handling payments for scam-adjacent binary options trading sites ($, WSJ). This illustrates why frauds can persist for a decade and then suddenly fall apart in a few weeks. There's a sort of magic-eye phenomenon with investor and regulator perception: they had one good explanation for a company's behavior, and suddenly another one was a better fit, and the new image of Wirecard as a fake company, rather than a real company that profited from some morally questionable business decisions, snapped into focus.
I expect future novel infections to be dramatically over-reported in the next few years. Something similar happened after 9/11, with lots of rumors, terrorism scares, buildings getting evacuated, etc. Even an impostor pilot who showed up at JFK two weeks after the attacks. At a normal time, that kind of story generates news-of-the-weird headlines about pretend pilots; after a terrorist attack, it's a big story. (Based on the outcome of the trial, a six-month prison sentence, this probably was a quirky story about an impostor who had phenomenally poor timing.) ↩
One prediction market where there are lots of noise traders is in national elections, where people tend to bet on their favorite politician without thinking too hard about the odds. And since electability is a factor in who gets nominated, the odds tend towards 50/50. These markets are pretty high-volume, and do indeed quickly reflect news. Sports betting markets are another case where there are noise traders, for roughly the same reason; they have the further benefit that athletic leagues want teams to be roughly evenly matched over time so the spectacle remains exciting, and there are biased transactors on both sides. But in both markets, the artificial forces that push odds close to even also make them make the outcomes less meaningful in the real world. Hotelling's Law predicts that political contests will tend to take place between pairs of near-centrists who dramatically overemphasize their differences. And this, with some obvious exceptions, is pretty true. ↩