Astral Codex Ten

superforecasting

18 min read

superforecasting

Article

superforecasting is a recurring concept in the Astral Codex Ten archive, appearing 2 times across 2 issues between December 20, 2022 and September 17, 2024. The archive places it in contexts such as “Aren’t prediction markets worse than superforecasting? “Superforecasting” refers to a variety of forecasting methods”; ““Superforecasting” refers to a variety of forecasting methods”; “I talk more about superforecasting and probability calibration than Tyler”. It most often appears alongside Kalshi, Manifold, Metaculus.

Metadata

  • Category: Concepts
  • Mention count: 2
  • Issue count: 2
  • First seen: December 20, 2022
  • Last seen: September 17, 2024

Appears In

Source Context

Recovered passages from the original issue text. When the raw archive preserved outbound links inside the source passage, they are listed directly under the quote.

Prediction Market FAQ
December 20, 2022 · Original source
After the event happens, use the outcome to update everyone’s reputation and refine the algorithm. Superforecasting uses some of the same ideas as prediction markets - probabilistic forecasts, incentives to get the right answer, aggregation methods that favor people with good track records. In studies comparing superforecasting tournaments to small prediction markets, the superforecasting tournaments have done equally well or even slightly better. My goal with this FAQ is not to claim that prediction markets are always better than superforecasting. I think of both as part of the same revolution in forecasting technology, and would be happy with policy-makers or other important people using either. Still, I do think that each has situations where they might be a better fit than the other. Superforecasting tournaments shine on questions so far in the future that financial incentives start to lose force (for example, people are unlikely to place bets on questions about 2100, when most of them will be dead anyway). They’re also good in situations where you can’t get a big prediction market together - superforecasting scales down more gracefully, since you can identify individuals as superforecasters and consult them even in situations where you can’t get a full tournament together. Prediction markets shine in avoiding advanced manipulation attempts, in providing a single canonical answer when someone might worry that any given tournament was biased, and in aggregating the results of superforecaster tournaments with each other and with other sources. Remember that a superforecasting tournament can be considered an “expert”, like Nate Silver. So by the argument in Part 2, we should expect that a big prediction market won’t consistently be worse than any given superforecasting tournament, as long as the tournament’s answers are public knowledge. If there were ever a superforecasting tournament that consistently outperformed prediction markets, that would be a simple mispricing, people would correct it, and the market would eventually agree with the tournament. 4.5: Aren’t prediction markets gambling? Isn’t gambling bad and addictive? Yes, sort of. But most countries allow forms of gambling that aren’t too addictive and have some social value. For example, investing in stocks, or investing in commodities futures. I think prediction markets are more like this than like traditional gambling in casinos. People who want to gamble can already buy cryptocurrencies, or trade stocks on Robin Hood, or (in 20 states) place online sports bets on sites like DraftKings. All these things seem more addictive than, and have less social utility than, prediction markets. I don’t think promoting or legalizing prediction markets is going to make the gambling situation much worse than it is already - so given how useful I think they are, I think they would be net positive. People who are more concerned about the gambling aspect might want to stick to play money prediction markets, which wouldn’t have this problem. 4.6: Where does the money in prediction markets come from? That is, if "you get a dollar when the Democrats win”, who provides the dollar? In the abstract, prediction markets pair up people who want to bet on different sides of a proposition. For example, if a market says that there’s a 75% chance that the Democrats win, then they pair up someone willing to buy a share in “The Democrats win” for $0.75 with someone willing to buy a share in “The Democrats lose” for $0.25, for a total of $1 spent on these two shares. Then, when the Democrats either win or lose, the person with the correct share gets the $1. In practice it’s annoying to have to wait for someone to take the opposite side of the trade, so some people (or bots!) play “market maker” and are willing to take your bet on the assumption that someone else will come along soon to take the other side. But it’s usually safe to abstract this step away and just imagine people betting with each other, using the market as an intermediary. 4.6.1: Then why should anyone play prediction markets, when on average they’ll only break even? It seems like this is a worse deal than stocks, which tend to go up over time. Every dollar someone wins on a prediction market corresponds to someone else’s loss; in expectation; across all participants, the average gain is 0. But the stock market tends to go up over time, as businesses expand to new areas and invent new products; across all participants, the average gain is about 4% per year. So why ever invest in prediction markets instead of stocks? Whatever the theoretical answer to this question, lots of people do invest in prediction markets instead of stocks sometimes; several existing prediction markets have questions with hundreds of thousands of dollars in trading volume. You would have to ask those people why they do it. Maybe it’s because it’s fun. Or maybe it’s because they think (rightly or wrongly) that they’re above average and can make a profit. This is no different than other zero-sum games like sports betting, which attracts billions of dollars each year. The futures and commodities markets are also zero-sum, but attract billions of dollars by giving companies an opportunity to hedge risk. For example, a nickel mine might get rich if the price of nickel goes up, but go bankrupt if the price of nickel goes down. And they might prefer a predictable world where they get a small but guaranteed profit no matter what happens to nickel prices. So they bet some amount of money on commodity markets that the price of nickel will go down, and then their income is the sum of what they make from their nickel mining and from their bets - which, if they handled their hedging correctly, should be a small but guaranteed profit. Prediction markets would allow hedging of other types of risk - for example, import-export businesses might want to hedge against the risk that a protectionist politician gets elected, or tourism companies might want to hedge against a pandemic that closes international borders. These people would inject enough money into the market to subsidize sophisticated speculators. Finally, I envision that someday people who want to know the answer to specific questions can subsidize prediction markets on them. For example, the Democratic Party might subsidize a conditional market (see 5.1) about which Democratic primary candidate is most likely to win the general election. Their money would go to giving the average investor a 4% (or some other number) rate of return - although of course winners would gain more than that and losers would still lose on net. I think this is the most likely way for prediction markets to become very big. 4.6.1.1: If people use prediction markets to hedge risk, won’t that distort them? That is, suppose that an import-export business spends millions of dollars betting that Trump will win in order to hedge against his protectionist policies. Since their bets aren’t based on the real chance of Trump winning, won’t that distort the market? No. Suppose that everyone knows Trump has a 50-50 chance of winning. And suppose the import-export business, in the process of hedging risk, bids it up to 90-10. Since you know Trump has a 50-50 chance of winning, you can get rich quick by bidding it back down to 50-50. From your point of view, the import-export business is (in expectation) giving you free money. But they’re still happy to do it, because they’re hedging their risk successfully. 4.7: Aren’t a lot of the questions we care about inherently subjective or hard to measure? This is a frequent problem for prediction markets. For example, we might want to know something like “will we get human-level AI before 2050?” But how do we define “human-level AI”? If there’s an AI that’s much better than humans at most tasks, but much worse at a few, is that “human-level”? If there’s an AI that seems human-level in demos, but the team that makes it won’t let it be independently tested, should that count? If it works through some kind of Frankenstein chip that combines vat-grown brain tissue with computing machinery, is that still an “AI”? Prediction markets have found a few ways around this problem. First, many groups (for example, Metaculus) try to define their resolution criteria very carefully. A typical Metaculus question on AI sounds like this: We will thus define "an AI system" as a single unified software system that can satisfy the following criteria, all completable by at least some humans. Able to reliably pass a 2-hour, adversarial Turing test during which the participants can send text, images, and audio files (as is done in ordinary text messaging applications) during the course of their conversation. An 'adversarial' Turing test is one in which the human judges are instructed to ask interesting and difficult questions, designed to advantage human participants, and to successfully unmask the computer as an impostor. A single demonstration of an AI passing such a Turing test, or one that is sufficiently similar, will be sufficient for this condition, so long as the test is well-designed to the estimation of Metaculus Admins.
Operate using play-money only. Here Manifold is the leader. You could also think of superforecasting tournaments like Metaculus as a version of this.
Inline links: Manifold, Metaculus
If we try this plan, then looking back on it ten years from now, will we agree it was a mistake? Prediction markets give us a way to get accurate and canonical answers to questions like these, and to short circuit the usual discussions about how biased different information sources are. See below for some clever, more exotic ways we can use prediction markets. 4. What are the most common objections to prediction markets? These are various objections, some wrongheaded, some true but nonfatal. There are many of them, making this section very long - you might want to skip over any objections you’re not worried about. 4.1: Would prediction markets be ruined by insider trading? That is, suppose there is a market on whether President Biden will resign before the end of his term. President Biden has special knowledge of this, so he could bet on the true outcome and make a lot of money unfairly. He could even change his behavior (eg resign at an unexpected time) just to make more money. Isn’t this unfair? One answer is that normal markets (eg the stock market) face these same problems, but manage them by making insider trading illegal. These laws don’t always work perfectly, but they work well enough that most people are happy to buy stocks. Another answer is that, while this is bad for other investors, it’s not bad for the accuracy of prediction markets, or their use in creating unbiased social consensuses. In fact, knowing that President Biden is insider-trading on a “Will President Biden resign?” prediction market should only increase your confidence in it getting the right answer! This is slightly too rosy, because if insider trading is bad enough for other investors, they might just not trade. This would be a partial effect: investors would be willing to overcome their fear for a big enough payday, meaning that concerns about insider trading probably would increase the likelihood of persistent small mispricings while still not allowing bigger ones (with the exact size depending on how frequent the insider trading was). It’s unclear whether this negative effect would be bigger or smaller than the positive effect from insiders having more information, so in different situations the market might end up either more or less accurate. Overall, economists are split on whether insider trading makes markets more or less accurate. Commodities markets don’t really have insider trading laws right now, and seem to be about as accurate as anything else. I hope prediction markets will experiment with different insider trading rules, and the ones that best satisfy all participants and create the most accurate results will win out. If for some reason this doesn’t work, I don’t expect it to make too much difference either way. 4.2: Would prediction markets encourage harmful or illegal activities? What about the risk of insider trading by committing harmful / illegal acts? That is, could President Biden’s doctor decide to poison him, then make money when he has to resign due to ill health? I think the strongest evidence against is that this basically never happens in stock markets. Tesla stock would plummet if Elon Musk died or resigned, but nobody realistically worries that Musk’s doctor will short Tesla and poison him. Lots of corporations’ stocks would sink to zero if you burned down their offices and factories, but nobody shorts them and then commits arson. Probably this is because there are laws against doing harmful and illegal things, and people have decided that stock market gains aren’t worth breaking the law and getting punished. Since prediction markets have only a tiny fraction of the amount of money that stock markets do, probably people won’t consider it worthwhile to commit harmful actions to manipulate them either. If you were going to murder someone to profit off a market, who would you rather kill: a US politician (the PredictIt market on the presidential election has a volume of about $600,000)? Or a Fortune 500 CEO (whose companies might have market caps in the hundreds of billions)? 4.2.1: What about prediction markets in very specific harmful or illegal activities? I guess if you created a market in “Will someone burn down the 7-11 on Main Street tomorrow at 3:32 AM?”, then bet a lot of money, then did it, that would be bad. I think realistically nobody would bet against you on that. But probably prediction markets should avoid hosting markets on these very specific bad things, just to make sure. 4.3: Would prediction markets give rich people more power? That is, suppose we used prediction markets to assess socially important questions like “will the climate change by such-and-such a number of degrees by 2030?” It would be bad if rich people could manipulate our social consensus on this. But you move prediction markets by buying shares, and rich people can afford more shares than poor people. So doesn’t this mean that rich people can manipulate how concerned we are by global warming? No. See 3.2 for the general reasons why it’s very difficult or impossible to successfully manipulate a prediction market. These reasons apply to rich people too. Suppose a rich person spent $100 million to buy NO shares in “will the climate be warmer in 2030 than today?”, pushing the market’s implicit chance of global warming down to 1%. That means if there is global warming, you could multiply your money by 100x by buying YES. I would immediately invest $10,000 in this market, so that I could get $1 million back in 2030 and retire rich. My $10,000 isn’t going to be enough to fully move this market all the way back - we already said the rich person spent $100 million manipulating it. But “you can get a free $1 million quickly with no downside at an evil rich person’s expense by correcting an obvious misconception about global warming” sounds like the sort of thing that could make it to the front page of Reddit (to put it lightly). I think more than enough people would learn about this to fully correct the mispricing. Is there any amount of money that could successfully manipulate a market? I think the answer is that you need to have more money than the sum total owned by everybody else in the world who wants to make $1 million quick. And at the limit, there’s always Goldman Sachs - who watch financial markets very closely, definitely want to make $1 million quick, and have a lot of money. So I think the most honest answer to this objection is: if you are an evil rich person reading this FAQ, then it will definitely work for you. Please sink $100 million into reducing a prediction market’s chance of global warming to 1%. And make sure you tell me first, so that I can fully marvel at your evil genius. This will work great for you and nothing will possibly go wrong. 4.3.1: But wouldn’t the subtle biases of rich people (which they might genuinely believe) still affect the market more, since they have more money? No. See 3.3 for the general reasons why we should expect prediction markets to be free from subtle biases which people genuinely believe. These reasons apply to rich people too. Suppose rich people have subtle biases which make them wrong more often than poor people. And suppose rich people (wrongly) believe global warming is 75% likely, but poor people (correctly) believe it’s 99% likely. This just reduces to the Nate Silver situation earlier, with poor people playing Nate Silver. The aggregated opinion of poor people is “an expert” which is right more often than the markets. It’s easy for someone to notice this and get rich quick (in expectation) by betting on what poor people think. Since lots of people can easily notice this and want to get rich quick, eventually they will correct the mispricing. Even if rich people have so much more money than poor people that no group of poor people, however large, can ever correct a rich person mispricing, eventually some smart rich person will hit upon this strategy themselves. If no individual rich person does it, Goldman Sachs will definitely do it. 4.3.1.1: What if both rich people and poor people have biases, and neither one is consistently more right than the other? Won’t the market still reflect rich people’s biases rather than poor people’s? Not if it’s possible for anybody to notice these biases and correct for them. Treating the aggregate opinion of poor people as an expert was just one example. If the winning strategy is something like “trust rich people on financial questions, poor people on environmental questions, and the point exactly halfway between them on social questions”, then whoever discovers that strategy can get rich quick. The more often people use prediction markets, the easier it should be to detect strategies like these. 4.4: Aren’t prediction markets worse than superforecasting? “Superforecasting” refers to a variety of forecasting methods similar to those pioneered by Philip Tetlock and the Good Judgment Project. Typically, they would do something like: Ask many smart people to give probabilistic answers to a very well-specified question
Inline links: economists are split
24
September 17, 2024 · Original source
2: Dean Ball has a sort of vague vision of LLMs betting on prediction markets at massive scale. I agree something like this is interesting and plausible; I agree that it’s hard to pin down exactly how it would work. One suggestion he makes is to have the bots shadow public intellectuals - for example, a bot “trained on” my writing would ask itself “how would Scott Alexander bet in this market?”, and if it made more money than a bot asking “how would Tyler Cowen bet in this market?”, then maybe you would trust me more than Tyler. This is cute but there are a lot of wrinkles to work out For example, I talk more about superforecasting and probability calibration than Tyler, my bot might simulate me by making good bets; if Tyler sometimes uses extreme or ideological language, his bot might make worse bets not because his ideas are worse, but because it “simulates” him as being an incautious better.
Inline links: a sort of vague vision of LLMs betting on prediction markets at massive scale