by the by, about Bolsonaro, I know a few reliable people who were in Brazil for the thick of Covid 19 there(which I am told was overblown in terms of media coverage), and no amount of telling their compatriots to stay home was going to be sufficient. What India did about a similar issue was to have non-compliant bar-goers bludgeoned with sticks by brute squads / Lou Ferrigno.
If Metaculus were a real-money market, you could imagine a world where in 2029 there existed some SMR nuclear reactors, but not in sufficient quantities or in economies where they made sense to deploy at scale for national electric power, but someone floated a few into whatever the smallest, poorest island nation is to supply 1% of its electric power and win the market.
But, you know, I guess that won't be worthwhile even if possible, unless Metaculus leaderboard positions become REAL valuable in the next 8 years.
Maybe we should create a prediction market about the real-world dollar value of a position on Metaculus leaderboard in 2030.
I don't have experience with google's new prediction market. However, when I worked there a decade ago, I did use prophit, which also had "play money with leaderboards" as the reward. It was not quite the right incentive system. Back then at least, leaderboards and play money were reset on a quarterly schedule, and the people at the top of the leaderboard got gift certificates (I think $75 was typical). This means it was best to make somewhat-risky bets with your finite play money and hope they paid out. If there were a market where the outcome were 100% known (but that would only pay out at end of quarter), it would still have been a losing strategy to bet at 70% odds, I'd guess.
On the questions of the form "If this party wins, what will be the value of this metric?", in a one-shot I don't think you learn much except the political prejudices of metaculus users.
But if you did this a really large number of times over a really large number of elections, you could probably learn something. If unemployment is consistently forecast to be 0.5% higher under Party X than Party Y, and unemployment is consistently higher than the Metaculus estimate in years when Party X wins and consistently lower than the Metaculus estimate in years when Party Y wins, then you might be able to figure out exactly how wrong Metaculus users are, subtract it, and figure out the actual effect of Party X vs Party Y on unemployment. (Of course you'd need to wait a very long time to have enough data for this to be useful.)
Regarding "looking for options" -- prediction markets could adopt methods similar to what financial markets use for cleared derivatives transactions. Lets say you want to buy 100 shares of "yes" on some question at a 50/50 chance. You are risking $50, and your total payout will be $100 if correct, in 2100. The daily volatility in the contract may be some fraction of 1%, and a really big price move in the prediction for something that far out might be 2%. The market could simply make you post $2 upfront instead of the entire $50. As the market moves either direction, you get credited or debited 100% of each market move. If you do not have the money to cover the margin call when the market moves against you, the clearing agent liquidates your position. That money reserve could be either posted at the prediction market exchange or take the form of say having a credit card on file.
The obvious benefit is that this method reduces the initial opportunity cost of capital by 96%. The downsides?
(1) People might be worried about getting margin called. Of course they can prevent this completely if they want to buy just keeping $48 in their account in my example, so they are no worse off than the status quo regardless. If they make predictions in many different questions, they likely have lots of diversification benefits and their risk of being proven hugely wrong on all their questions at the same time is quite low. Over the life of their predictions they will need to keep some cash buffer, so the lifetime opportunity cost of capital isn't 96% saved, but could be pretty close anyway.
(2) The market might move by more than the required $2 margin amount, and the people on the losing side might opt not to make good on their payment obligation to cover the difference (i.e. they default). This risk is mitigated when predictors have lots of different positions on different prediction contracts on the same exchange (cross-collateralization). In financial markets, the clearinghouse itself acts as guarantor against defaults, so no predictor would be taking the credit risk of the other predictors. The clearinghouse takes the risk, and charges some sort of fee to fund a pot of money from which to pay for any future credit defaults.
It's not a new problem and a complex market has taken a pretty good shot at an efficient solution. If you have a better solution you can probably make at least hundreds of millions of dollars displacing the CME or LCH from their current gigs.
Say I make a prediction one year out, and then drastically adjust it with 1 month left (closer to what will actually happen), then my Brier score is going to be pretty bad right? But if the majority of votes came in with only 1 month left, being close to the actual outcome, does that mean overall Brier score will be good?
And I think predictions also need a complexity score. So for example predicting 2022 inflation accurately is probably more impressive than predicting the outcome of a sports match (which is a lot more local). And predicting something further out is generally more impressive as well, the more global it is.
A value measure for each prediction could sort of solve this, where each person has a limited number of upvotes they can give to predictions. Where predictions with higher scores count more towards overall prediction score.
Black and Scholes showed in their famous model that (under some assumptions that have been made less restrictive over time) you can replicate the performance of an option using a mixture of cash (or a loan) and holding the underlying asset, then re-balancing the portfolio of the underlying and the cash over time. If these sorts of assumptions can hold here, then this is wrong: "But if nobody is buying shares directly and all the action is in options trading on the side, that won’t work. " People buy options, market making option sellers price those options and hedge their risk using the underlying prediction market, and the market can reflect the beliefs of the people buying the options. The hedging portfolios that the market makers create (a replicating portfolio) brings the information from the option buyer into the market prices of the underlying asset.
Surely the obvious problem with political prediction markets is that money that wanted to make The Other Party look bad would drown out smart money that actually wanted to predict for a return?
In re obesity predictions: Are you betting that the definition of obesity won't change? Or you'll go with whatever the definition is then?
What's the argument for more Americans being obese? The drugs don't work that well? They don't work that well on a lot of people? There are side effects and the drugs are withdrawn? The drugs are expensive and/or otherwise not available? Odds of a black market if they aren't available?
I agree that politicians will probably try to wave away bad looking results and the result of tradeoffs, but isn't that still better than the political discourse we have now?
I guess I don't know the British talking points, so I'm switching to American politics here. Like I can imagine Republican politicians responding to conditional markets saying that Democrats will lead to better test scores by saying that it's only because they'll tax you dry to fund education (cue a discussion of the conditional markets on tax rates). But today I'd imagine the Republican saying that they'll produce higher test scores *even though* the Democrats are spending more money on it because the Democrats will be spending money on teaching white kids to hate themselves or something rather than any useful educational goal.
These prediction markets are worthless because the only predictions that really matter are the non-linear ones.
AI/machine learning can do linear extrapolations just fine.
The 2nd, more fundamental problem with prediction markets is that they are ultimately flawed. If any market shows any promise - it is then trivially skewed by a small amount of money/effort. If y
ou can't trust them early, and you can't trust them once they've proven themselves, that is the literal definition of useless.
1. "The whole point of prediction markets is that their prices correspond to the chance of something happening. But if nobody is buying shares directly and all the action is in options trading on the side, that won’t work..."
It is possible to arbitrage any price discrepancies between options and the underlying stock. For example, if you buy a put and sell a call at the same strike and same expiration date, that is a synthetic short position that can be perfectly hedged by buying the underlying. If the options are already cheaper than the underlying, you can arb it by doing the opposite thing (buy call, sell put, sell underlying)
People will quickly arb any price discrepancies, so there's no risk of the prediction market getting far out of whack that way. The biggest problem I see is that options will be too illiquid. On wall street, companies under $5B don't have liquid options for next month, and even $500B megacorporations have 20% bid-ask spreads on options expiring 2 years in the future. That's even worse than Predictit fees! People can put limit orders in the middle and wait a long time for someone to fill them, and hope they don't get past-posted when news comes out, but that experience kinda sucks.
2. Caution on conditional prediction markets.
Betonline has some hilarious alien-themed markets (https://www.betonline.ag/sportsbook/futures-and-props/entertainment/aliens) for who will get abducted by aliens first, which country will get attacked by aliens first, and who will win the alien v human war. All bets are void if the aliens don't show up by 12/31/2021. So they're basically getting free loans from whoever bets these things. People have nothing to lose except interest, so they don't care what the lines are and might bet it just to generate reward points or something.
The more improbable the condition, the less skin in the game people have. If the condition is <1%, then probably nobody really wants to bet it except for ideological reasons or silly reasons that have nothing to do with the correct conditional probability.
If you want a prediction market for something that will happen (or not) in 2100, then that market needs to trade not in 202x dollars but in safe bonds maturing at 2100. This is important both because of the time value of money (since a 100% certain win of $100 at 2100 is likely worth just a few dollars now, so you should not have to put up $100 so that the other side would expect to get $100 at the end of the bet) and because of the need for the market to decouple the fluctuations of the actual bet from the fluctuations on the expected interest/discount rate for that decades-long bond, which - at least when the payoff is still distant - would be much more influential than the actual expected value of the bet.
How do they define 'Nation'? Per the question, they have 'Resolution may come from credible media reports, government agencies, or energy industry researchers such as the IEA.' as a resolution criteria, but that isn't clear on who exactly counts. Does Israel count? They still aren't fully recognized. Kosovo? Less recognized, but recognized by some big players. Taiwan? Barely recognized. Bougainville? Aiming for independence but arguably not there yet. Scotland? In no way independent but may or may not want to be, also still labelled a kingdom.
Can I declare independence as dictator of the Russian city Metaculus links as having a Small Nuclear Reactor already, have an article written about me as a crackpot trying to secede from Russia, and win?
I think the closest correlate to you option market conundrum is VIX futures. Since there is no such thing as trade able spot vix, the curve exhibits some interesting behavior.
I don’t know if the following helps re options since I didn’t read the previous post in depth but here’s something:
Options can be arbitraged with their underlying asset in the following way: if you buy a call and sell a put at the same strike , that’s equivalent to buying the stock for the premiums plus the strike since you will either exercise the call or be forced to buy by the holder of the put depending on where them price of the underlying ends up.
Wegovy. The drug was approved by the FDA on June 4. It is very expensive (>$300/week). After a failed attempt to obtain insurance coverage, I began a course of treatment in September. The course of treatment described by the package insert is to begin at 0.25 mg and escalate by steps to the therapeutic dose of 2.4 mg, which will be my next step. I have had none of the side effects described in the warnings. By the time I completed the third step, I had already lost 7% of my body weight. When I started my BMI was almost 40. My goal is to get that down below 30, at which point i will be merely overweight. Right now it seems to be very doable.
The biggest problem with mass use of Wegovy are 1. cost/insurance, which in the long run will be solved. Every one of the five other prescriptions I take was once patented and expensive. They are now all off patent (one is now OTC) and are a lot cheaper. And 2. the supply chain (sigh):
"Novo Nordisk (NOVOb.CO) shares fell as much as 16% on Monday after the Danish drugmaker was hit by U.S. supply issues for its new obesity drug as it seeks to establish a market foothold before the launch of a rival drug by Eli Lilly.
"The obesity market has proved difficult for pharmaceuticals companies, but Novo made a breakthrough with its Wegovy drug. The company hopes the drug, which helps to achieve weight loss of 17% on average over almost two years, will offset growing pressure on its core insulin business. read more
"In a stock announcement after Friday's market close, Novo said that a contract manufacturer filling syringes for pens to inject the drug had halted deliveries and manufacturing temporarily after issues relating to good manufacturing practice.
"The Danish drugmaker was already facing separate supply constraints after being "overwhelmed" by initial uptake of the drug. read more
"As a consequence, fewer new patients will start treatment with the drug in the first half of next year. Novo now expects to be able to meet U.S. demand in the second half of next year, having previously estimated it would do so at the beginning of the year."
"I don’t know, maybe you wouldn’t learn anything there either. If Conservatives had better GDP growth, maybe leftists would say “of course, they’re the more capitalist party, they trade off environmental damage and inequality for a slightly hotter economy”. If Labour had better test scores, maybe rightists would say “of course, they’re the more socialist party, they’ll tax you dry and throw some of the money at schools but it will still be inefficient.”"
I think this is pessimistic. Yes, of course this is the kind of thing each side will say, but isn't that how it should be? Imagine a world where there are established prediction markets for public policies like this with strong track-records of accuracy. Then the public debate could be about what what values people want to support: crime reduction vs. mass incarceration; economy vs. environment, etc. rather than the byzantine factual/counterfactual arguments about which side's policies would *really* be better for crime reduction or economy that we have today. Also it would encourage parties to drop policies that suck at a achieving their intended/stated ends.
> Here I admit I don’t know much about markets or options - is there some way to combine regular trading and options trading into a single price, so that we could get the advantages of options trading, but traders would still be betting on the regular market and increasing our predictive accuracy?
Yes, this would happen naturally and automatically via arbitrage. See:
> That peak at the end is weird-looking and shows up on many Metaculus questions. I think this is something like all dates after April 1 getting crammed into April 1 for the graph, and even though very few people think it will be after April 1, that’s still a really high number of guesses when it’s all concentrated into an individual day.
I assumed this was happening because users really put their prediction on the last day of the time window to signal “I don’t believe this will happen within the time window of this question”. Many questions include “If this does not occur by [last selectable date], it resolves as > [last selectable date]” in the fine print.
Did Robin Hanson really invent prediction markets? On the Wikipedia article on prediction markets, I see that he made the first known corporate prediction market in 1990, but it also says that the University of Iowa's market first ran in the 1988 presidential election. Did he make one earlier?
> Suppose you want to predict something in 2100. It’s hard. Nobody wants to lock their money up for 80 years to get a 5% or 10% or even a 100% rate of return.
A partial solution to this problem:
Prediction markets presumably provide a little bit of diversification, so the mean-variance-optimal portfolio contains at least a small allocation to prediction markets, regardless of how low the expected return is (as long as it's still positive). Investors can add leverage until they reach the desired level of risk.
This solution is not complete because it only works if you can borrow money at the risk-free rate. Pretty much nobody can do that. Even if you're only paying an extra 0.5% for leverage (which is something like what institutional investors pay), that could be enough to wipe out your expected return. So long-dated predictions are still worth buying *if* you can get cheap enough leverage. The question then becomes how to get sufficiently cheap leverage.
"My next question is: is there a structure where options directly move the market? The whole point of prediction markets is that their prices correspond to the chance of something happening. But if nobody is buying shares directly and all the action is in options trading on the side, that won’t work. Here I admit I don’t know much about markets or options - is there some way to combine regular trading and options trading into a single price, so that we could get the advantages of options trading, but traders would still be betting on the regular market and increasing our predictive accuracy?"
Options move the stock market through arbitrage and the hedging of market makers (they don't want to make or lose money on the outcome of the options they buy/sell, so they buy/sell the underlying security to offset price moves and stay "delta neutral.") Presumably something similar could work for a prediction market that had a liquid market supported by market making firms.
On the subject of options effects on security prices when they dominate the flows of their underlying, investory/day trader types might be intersted in googling about NOPE, the Net Options Pricing Effect, a calculation invented in 2019 by Lily Francus and some cohorts that tries to guess when stock prices (specifically SPX index or SPY ETF given their liquidity) will have a near term reversion due to market maker hedging - if too many calls or puts are bought relative to the actual stock, they eventually have too much built up pressure and risk and have to engage in hedging purchases/sales (usually through /ES futures, which are abritraged to effect the SPX/SPY prices) that will cause a short term "snap back" in the underlying price until the put/call flow becomes more balanced.
It's had some useful backtesting as well as identifying conditions where it doesn't work well (actual market moving news is going to run over relatively modest hedging flows, as will high options gamma structures).
I'm quite surprised that the Metaculus article doesn't mention the recently FDA-approved Setmelanotide nor any of the other melanocortin small molecule agonists currently in development (e.g. LR-19021).
Guyenet argues that the obesity epidemic is caused by modern hyper-palatable processed foods that entice us to overeat. The Big Food industry engineers these foods in labs to be enticing. As the industry has improved its techniques, it has surrounded us with more and more tempting food at lower and lower prices.
A negative scenario is that the palatability of processed foods evolves in response to the introduction of these GLP+ drugs. What if the current performance of semaglutide is like the initial performance of a new antibiotic? It's a miracle drug at first, but Big Food responds by updating its processed food recipes, producing something analogous to antibiotic-resistant bacteria.
Through the Darwinian selection of capitalism, processed foods that partially overcome the power of the new drugs attain more shelf space and, eventually, "share of stomach". Big Pharma responds with new drugs that, for a while, have stronger efficacy. Etc. So we end up with something closer to a dynamic stalemate than to a total victory over obesity.
Great to see coverage on Google's new prediction market!
[Disclaimer: I created and run Gleangen. Opinions are mine, not Google's.]
First, you're right, the accuracy graph is confusing, sorry about that. The red line counts the number of markets in each forecasted probability decile that actually happened. The first point means that for markets that Googlers gave a 0-10% chance of happening (at various time points before expiry), slightly more than 5% happened. By this metric, markets Googlers give a low probability happen more often than predicted, and markets Googlers give a high probability happen less often.
Second, on the Hal Varian quote: Personally, I think there's a very large space for questions that are (a) are of interest to the company but (b) do not reveal financially critical information. The most basic case is forecasting engineering timelines. More interesting is predicting tech advances, such as Google's quantum computing milestones: https://quantumai.google/learn/map.
In general, I agree with Scott's take. So far all we’ve demonstrated is that a prediction market can be popular, employees value the information it generates, and it can be fairly accurate. Whether it can improve Google’s high-level strategy remains to be seen.
With regard to the SAT/ACT thing: I think some of the higher predictions may have to do with the wording of the question i.e. there could be an alternative standardised test that would technically satisfy the question.
I haven't read Hayek, but it seems like there's an obvious problem with this:
> COWEN: Why doesn’t business use more prediction markets? They would seem to make sense, right? Bet on ideas. Aggregate information. We’ve all read Hayek.
A prediction market might provide info on, e.g., "will China invade Taiwan in the next 10 years". Markets might provide info on, e.g., the market-clearing price of a pound of copper. But it makes sense for the market to provide more info on the market-clearing price of a pound of copper, because that price is determined by the desires of people to buy/sell copper (or stuff that is made with copper). If more people suddenly want copper for some reason, then that increases both the free market's price of copper, and the market-clearing price of copper, because they're the same thing; whereas "a central planner's estimate as to the market-clearing price of copper" is very much not the same thing. The reason the market "works" isn't some generic "wisdom of crowds" idea, it's that the desires of the crowd - even if the crowd is all dumbasses! - directly defines what the price should be in the first place
But for the "invade Taiwan" question there's no link between the prediction market and the reality. I don't see why it would provide any information on that question at all.
There can be a more direct link in the case of, say, an election, where the result really is an aggregation of what people want. But even then there are some obvious differences between what a prediction market would say, and what the election result would be, that make it not like markets and Hayek.
Options are tied to the underlying market by market-makers (or "arbitrageurs," who are functionally equivalent to market-markets in derivative markets). But market-makers only show up when there is both (a) enough money to pay them for their service, and (b) enough liquidity in the underlying for them to eliminate their risk. The second criterion prevents an options market from creating a market in the underlying (long-term prediction) that doesn't otherwise exist. The market in the underlying has to exist and trade with sufficient liquidity to support the derivatives market. Otherwise the options, if created, will not be derivatives but rather will be their own untethered speculative arena.
Regarding options, in financial markets they can move the market itself. This happens in the "manufacturing" of these options and the hedging that the seller has to make to balance their obligations in the options.
I am not sure how this would work in a prediction market, as I don't know what kind of margin requirements you would have in such a case.
"Lots of people think it will be 100%, with a long tail of people predicting various other things."
This doesn't seem right to me. That graph corresponds to the aggregate thinking 100% is quite likely, but this could be because everybody thinks there's a, say, 30% chance of that, rather than 30% of people thinking it a certainty.
How much can a single whale move the cursor on Metaculus if they're willing to buy/sell whatever it takes at a loss ? If you start using it to make predictions conditional on election results, and use the results as alectoral arguments as in the prison thing, I'd say it becomes a real hazard or at least a temptation.
You are asking if "is there some way to combine regular trading and options trading into a single price, so that we could get the advantages of options trading, but traders would still be betting on the regular market and increasing our predictive accuracy?"
If I understand your question correctly, by "regular trading" you mean the prediction of the event itself, such as "Bolsonaro is on trial before 2025", while by "option trading" you mean trading in contracts like "the market for <<Bolsonaro on trial before 2025>> is pricing probability more than 30% as of 1st of July 2023". In that case, the answer is that the option market already contains the regular market, so there is nothing to combine.
To see why that is the case, let C be the base contract "Pay $1 on 01/01/2025 if Bolsonaro is on trial before 2025, pay $0 otherwise". As of today, there is a simple arbitrage between C and the contract C23 defined as "Pay the price of C as observed on the market on 1st July 2023". If we ignore interest rates and counterparty risk, the price of these two contracts should be simply equal until 1st of July 2023 and if it is not, one can buy the cheaper one and sell the more expensive one to lock in a riskless profit. With non-zero interest rates and appropriate counterparty risk mitigation, the prices of C and C23 will diverge but an arbitrage link will persist.
Now, C23 is a special case of an option (namely, it is a call option on C with strike of $0 and maturity 07/01/2023.) So the options market already contains the market for C.
In fact, traders on an options market will be keen watchers of whatever is the most liquid market in the base contract C or anything equivalent to it, like C23 or C22 or whatever and will be not just updating their prices with movements in the base market but trading actively in the base market as it moves to hedge their option positions. So the market for C will be closely linked with the options market as a whole, not just with the market in the 0-strike C23.
Re. "Suppose you want to predict something in 2100. It’s hard. Nobody wants to lock their money up for 80 years to get a 5% or 10% or even a 100% rate of return. ... One option we talked about last time is chained prediction markets."
When you buy a stock, you're predicting the time-discounted returns of that stock to infinity. So use the stock solution: Instead of giving a probability, have prediction-market users buy a "true" or a "false". The relative prices of a T and a F are determined by the (Laplace adjusted) ratio of Ts and Fs outstanding. The question is capitalized by both "true" and "false" purchases, and at resolution all money is divided up among holders of stock in the correct prediction. The price of a "true" vs a "false" should translate directly into an odds ratio whenever you want an odds ratio or a probability, and anyone who wishes to exit the market can sell their shares back to the now-capitalized question at the current price.
(I'd always assumed that was how prediction markets worked; but Scott's statement, quoted above, implies that they don't work that way.)
Re. "Conservatives are right-wing": NO. Stop saying that. It's extremist, left-wing and right-wing propaganda, whose purpose is to present our political choices as between either left or right, thus obscuring the fact that left-wing and right-wing are, for all practical purposes, just different flavors of the same mindless Hegelian poison.
"Right wing" has many distinct meanings, including these:
- The nobility sat to the right of the presiding officer's chair in the Estates General of 1789 at the start of the French Revolution. "Right-wing", using this precedent, means "in favor of hereditary aristocracy", and more-generally, "in favor of class hierarchy without social mobility".
- Left Hegelians (today you can read this as Marxists, Unitarians and other non-literalist Christians, and self-described "Progressives") pursue a religious, apocalyptic, historicist (teleological) program based on ideological purity; while Right Hegelians (nationalists, fascists, Nazis) pursue a religious, apocalyptic, historicist (teleological) program based on racial or ethnic purity. Right Hegelians historically formed temporary alliances with conservative religious groups, primarily pre-Vatican 2 Catholics and the Orthodox and Russian Orthodox Churches, as they did in fascist Spain, Italy, Germany, and Yugoslavia.
- According to Wikipedia, "Right-wing_politics", the US Dept of Homeland Security defines "right-wing" as being either hate-based, or preferring conflicts to be resolved at the state level rather than at the federal level--IMHO a shockingly biased definition which attempts to associate local political autonomy with racial hate.
None of these meanings of "right-wing" correlate positively with conservatism. Conservatism today in the US means Enlightenment liberalism: supporting the continuation of free markets, free speech, mistake theory, and liberal democracy. Both left and right wings have always been resolutely anti-liberal and pro-conflict-theory.
The French nobility were conservative in their time, but would be radical today; and the centrally-planned monarchy they believed in was much closer to communism than to an individualistic, free-market democracy. Fascists were never conservative. They sought to create a new society, using the fantasy of an imaginary distant past as propaganda, just as Marxists do.
The Social Justice movement today can't be classified as either left or right. It has strong ties to Marxism through the 1960s student movements and the resulting post-1990 aging-hippy university faculties; and to fascism through Nietzsche, continental phenomenology and existentialism, post-modernist philosophers who studied under Nazi professors (e.g., Michel Foucault and Herbert Marcuse), and the associated resurgence of Plato (who literally *invented* racial essentialism, for use in creating his perfect State). It views culture through the Nazi lens of racial essentialism, assuming that culture is a manifestation of a racial essence and therefore culture maps 1-1 onto race. This is why it calls cultural assimilation racially oppressive (e.g., blacks who live in the suburbs are being oppressed by a false racial consciousness).
The social division in the US today is between urban and rural culture. Rural culture doesn't align with nationalism; it correlates with nationalism only because the dominant ideology of urban culture today is actively anti-American, and saying you're proud of America can get you Twitter-mobbed. Rural culture is more religious; yet the right-wing religious--Catholic, Orthodox, Muslim, and Hindu--all now lean Democrat (https://www.pewforum.org/religious-landscape-study/party-affiliation/). Except for Mormons, the only majority-Republican religious sects are Protestant sects formed as left-wing dissenters from the right-wing Catholic hierarchy.
[Epistemic status: I am a retail investor who's been pretty serious about it for 10 years plus moderately active on predictit/metaculus. 75% confidence]
I thought about prediction market options some more, and I think liquidity concerns make me prefer directly using leverage on the underying contracts, instead of options.
Example: Suppose JGalt wants to bet $100 each on 100 different markets. Total bet value is $10k. How much should JGalt have to deposit to cover this if the markets are all uncorrelated coin flips? He wins a mean of 50 contracts with a standard deviation of 5, so to cover 4 standard deviations of negative variance, he needs 4*5*$100=$2000 in his account, so that he is getting 5x leverage. This is conservatively assuming we let all the contracts run to completion, but the system could get some additional protection by preemptively liquidating some of JGalt's contracts if the aggregate of their market prices moves unfavorably and he gets down to <2SD of margin.
This math can be generalized to cover a set of contracts with arbitrary variance and covariance. To protect the site from people who might lose a bunch of correlated bets to get a negative balance and then not pay their debt, admins would need to estimate covariance. Each time they create a new market, they need to run a search for related markets and override the default-zero covariance appropriately. In the limit as you assume every market has 100% covariance with every other market, you get what most prediction markets do, which is to not let anyone use any margin at all (except within a single market where it's logically impossible to lose both contracts).
What could go wrong:
* Maybe admins are bad at estimating covariance, so someone is allowed to make a bunch of correlated bets that all lose, and his account goes negative, and he doesn't pay his debt.
But options based systems have all the same margin problems, *plus* liquidity problems that make it much harder to liquidate accounts that fall below an adequate margin. So I'm 75% confident that leverage on the underlying would make a better prediction market than options.
Re. weight-loss drugs, this ignores the fact that the medical establishment is morally opposed to weight-loss drugs. Most doctors feel that prescribing a fat person a pill that would make them thin, instead of teaching them good Puritan self-control, would be immoral. This is why, for perhaps the past 40 years, the foremost weight-loss pill has always been subjected to a withering attack from the medical establishment until it was effectively banned.
Today, the most-effective weight-loss pill I know of is phentermine. It was developed in the 1960s as an alternative to amphetamines. Studies done at the time showed it was non-addictive. Recent studies have also all concluded it's non-addictive. But as soon as the previous effective weight-loss pill (IIRC it was ephedrine) was banned, phentermine came under attack. If you read Wikipedia today, or for-the-public websites like WebMD, they'll say that phentermine is an addictive amphetamine, even though it is not an amphetamine and is not addictive, except under the new "my patient resists stopping this drug because it works" definition of "addictive". It's now a scheduled drug with stringent requirements even when issued by prescription, which were tightened just this year to ensure that you have to be obese to be allowed to use it, and will have to stop using it before reaching a healthy weight.
verywellmind.com says (https://www.verywellmind.com/how-long-does-adipex-stay-in-your-system-80217), "Phentermine increases levels of norepinephrine, dopamine, and serotonin, producing an effect similar to amphetamine. For this reason, phentermine is a controlled substance and only available legally via a prescription." In reality, the stimulant effects of phentermine are like those of caffeine, which also effectively increases increases levels of norepinephrine, dopamine, and serotonin. (I wrote "effectively" because this may in part be mediated by increasing dopamine receptor availability; https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4462609 .)
(I don't recommend ephedrine because it works by speeding up your metabolism, which surely shortens lifespan. I don't recommend phen-fen because apparently it really is dangerous. Phentermine enabled me to permanently lower my stable body fat from over 14% of body weight to perhaps 12%, with a total weight loss around 3%, but weight loss did not continue past that level. That was on 19mg EOD, which is 1/4 of the usual dose. You can lose more weight by combining it with Topamax, but I found that made me feel tired all day long. I eventually stopped using phentermine entirely because, like caffeine, it gives me migraine headaches.)
If the market is large enough to support statistical arbitrage, then options traffic will be reflected in the underlying. AFAIK no such liquid options markets exist today.
> I bet a lot of people would care a lot if the conservatives could produce 0.01% higher GDP growth vs. 5%.
What if we want to know how much GDP is actually driven by policy, if at all? Can we use prediction markets to get a sense of the correlation coefficient?
We might naively expect we could just look at the distribution of the predictions to determine the variance of outcomes, but I'm really nervous about that. Prediction markets combine wisdom of crowds and incentives for knowledge to chase out bad predictions. But there's no direct and obvious way for good distributional information to chase out bad distributional information on this second layer. If a bunch of crazy people are guessing randomly all over the place, and you're the smart money and you're really sure the correlational coefficient is much tighter, there's no additional incentive for fixing the shape of the graph after you've committed the appropriate amount for where the outcome will land. If the smart money has already gotten in, there's no payout for you to go in and tweak the shape of the graph.
So I think we'd need a new tool on top of prediction markets to tease this out and I'm not sure what it would be.
If you had an infinite number of players, then maybe you could have a bunch of markets with different precisions. Predict the impact on next year's GDP to the trillion, to the hundred billion, down to the individual dollar if you like. Nobody's going to take the individual dollar bet. Nobody's going to bet in the tens of trillions because it will be a fixed market. The market with the most action might tell us where the biggest area of debate is that is plausible to predict, which might tell us something about the likely variance in outcomes?
Aside from being utterly impractical, that might be an interesting direction to explore. :D
Anyone have any other market technology to help predict how tightly A causes B? Maybe this is a solved problem already and I'm just missing it.
Options Markets today are hedged. When you buy an options contract a market maker hedges their risk by buying (some of) the underlying contact, this moves the market.
I predict that the UK's GDP will be higher five years after the next election if the Conservatives win than if Labour does. Why? Nothing to do with what either party would do if they win; it's just that the Conservatives are in government now, and economic growth is correlated with incumbent parties getting re-elected.
What about GDP growth starting from after the next parliament takes office? Perhaps the expected GDP growth should be higher if Labour wins than if the Conservatives do, again not because of their policies, but because expected GDP at the time of the election will be higher if the Conservatives win than if Labour does, and it could regress to the mean.
This sort of thing will always be a problem when trying to use conditional prediction markets to influence decisions that are also influenced by other things. The differences in conditional expectations don't tell you the direction of causation.
by the by, about Bolsonaro, I know a few reliable people who were in Brazil for the thick of Covid 19 there(which I am told was overblown in terms of media coverage), and no amount of telling their compatriots to stay home was going to be sufficient. What India did about a similar issue was to have non-compliant bar-goers bludgeoned with sticks by brute squads / Lou Ferrigno.
If Metaculus were a real-money market, you could imagine a world where in 2029 there existed some SMR nuclear reactors, but not in sufficient quantities or in economies where they made sense to deploy at scale for national electric power, but someone floated a few into whatever the smallest, poorest island nation is to supply 1% of its electric power and win the market.
But, you know, I guess that won't be worthwhile even if possible, unless Metaculus leaderboard positions become REAL valuable in the next 8 years.
Maybe we should create a prediction market about the real-world dollar value of a position on Metaculus leaderboard in 2030.
I don't have experience with google's new prediction market. However, when I worked there a decade ago, I did use prophit, which also had "play money with leaderboards" as the reward. It was not quite the right incentive system. Back then at least, leaderboards and play money were reset on a quarterly schedule, and the people at the top of the leaderboard got gift certificates (I think $75 was typical). This means it was best to make somewhat-risky bets with your finite play money and hope they paid out. If there were a market where the outcome were 100% known (but that would only pay out at end of quarter), it would still have been a losing strategy to bet at 70% odds, I'd guess.
On the questions of the form "If this party wins, what will be the value of this metric?", in a one-shot I don't think you learn much except the political prejudices of metaculus users.
But if you did this a really large number of times over a really large number of elections, you could probably learn something. If unemployment is consistently forecast to be 0.5% higher under Party X than Party Y, and unemployment is consistently higher than the Metaculus estimate in years when Party X wins and consistently lower than the Metaculus estimate in years when Party Y wins, then you might be able to figure out exactly how wrong Metaculus users are, subtract it, and figure out the actual effect of Party X vs Party Y on unemployment. (Of course you'd need to wait a very long time to have enough data for this to be useful.)
Regarding "looking for options" -- prediction markets could adopt methods similar to what financial markets use for cleared derivatives transactions. Lets say you want to buy 100 shares of "yes" on some question at a 50/50 chance. You are risking $50, and your total payout will be $100 if correct, in 2100. The daily volatility in the contract may be some fraction of 1%, and a really big price move in the prediction for something that far out might be 2%. The market could simply make you post $2 upfront instead of the entire $50. As the market moves either direction, you get credited or debited 100% of each market move. If you do not have the money to cover the margin call when the market moves against you, the clearing agent liquidates your position. That money reserve could be either posted at the prediction market exchange or take the form of say having a credit card on file.
The obvious benefit is that this method reduces the initial opportunity cost of capital by 96%. The downsides?
(1) People might be worried about getting margin called. Of course they can prevent this completely if they want to buy just keeping $48 in their account in my example, so they are no worse off than the status quo regardless. If they make predictions in many different questions, they likely have lots of diversification benefits and their risk of being proven hugely wrong on all their questions at the same time is quite low. Over the life of their predictions they will need to keep some cash buffer, so the lifetime opportunity cost of capital isn't 96% saved, but could be pretty close anyway.
(2) The market might move by more than the required $2 margin amount, and the people on the losing side might opt not to make good on their payment obligation to cover the difference (i.e. they default). This risk is mitigated when predictors have lots of different positions on different prediction contracts on the same exchange (cross-collateralization). In financial markets, the clearinghouse itself acts as guarantor against defaults, so no predictor would be taking the credit risk of the other predictors. The clearinghouse takes the risk, and charges some sort of fee to fund a pot of money from which to pay for any future credit defaults.
It's not a new problem and a complex market has taken a pretty good shot at an efficient solution. If you have a better solution you can probably make at least hundreds of millions of dollars displacing the CME or LCH from their current gigs.
Say I make a prediction one year out, and then drastically adjust it with 1 month left (closer to what will actually happen), then my Brier score is going to be pretty bad right? But if the majority of votes came in with only 1 month left, being close to the actual outcome, does that mean overall Brier score will be good?
And I think predictions also need a complexity score. So for example predicting 2022 inflation accurately is probably more impressive than predicting the outcome of a sports match (which is a lot more local). And predicting something further out is generally more impressive as well, the more global it is.
A value measure for each prediction could sort of solve this, where each person has a limited number of upvotes they can give to predictions. Where predictions with higher scores count more towards overall prediction score.
Re: Looking For Options
Black and Scholes showed in their famous model that (under some assumptions that have been made less restrictive over time) you can replicate the performance of an option using a mixture of cash (or a loan) and holding the underlying asset, then re-balancing the portfolio of the underlying and the cash over time. If these sorts of assumptions can hold here, then this is wrong: "But if nobody is buying shares directly and all the action is in options trading on the side, that won’t work. " People buy options, market making option sellers price those options and hedge their risk using the underlying prediction market, and the market can reflect the beliefs of the people buying the options. The hedging portfolios that the market makers create (a replicating portfolio) brings the information from the option buyer into the market prices of the underlying asset.
Surely the obvious problem with political prediction markets is that money that wanted to make The Other Party look bad would drown out smart money that actually wanted to predict for a return?
In re obesity predictions: Are you betting that the definition of obesity won't change? Or you'll go with whatever the definition is then?
What's the argument for more Americans being obese? The drugs don't work that well? They don't work that well on a lot of people? There are side effects and the drugs are withdrawn? The drugs are expensive and/or otherwise not available? Odds of a black market if they aren't available?
About the conditional markets and politics:
I agree that politicians will probably try to wave away bad looking results and the result of tradeoffs, but isn't that still better than the political discourse we have now?
I guess I don't know the British talking points, so I'm switching to American politics here. Like I can imagine Republican politicians responding to conditional markets saying that Democrats will lead to better test scores by saying that it's only because they'll tax you dry to fund education (cue a discussion of the conditional markets on tax rates). But today I'd imagine the Republican saying that they'll produce higher test scores *even though* the Democrats are spending more money on it because the Democrats will be spending money on teaching white kids to hate themselves or something rather than any useful educational goal.
This sums up beautifully what I think of the whole prediction market biz:
"You’re not really learning anything you wouldn’t have guessed otherwise. But it’s a proof of concept."
These prediction markets are worthless because the only predictions that really matter are the non-linear ones.
AI/machine learning can do linear extrapolations just fine.
The 2nd, more fundamental problem with prediction markets is that they are ultimately flawed. If any market shows any promise - it is then trivially skewed by a small amount of money/effort. If y
ou can't trust them early, and you can't trust them once they've proven themselves, that is the literal definition of useless.
1. "The whole point of prediction markets is that their prices correspond to the chance of something happening. But if nobody is buying shares directly and all the action is in options trading on the side, that won’t work..."
It is possible to arbitrage any price discrepancies between options and the underlying stock. For example, if you buy a put and sell a call at the same strike and same expiration date, that is a synthetic short position that can be perfectly hedged by buying the underlying. If the options are already cheaper than the underlying, you can arb it by doing the opposite thing (buy call, sell put, sell underlying)
People will quickly arb any price discrepancies, so there's no risk of the prediction market getting far out of whack that way. The biggest problem I see is that options will be too illiquid. On wall street, companies under $5B don't have liquid options for next month, and even $500B megacorporations have 20% bid-ask spreads on options expiring 2 years in the future. That's even worse than Predictit fees! People can put limit orders in the middle and wait a long time for someone to fill them, and hope they don't get past-posted when news comes out, but that experience kinda sucks.
2. Caution on conditional prediction markets.
Betonline has some hilarious alien-themed markets (https://www.betonline.ag/sportsbook/futures-and-props/entertainment/aliens) for who will get abducted by aliens first, which country will get attacked by aliens first, and who will win the alien v human war. All bets are void if the aliens don't show up by 12/31/2021. So they're basically getting free loans from whoever bets these things. People have nothing to lose except interest, so they don't care what the lines are and might bet it just to generate reward points or something.
The more improbable the condition, the less skin in the game people have. If the condition is <1%, then probably nobody really wants to bet it except for ideological reasons or silly reasons that have nothing to do with the correct conditional probability.
If you want a prediction market for something that will happen (or not) in 2100, then that market needs to trade not in 202x dollars but in safe bonds maturing at 2100. This is important both because of the time value of money (since a 100% certain win of $100 at 2100 is likely worth just a few dollars now, so you should not have to put up $100 so that the other side would expect to get $100 at the end of the bet) and because of the need for the market to decouple the fluctuations of the actual bet from the fluctuations on the expected interest/discount rate for that decades-long bond, which - at least when the payoff is still distant - would be much more influential than the actual expected value of the bet.
>SMR Metaculus
How do they define 'Nation'? Per the question, they have 'Resolution may come from credible media reports, government agencies, or energy industry researchers such as the IEA.' as a resolution criteria, but that isn't clear on who exactly counts. Does Israel count? They still aren't fully recognized. Kosovo? Less recognized, but recognized by some big players. Taiwan? Barely recognized. Bougainville? Aiming for independence but arguably not there yet. Scotland? In no way independent but may or may not want to be, also still labelled a kingdom.
Can I declare independence as dictator of the Russian city Metaculus links as having a Small Nuclear Reactor already, have an article written about me as a crackpot trying to secede from Russia, and win?
I think the closest correlate to you option market conundrum is VIX futures. Since there is no such thing as trade able spot vix, the curve exhibits some interesting behavior.
Scott,
I don’t know if the following helps re options since I didn’t read the previous post in depth but here’s something:
Options can be arbitraged with their underlying asset in the following way: if you buy a call and sell a put at the same strike , that’s equivalent to buying the stock for the premiums plus the strike since you will either exercise the call or be forced to buy by the holder of the put depending on where them price of the underlying ends up.
Wegovy. The drug was approved by the FDA on June 4. It is very expensive (>$300/week). After a failed attempt to obtain insurance coverage, I began a course of treatment in September. The course of treatment described by the package insert is to begin at 0.25 mg and escalate by steps to the therapeutic dose of 2.4 mg, which will be my next step. I have had none of the side effects described in the warnings. By the time I completed the third step, I had already lost 7% of my body weight. When I started my BMI was almost 40. My goal is to get that down below 30, at which point i will be merely overweight. Right now it seems to be very doable.
The biggest problem with mass use of Wegovy are 1. cost/insurance, which in the long run will be solved. Every one of the five other prescriptions I take was once patented and expensive. They are now all off patent (one is now OTC) and are a lot cheaper. And 2. the supply chain (sigh):
"Novo Nordisk (NOVOb.CO) shares fell as much as 16% on Monday after the Danish drugmaker was hit by U.S. supply issues for its new obesity drug as it seeks to establish a market foothold before the launch of a rival drug by Eli Lilly.
"The obesity market has proved difficult for pharmaceuticals companies, but Novo made a breakthrough with its Wegovy drug. The company hopes the drug, which helps to achieve weight loss of 17% on average over almost two years, will offset growing pressure on its core insulin business. read more
"In a stock announcement after Friday's market close, Novo said that a contract manufacturer filling syringes for pens to inject the drug had halted deliveries and manufacturing temporarily after issues relating to good manufacturing practice.
"The Danish drugmaker was already facing separate supply constraints after being "overwhelmed" by initial uptake of the drug. read more
"As a consequence, fewer new patients will start treatment with the drug in the first half of next year. Novo now expects to be able to meet U.S. demand in the second half of next year, having previously estimated it would do so at the beginning of the year."
https://www.reuters.com/business/healthcare-pharmaceuticals/novo-nordisk-shares-slump-us-supply-challenges-obesity-drug-2021-12-20/
> You could do this for [...] school test scores
And if you did, the question you'd be predicting is "which party will fiddle the figures more by making exams easier".
"I don’t know, maybe you wouldn’t learn anything there either. If Conservatives had better GDP growth, maybe leftists would say “of course, they’re the more capitalist party, they trade off environmental damage and inequality for a slightly hotter economy”. If Labour had better test scores, maybe rightists would say “of course, they’re the more socialist party, they’ll tax you dry and throw some of the money at schools but it will still be inefficient.”"
I think this is pessimistic. Yes, of course this is the kind of thing each side will say, but isn't that how it should be? Imagine a world where there are established prediction markets for public policies like this with strong track-records of accuracy. Then the public debate could be about what what values people want to support: crime reduction vs. mass incarceration; economy vs. environment, etc. rather than the byzantine factual/counterfactual arguments about which side's policies would *really* be better for crime reduction or economy that we have today. Also it would encourage parties to drop policies that suck at a achieving their intended/stated ends.
> Here I admit I don’t know much about markets or options - is there some way to combine regular trading and options trading into a single price, so that we could get the advantages of options trading, but traders would still be betting on the regular market and increasing our predictive accuracy?
Yes, this would happen naturally and automatically via arbitrage. See:
https://en.wikipedia.org/wiki/Put%E2%80%93call_parity
https://en.wikipedia.org/wiki/Replicating_portfolio
> That peak at the end is weird-looking and shows up on many Metaculus questions. I think this is something like all dates after April 1 getting crammed into April 1 for the graph, and even though very few people think it will be after April 1, that’s still a really high number of guesses when it’s all concentrated into an individual day.
I assumed this was happening because users really put their prediction on the last day of the time window to signal “I don’t believe this will happen within the time window of this question”. Many questions include “If this does not occur by [last selectable date], it resolves as > [last selectable date]” in the fine print.
Are there any prediction markets that predict their own continued existence or eventual demise?
Did Robin Hanson really invent prediction markets? On the Wikipedia article on prediction markets, I see that he made the first known corporate prediction market in 1990, but it also says that the University of Iowa's market first ran in the 1988 presidential election. Did he make one earlier?
> Suppose you want to predict something in 2100. It’s hard. Nobody wants to lock their money up for 80 years to get a 5% or 10% or even a 100% rate of return.
A partial solution to this problem:
Prediction markets presumably provide a little bit of diversification, so the mean-variance-optimal portfolio contains at least a small allocation to prediction markets, regardless of how low the expected return is (as long as it's still positive). Investors can add leverage until they reach the desired level of risk.
This solution is not complete because it only works if you can borrow money at the risk-free rate. Pretty much nobody can do that. Even if you're only paying an extra 0.5% for leverage (which is something like what institutional investors pay), that could be enough to wipe out your expected return. So long-dated predictions are still worth buying *if* you can get cheap enough leverage. The question then becomes how to get sufficiently cheap leverage.
"My next question is: is there a structure where options directly move the market? The whole point of prediction markets is that their prices correspond to the chance of something happening. But if nobody is buying shares directly and all the action is in options trading on the side, that won’t work. Here I admit I don’t know much about markets or options - is there some way to combine regular trading and options trading into a single price, so that we could get the advantages of options trading, but traders would still be betting on the regular market and increasing our predictive accuracy?"
Options move the stock market through arbitrage and the hedging of market makers (they don't want to make or lose money on the outcome of the options they buy/sell, so they buy/sell the underlying security to offset price moves and stay "delta neutral.") Presumably something similar could work for a prediction market that had a liquid market supported by market making firms.
On the subject of options effects on security prices when they dominate the flows of their underlying, investory/day trader types might be intersted in googling about NOPE, the Net Options Pricing Effect, a calculation invented in 2019 by Lily Francus and some cohorts that tries to guess when stock prices (specifically SPX index or SPY ETF given their liquidity) will have a near term reversion due to market maker hedging - if too many calls or puts are bought relative to the actual stock, they eventually have too much built up pressure and risk and have to engage in hedging purchases/sales (usually through /ES futures, which are abritraged to effect the SPX/SPY prices) that will cause a short term "snap back" in the underlying price until the put/call flow becomes more balanced.
It's had some useful backtesting as well as identifying conditions where it doesn't work well (actual market moving news is going to run over relatively modest hedging flows, as will high options gamma structures).
See here, and the whitepaper link at the top.
https://nopechart.com/
I'm quite surprised that the Metaculus article doesn't mention the recently FDA-approved Setmelanotide nor any of the other melanocortin small molecule agonists currently in development (e.g. LR-19021).
Highly speculative re weight loss drugs:
Guyenet argues that the obesity epidemic is caused by modern hyper-palatable processed foods that entice us to overeat. The Big Food industry engineers these foods in labs to be enticing. As the industry has improved its techniques, it has surrounded us with more and more tempting food at lower and lower prices.
A negative scenario is that the palatability of processed foods evolves in response to the introduction of these GLP+ drugs. What if the current performance of semaglutide is like the initial performance of a new antibiotic? It's a miracle drug at first, but Big Food responds by updating its processed food recipes, producing something analogous to antibiotic-resistant bacteria.
Through the Darwinian selection of capitalism, processed foods that partially overcome the power of the new drugs attain more shelf space and, eventually, "share of stomach". Big Pharma responds with new drugs that, for a while, have stronger efficacy. Etc. So we end up with something closer to a dynamic stalemate than to a total victory over obesity.
Great to see coverage on Google's new prediction market!
[Disclaimer: I created and run Gleangen. Opinions are mine, not Google's.]
First, you're right, the accuracy graph is confusing, sorry about that. The red line counts the number of markets in each forecasted probability decile that actually happened. The first point means that for markets that Googlers gave a 0-10% chance of happening (at various time points before expiry), slightly more than 5% happened. By this metric, markets Googlers give a low probability happen more often than predicted, and markets Googlers give a high probability happen less often.
Second, on the Hal Varian quote: Personally, I think there's a very large space for questions that are (a) are of interest to the company but (b) do not reveal financially critical information. The most basic case is forecasting engineering timelines. More interesting is predicting tech advances, such as Google's quantum computing milestones: https://quantumai.google/learn/map.
In general, I agree with Scott's take. So far all we’ve demonstrated is that a prediction market can be popular, employees value the information it generates, and it can be fairly accurate. Whether it can improve Google’s high-level strategy remains to be seen.
With regard to the SAT/ACT thing: I think some of the higher predictions may have to do with the wording of the question i.e. there could be an alternative standardised test that would technically satisfy the question.
I haven't read Hayek, but it seems like there's an obvious problem with this:
> COWEN: Why doesn’t business use more prediction markets? They would seem to make sense, right? Bet on ideas. Aggregate information. We’ve all read Hayek.
A prediction market might provide info on, e.g., "will China invade Taiwan in the next 10 years". Markets might provide info on, e.g., the market-clearing price of a pound of copper. But it makes sense for the market to provide more info on the market-clearing price of a pound of copper, because that price is determined by the desires of people to buy/sell copper (or stuff that is made with copper). If more people suddenly want copper for some reason, then that increases both the free market's price of copper, and the market-clearing price of copper, because they're the same thing; whereas "a central planner's estimate as to the market-clearing price of copper" is very much not the same thing. The reason the market "works" isn't some generic "wisdom of crowds" idea, it's that the desires of the crowd - even if the crowd is all dumbasses! - directly defines what the price should be in the first place
But for the "invade Taiwan" question there's no link between the prediction market and the reality. I don't see why it would provide any information on that question at all.
There can be a more direct link in the case of, say, an election, where the result really is an aggregation of what people want. But even then there are some obvious differences between what a prediction market would say, and what the election result would be, that make it not like markets and Hayek.
Yowza. Semaglutide sounds really exciting except for the part where it costs almost as much as my mortgage payment per month.
Options are tied to the underlying market by market-makers (or "arbitrageurs," who are functionally equivalent to market-markets in derivative markets). But market-makers only show up when there is both (a) enough money to pay them for their service, and (b) enough liquidity in the underlying for them to eliminate their risk. The second criterion prevents an options market from creating a market in the underlying (long-term prediction) that doesn't otherwise exist. The market in the underlying has to exist and trade with sufficient liquidity to support the derivatives market. Otherwise the options, if created, will not be derivatives but rather will be their own untethered speculative arena.
This could be a usecase for perpetual options? Or a new derivative I learned about recently, "squeeth" https://www.paradigm.xyz/2021/08/power-perpetuals/
Regarding options, in financial markets they can move the market itself. This happens in the "manufacturing" of these options and the hedging that the seller has to make to balance their obligations in the options.
I am not sure how this would work in a prediction market, as I don't know what kind of margin requirements you would have in such a case.
"Lots of people think it will be 100%, with a long tail of people predicting various other things."
This doesn't seem right to me. That graph corresponds to the aggregate thinking 100% is quite likely, but this could be because everybody thinks there's a, say, 30% chance of that, rather than 30% of people thinking it a certainty.
How much can a single whale move the cursor on Metaculus if they're willing to buy/sell whatever it takes at a loss ? If you start using it to make predictions conditional on election results, and use the results as alectoral arguments as in the prison thing, I'd say it becomes a real hazard or at least a temptation.
You are asking if "is there some way to combine regular trading and options trading into a single price, so that we could get the advantages of options trading, but traders would still be betting on the regular market and increasing our predictive accuracy?"
If I understand your question correctly, by "regular trading" you mean the prediction of the event itself, such as "Bolsonaro is on trial before 2025", while by "option trading" you mean trading in contracts like "the market for <<Bolsonaro on trial before 2025>> is pricing probability more than 30% as of 1st of July 2023". In that case, the answer is that the option market already contains the regular market, so there is nothing to combine.
To see why that is the case, let C be the base contract "Pay $1 on 01/01/2025 if Bolsonaro is on trial before 2025, pay $0 otherwise". As of today, there is a simple arbitrage between C and the contract C23 defined as "Pay the price of C as observed on the market on 1st July 2023". If we ignore interest rates and counterparty risk, the price of these two contracts should be simply equal until 1st of July 2023 and if it is not, one can buy the cheaper one and sell the more expensive one to lock in a riskless profit. With non-zero interest rates and appropriate counterparty risk mitigation, the prices of C and C23 will diverge but an arbitrage link will persist.
Now, C23 is a special case of an option (namely, it is a call option on C with strike of $0 and maturity 07/01/2023.) So the options market already contains the market for C.
In fact, traders on an options market will be keen watchers of whatever is the most liquid market in the base contract C or anything equivalent to it, like C23 or C22 or whatever and will be not just updating their prices with movements in the base market but trading actively in the base market as it moves to hedge their option positions. So the market for C will be closely linked with the options market as a whole, not just with the market in the 0-strike C23.
Re. "Suppose you want to predict something in 2100. It’s hard. Nobody wants to lock their money up for 80 years to get a 5% or 10% or even a 100% rate of return. ... One option we talked about last time is chained prediction markets."
When you buy a stock, you're predicting the time-discounted returns of that stock to infinity. So use the stock solution: Instead of giving a probability, have prediction-market users buy a "true" or a "false". The relative prices of a T and a F are determined by the (Laplace adjusted) ratio of Ts and Fs outstanding. The question is capitalized by both "true" and "false" purchases, and at resolution all money is divided up among holders of stock in the correct prediction. The price of a "true" vs a "false" should translate directly into an odds ratio whenever you want an odds ratio or a probability, and anyone who wishes to exit the market can sell their shares back to the now-capitalized question at the current price.
(I'd always assumed that was how prediction markets worked; but Scott's statement, quoted above, implies that they don't work that way.)
Re. "Conservatives are right-wing": NO. Stop saying that. It's extremist, left-wing and right-wing propaganda, whose purpose is to present our political choices as between either left or right, thus obscuring the fact that left-wing and right-wing are, for all practical purposes, just different flavors of the same mindless Hegelian poison.
"Right wing" has many distinct meanings, including these:
- The nobility sat to the right of the presiding officer's chair in the Estates General of 1789 at the start of the French Revolution. "Right-wing", using this precedent, means "in favor of hereditary aristocracy", and more-generally, "in favor of class hierarchy without social mobility".
- Left Hegelians (today you can read this as Marxists, Unitarians and other non-literalist Christians, and self-described "Progressives") pursue a religious, apocalyptic, historicist (teleological) program based on ideological purity; while Right Hegelians (nationalists, fascists, Nazis) pursue a religious, apocalyptic, historicist (teleological) program based on racial or ethnic purity. Right Hegelians historically formed temporary alliances with conservative religious groups, primarily pre-Vatican 2 Catholics and the Orthodox and Russian Orthodox Churches, as they did in fascist Spain, Italy, Germany, and Yugoslavia.
- According to Wikipedia, "Right-wing_politics", the US Dept of Homeland Security defines "right-wing" as being either hate-based, or preferring conflicts to be resolved at the state level rather than at the federal level--IMHO a shockingly biased definition which attempts to associate local political autonomy with racial hate.
None of these meanings of "right-wing" correlate positively with conservatism. Conservatism today in the US means Enlightenment liberalism: supporting the continuation of free markets, free speech, mistake theory, and liberal democracy. Both left and right wings have always been resolutely anti-liberal and pro-conflict-theory.
The French nobility were conservative in their time, but would be radical today; and the centrally-planned monarchy they believed in was much closer to communism than to an individualistic, free-market democracy. Fascists were never conservative. They sought to create a new society, using the fantasy of an imaginary distant past as propaganda, just as Marxists do.
The Social Justice movement today can't be classified as either left or right. It has strong ties to Marxism through the 1960s student movements and the resulting post-1990 aging-hippy university faculties; and to fascism through Nietzsche, continental phenomenology and existentialism, post-modernist philosophers who studied under Nazi professors (e.g., Michel Foucault and Herbert Marcuse), and the associated resurgence of Plato (who literally *invented* racial essentialism, for use in creating his perfect State). It views culture through the Nazi lens of racial essentialism, assuming that culture is a manifestation of a racial essence and therefore culture maps 1-1 onto race. This is why it calls cultural assimilation racially oppressive (e.g., blacks who live in the suburbs are being oppressed by a false racial consciousness).
The social division in the US today is between urban and rural culture. Rural culture doesn't align with nationalism; it correlates with nationalism only because the dominant ideology of urban culture today is actively anti-American, and saying you're proud of America can get you Twitter-mobbed. Rural culture is more religious; yet the right-wing religious--Catholic, Orthodox, Muslim, and Hindu--all now lean Democrat (https://www.pewforum.org/religious-landscape-study/party-affiliation/). Except for Mormons, the only majority-Republican religious sects are Protestant sects formed as left-wing dissenters from the right-wing Catholic hierarchy.
[Epistemic status: I am a retail investor who's been pretty serious about it for 10 years plus moderately active on predictit/metaculus. 75% confidence]
I thought about prediction market options some more, and I think liquidity concerns make me prefer directly using leverage on the underying contracts, instead of options.
Example: Suppose JGalt wants to bet $100 each on 100 different markets. Total bet value is $10k. How much should JGalt have to deposit to cover this if the markets are all uncorrelated coin flips? He wins a mean of 50 contracts with a standard deviation of 5, so to cover 4 standard deviations of negative variance, he needs 4*5*$100=$2000 in his account, so that he is getting 5x leverage. This is conservatively assuming we let all the contracts run to completion, but the system could get some additional protection by preemptively liquidating some of JGalt's contracts if the aggregate of their market prices moves unfavorably and he gets down to <2SD of margin.
This math can be generalized to cover a set of contracts with arbitrary variance and covariance. To protect the site from people who might lose a bunch of correlated bets to get a negative balance and then not pay their debt, admins would need to estimate covariance. Each time they create a new market, they need to run a search for related markets and override the default-zero covariance appropriately. In the limit as you assume every market has 100% covariance with every other market, you get what most prediction markets do, which is to not let anyone use any margin at all (except within a single market where it's logically impossible to lose both contracts).
What could go wrong:
* Maybe admins are bad at estimating covariance, so someone is allowed to make a bunch of correlated bets that all lose, and his account goes negative, and he doesn't pay his debt.
But options based systems have all the same margin problems, *plus* liquidity problems that make it much harder to liquidate accounts that fall below an adequate margin. So I'm 75% confident that leverage on the underlying would make a better prediction market than options.
Re. weight-loss drugs, this ignores the fact that the medical establishment is morally opposed to weight-loss drugs. Most doctors feel that prescribing a fat person a pill that would make them thin, instead of teaching them good Puritan self-control, would be immoral. This is why, for perhaps the past 40 years, the foremost weight-loss pill has always been subjected to a withering attack from the medical establishment until it was effectively banned.
Today, the most-effective weight-loss pill I know of is phentermine. It was developed in the 1960s as an alternative to amphetamines. Studies done at the time showed it was non-addictive. Recent studies have also all concluded it's non-addictive. But as soon as the previous effective weight-loss pill (IIRC it was ephedrine) was banned, phentermine came under attack. If you read Wikipedia today, or for-the-public websites like WebMD, they'll say that phentermine is an addictive amphetamine, even though it is not an amphetamine and is not addictive, except under the new "my patient resists stopping this drug because it works" definition of "addictive". It's now a scheduled drug with stringent requirements even when issued by prescription, which were tightened just this year to ensure that you have to be obese to be allowed to use it, and will have to stop using it before reaching a healthy weight.
verywellmind.com says (https://www.verywellmind.com/how-long-does-adipex-stay-in-your-system-80217), "Phentermine increases levels of norepinephrine, dopamine, and serotonin, producing an effect similar to amphetamine. For this reason, phentermine is a controlled substance and only available legally via a prescription." In reality, the stimulant effects of phentermine are like those of caffeine, which also effectively increases increases levels of norepinephrine, dopamine, and serotonin. (I wrote "effectively" because this may in part be mediated by increasing dopamine receptor availability; https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4462609 .)
(I don't recommend ephedrine because it works by speeding up your metabolism, which surely shortens lifespan. I don't recommend phen-fen because apparently it really is dangerous. Phentermine enabled me to permanently lower my stable body fat from over 14% of body weight to perhaps 12%, with a total weight loss around 3%, but weight loss did not continue past that level. That was on 19mg EOD, which is 1/4 of the usual dose. You can lose more weight by combining it with Topamax, but I found that made me feel tired all day long. I eventually stopped using phentermine entirely because, like caffeine, it gives me migraine headaches.)
If the market is large enough to support statistical arbitrage, then options traffic will be reflected in the underlying. AFAIK no such liquid options markets exist today.
> I bet a lot of people would care a lot if the conservatives could produce 0.01% higher GDP growth vs. 5%.
What if we want to know how much GDP is actually driven by policy, if at all? Can we use prediction markets to get a sense of the correlation coefficient?
We might naively expect we could just look at the distribution of the predictions to determine the variance of outcomes, but I'm really nervous about that. Prediction markets combine wisdom of crowds and incentives for knowledge to chase out bad predictions. But there's no direct and obvious way for good distributional information to chase out bad distributional information on this second layer. If a bunch of crazy people are guessing randomly all over the place, and you're the smart money and you're really sure the correlational coefficient is much tighter, there's no additional incentive for fixing the shape of the graph after you've committed the appropriate amount for where the outcome will land. If the smart money has already gotten in, there's no payout for you to go in and tweak the shape of the graph.
So I think we'd need a new tool on top of prediction markets to tease this out and I'm not sure what it would be.
If you had an infinite number of players, then maybe you could have a bunch of markets with different precisions. Predict the impact on next year's GDP to the trillion, to the hundred billion, down to the individual dollar if you like. Nobody's going to take the individual dollar bet. Nobody's going to bet in the tens of trillions because it will be a fixed market. The market with the most action might tell us where the biggest area of debate is that is plausible to predict, which might tell us something about the likely variance in outcomes?
Aside from being utterly impractical, that might be an interesting direction to explore. :D
Anyone have any other market technology to help predict how tightly A causes B? Maybe this is a solved problem already and I'm just missing it.
Re: Options Markets.
Options Markets today are hedged. When you buy an options contract a market maker hedges their risk by buying (some of) the underlying contact, this moves the market.
I predict that the UK's GDP will be higher five years after the next election if the Conservatives win than if Labour does. Why? Nothing to do with what either party would do if they win; it's just that the Conservatives are in government now, and economic growth is correlated with incumbent parties getting re-elected.
What about GDP growth starting from after the next parliament takes office? Perhaps the expected GDP growth should be higher if Labour wins than if the Conservatives do, again not because of their policies, but because expected GDP at the time of the election will be higher if the Conservatives win than if Labour does, and it could regress to the mean.
This sort of thing will always be a problem when trying to use conditional prediction markets to influence decisions that are also influenced by other things. The differences in conditional expectations don't tell you the direction of causation.