Speaker 1: What are prediction markets, you may wonder? Consider an ordinary market where people trade money for some standard item, like this metal bar. Prices in such markets tend to fluctuate up and down. And when trading and storing items is cheap, this attracts speculators who try to guess, will the price rise or fall? Because if they can guess this, they can turn these markets into money pumps. That is, if they can buy low and then sell high, or sell high and then buy low, either way they walk away with cash. Hoping for this outcome, speculators search long and hard to find information to help them guess future prices. Do prices tend to fall when it rains? Do they rise on Tuesdays? If speculators think that prices tend to rise on Tuesdays, they will buy on Mondays and sell on Wednesdays. But doing this makes prices rise less on Tuesdays, so it turns off the money pump. In general, speculators who bet on information not currently reflected in market prices tend to push that information into the prices. Thus, in active markets, prices tend to embody a lot of information about future prices. So much so, in fact, that it is usually quite hard to consistently make money by betting in these markets. This is a mild version of the so-called efficient markets hypothesis. All this works not just for markets in natural objects like metal bars, but also for markets in artificial objects, like this piece of paper, which says pays $1 if Biden wins the 2024 U.S. presidential election. The future price here is set by the outcome of the election, so if you can get enough people to trade here, then the current market price should become an accurate estimate of that future price, that is, of the chance that Biden will win the election. Now, while we've had speculative markets for many centuries, they have mostly been created for other reasons and have only aggregated information as a side effect. However, in the last few decades, people have started to make prediction markets, which are speculative markets created to harness this information effect on topics of interest. So far, we've seen at least a dozen head-to-head comparisons wherein both speculative markets and some other familiar institution, like polls or committees, estimated the same thing at the same time and using comparable resources. The usual result? Speculative markets are either substantially more accurate or about as accurate as other methods. They're almost never much worse. For example, out of 1,000 comparisons between national election polls and markets, the markets were closer to the truth about three-quarters of the time, even though market participants were not at all representative of the electorate and markets moved a lot less than polls did in response to the debates. This strong track record has inspired many organizations to try to create internal markets to answer questions of interest to them, such as whether they will make project deadlines or which new projects they should start. Prediction markets have many advantages and many other ways to get people to aggregate information into consensus estimates, such as polls or committees. Markets give numerically precise estimates instead of vague weasel words. When there are markets on many questions, their estimates will tend to be consistent across all those questions. Price estimates are frequently updated, at least if trading is frequent. Compared to other ways to create consensus estimates, prices tend to be harder for interested parties to manipulate. Markets are also robust, in contrast to other mechanisms, and can be very sensitive to the quality of the participants. For example, a committee made mostly of bozos may go quite badly, but allowing many bozos to trade in speculative markets isn't a problem, as long as a few well-informed people can also trade. Some have framed the issue here as prediction markets revealing a wisdom of crowds, with crowds supposedly better than experts, but that's not right. The idea here instead is to entice participation by whomever really knows things on each particular topic. Sometimes that will be the highly credentialed, and other times that will be more ordinary-looking folks. The main argument in favor of speculative markets is empirical. When tested, they work. But theory also supports them, as we do roughly understand why they work. In contrast to other mechanisms, speculative markets offer simple, direct incentives for accuracy. Not just incentives to choose your trades well, but even more important, incentives to not trade at all if you don't honestly think you are better informed than the others who might trade. Furthermore, traders have incentives to find and fix any biases they can find in market prices. Markets do not need individual traders to be especially rational. Prediction markets are a simple, powerful way to let a large, diverse community collect their information and merge it into precise estimates, which everyone can then use to make their decisions. Prediction markets may look simple, and many of them are. But we can do a lot more with them once we learn a few tricks. In the simplest prediction markets, one trades dollars for assets of the form pays $1 if A, where A is some claim. The asset ratio in these trades is the market price, which can be treated as an estimate of the probability of A. Now if we use a conjunct A and B, where B is another claim instead of A, we can get a probability for that conjunct. And if users trade the A and B conjunct asset for the A asset, then that price should be the ratio of the other two prices, which is the conditional probability of A given B. Thus, we can make markets in conditional probabilities. Instead of starting with an asset that pays $1 if some discrete event A, we can use an asset that pays in proportion to some continuous variable x that is restricted to lie in the range 0 to 1. In this case, prices can give us expected values for x and conditional expected values for x. In an ordinary market, traders must either post an offer and then wait for someone to accept it or accept an offer that someone else has posted. When traders are rare, this can lead to long delays. A solution is to create an automated market maker who always has two posted offers, one to buy and one to sell with a small price difference between them and a simple rule for how prices move in response to trades. With an automated market maker, any trader can always make a trade fast even when other traders are rare. In an ordinary prediction market, trades are zero-sum. Whatever one side gains, the other loses. Theory says that rational risk-averse traders without insurance needs would not trade in such markets. Yes, humans do often trade in such situations, but what if they get wise? Here are two ways to subsidize these markets to make sure that all traders can gain. First, we can set the parameters of an automated market maker so that it loses money on average as the price moves toward its final resting place and set a limit on how much it would ever lose. Such a market maker could be funded by the parties who want the information that the market prices can reveal. Alternatively, a subsidy could come from manipulative traders, that is, traders who are willing to lose some money in order to distort the consensus price that everyone else will see. Not only do such manipulative traders subsidize other traders, but on average their presence increases price accuracy. While manipulation can be a serious problem for other institutions, it just isn't a problem for speculative markets. Finally, imagine a simple prediction market on whether a project will make its deadline. We might worry that project personnel will sabotage the project in order to win some bets. A simple fix here is to start everyone who could harm the project with a positive stake in that project and then prevent them from letting their stake go negative. Now traders can reveal information via their trades and yet have no incentive to sabotage. By using automated market makers, we can create an edit-based interface wherein users think in terms of editing a probability distribution instead of trading. For example, a user might see this distribution, grab it at this point, and then drag it to a new point. Their trading assets would change automatically. The next user would then see this new curve and could reverse this edit if they desired. We could also let users edit a probability distribution over a large tree of possibilities. For example, here is a tree of physics topics for the next Nobel Prize. Using this tree, specialists might dive deep and edit the very fine-grained distinctions while the rest of us might stick with editing the broadest categories. In a combinatorial prediction market, users can bet on any of a very large set of possible combinations of some basic set of questions. For example, consider this edit-based interface for a firm that will soon sell some widget, a firm with two competitors, A and B. This screen shows current estimates for a wide range of parameters for all these firms' products. A user could pick any one of these numbers and see a more detailed probability distribution about it. A user could also pick any number and change that number if they were willing to make the corresponding bet. In addition to changing any number, a user could assume any number in their interface by setting that number to some assumed value. After doing that, the entire rest of their screen would then show estimates conditional on this assumption. The user might even make a second assumption and then all the numbers would change again. They might also see some indication, like this red scenario probability number, of how unlikely is the situation that they have drilled down into. While viewing conditional estimates, a user could still edit any number, but now these edits are conditional on the assumptions made. If these assumptions turn out false, it will be as if these edits never happened. A few years ago, some colleagues and I created a conditional prediction market wherein all the probabilities were represented by a Bayesian network. Users could edit any number via radio buttons or by moving a slider, and user changes would then be propagated through the rest of the Bayesian network using Bayes' rule to exactly update both the probabilities shared by all users and a representation of this user's assets. We made sure that no one could ever make a change requiring a bet that they couldn't pay for in all possible scenarios. Here is a map of our Bayesian network at one point. Users could not only edit marginal and conditional probabilities, they could also make changes to the network structure, and all this was computed efficiently. While most all real prediction markets today are quite simple, I hope you can see now that the field of prediction markets contains a lot of possibilities. A variation on prediction markets can directly inform decisions and be the basis of a new form of government that I'm going to call futarky. Consider this diagram. The two arrows on the left represent two markets wherein users can bet on the claim S and not S where S says whether some particular region adopts a single-player medical system. Those markets estimate the probabilities of S and not S. The two arrows on the right represent markets which bet on the claim L, which says whether lifespans will rise in this region. But these right-side markets bet on L conditional on S and thus give conditional chances of increased lifespan if and if not a single-payer plan is adopted. If these two estimates differ, that difference offers direct decision advice on which single-player choice is more likely to increase lifespans. I call this sort of setup a decision market, and it has many promising applications. All we need is a discrete decision that we expect to make and some outcome that we hope that decision may influence. Then speculators can tell us which decision option gives the most of that outcome. Note that we don't ever need to know if this decision caused this outcome. It is enough to know speculators' prior conditional expectations. Here's a real example. Early on in the 2016 U.S. presidential election, there were betting markets in many candidates, markets that gave the chance for each candidate both of becoming their party's nominee and of becoming president. The ratio between these gives the conditional chance of becoming president if nominated, and this gives direct advice to these parties on who to nominate if they want to win the election. As another example, at one point, Bitcoin was considering making a change to its block size, and a decision market was created in our Sycast system on if this would increase the total value of all Bitcoins. Alas, I have no idea how much this influenced their actual choice. Here is how I would spend a million dollars if I had it. Please give it to me. In an ordinary stock market, one trades cash for stock. To estimate the stock price, traders try to think of all the scenarios that the firm might later end up in and ask how much revenue the company would earn in each scenario and then weigh them all together. For each firm in the Fortune 500, I would create two new conditional stock markets. These markets would also trade stock for cash, but these trades are called off if or if not the CEO stays in power until the end of the current quarter. When trading in these markets, traders should still ask themselves how much revenue the company will earn in each scenario, but now they will only weigh the scenarios consistent with each market's condition. The difference between the prices in these two markets offers direct advice to the firm's board of directors. Does the CEO add value to this firm or should they be replaced? After a few years of running these markets, we could compare the average financial return of firms that followed this market advice to the firms who did not, and then maybe shame or sue boards of directors into following such market advice. If firing CEOs isn't big enough to excite you, let me raise the stakes. Let me say how we could use speculative markets to run the government. I call this futarky. I do this to try to inspire you with just how far this idea could go, but I'm not proposing that we do this immediately. Instead, I think we should try small-scale trials of this sort of mechanisms and then work our way up to bigger trials as the smaller ones seem to work. Okay, here's a simple argument for futarky. First, in the long run, it is not actually that hard to tell rich, happy nations from poor, miserable ones. If you travel around the world today, you can roughly see. Second, governments largely fail by adopting bad policies, which they adopt because they don't listen to the people who did or could know that these are bad policies, which were more likely to lead to poor, miserable outcomes. Third, speculative markets are a best-known way to get everyone to tell us what they do or could know. Put this all together, we should put speculative markets in charge of picking policies. Here's a diagram of our current form of government. Voters at the bottom elect representatives who form a legislature that passes bills, bills that become law if courts approve their constitutionality. These voters and legislators are informed by a variety of information institutions, including academia, news media, and organized interest groups. Now here is a diagram of my alternative. On the left side, we use the same structure as today to vote on values. An elected legislature passes bills, but now these bills only define national welfare and oversee an agency like our Bureau of Labor Statistics to measure this number. National welfare is somewhat like GDP, but may include more kinds of statistics. For every month, the agency declares a precise number saying what our welfare was for that month, and national welfare is defined as some weighted average over these future months. On the right side of this diagram, we have decision markets where speculators decide if they expect each particular bill to increase national welfare. If they so expect and if courts approve, the bill becomes law. Every day, there are several time slots for considering a new bill, and there's even an auction to pick which bill is considered in each slot. So now we vote to decide what outcomes we want, but we trust market speculators to tell us which policies are most likely to get us what we want. Yes, there are some complications, some of which are addressed here. For example, there might be errors in the national welfare definition, errors that are only revealed when we see particular proposed bills. To prevent us from taking immediate and harmful action based on such errors, I've included a way that speculators can veto bills today based on a future-defined national welfare. Now, I know from long experience that as you've been listening, most of you have collected a few concerns in your mind about my proposal, and you won't be comfortable until those are addressed. Some of your concerns may be listed here. And some of them may be listed here. I think I've heard them all and have decent responses to most, but you won't know that until you hear them, which is why my main academic paper on this topic is mostly filled with responses to 25 different objections. Let me conclude by admitting what I see as futarki's biggest problem. It is bad at hypocrisy. Today, voters can pretend that they care a lot about trees, for example, but then consistently elect politicians who know that voters don't really mean it and would punish politicians who promoted trees too much. If low support for trees in policy were ever clearly exposed, voters could declare outrage and then blame corrupt politicians. But under futarki, we'd have to declare our tree priorities more clearly in our national welfare definition. So we'd either have to admit that we don't really like trees that much or continue the pretense at a much higher cost, as futarki would be effective at giving us what we say we want, even if that isn't what we really want.