Prediction markets, sometimes called information markets, idea futures, or virtual stock markets, are elaborations of this simple device for identifying sincere belief. By insisting that individuals back their views with money, they eliminate cheap talk. At bottom, the simplest form of prediction market is a forum for repeated challenges of this type. Someone who thinks that an event has more than a 75 percent chance of occurring offers to place three quarters against the single quarter of any challenger. Someone else who thinks that an event has less than a 25 percent chance of occurring would have a financial incentive to take that bet. The acceptance of such a bet indicates the apparent existence of genuine disagreement.

Still more useful than such bets in conveying probability estimates to the outside world are the challenges that go unanswered. Suppose that no one responds to the challenge described above. Meanwhile, someone who thinks that the event has a less than 80 percent chance of occurring offers to place $0.20 on the event’s not occurring against anyone else’s $0.80. No one responds to this offer, either, not even the first challenger. There are then two unanswered challenges. No one is sufficiently confident that the event has a less than 75 percent chance or a more than 80 percent chance of occurring to take either of the bets. One could infer no one sincerely believes that the probability is outside the 75 percent to 80 percent range. At the least, anyone who initially would have guessed the probability to be outside that range does not think it sufficiently clearly outside that range to be worth accepting a challenge. The actual probability, based on information known to potential bettors, thus may be between 75 percent and 80 percent. A probability of 77.5 percent might be a good guess, if a precise number is required.

Prediction markets do not merely show that those who make bets are sincere but identify the most committed predictors in a group. At any time, these are the predictors who have made forecasts that have not yet been challenged. Admittedly, the most committed challenger will not always be the most sincere, but at least he or she will appear to others to be the most sincere. For if it is obvious that a challenge is bluster, perhaps intended to influence a decision or simply to make a point, others can be expected to counter the challenge. This is especially so for prediction markets in which anyone is permitted to participate. If, after repeated challenges, no one agrees to accept a bet and meet a challenge, then even those who initially had different views must have developed sufficient doubt that they no longer think it in their interest to bet. In part, prediction market participants revise their views on the basis of the willingness of others to advance contrary positions.

In a recent book titled The Wisdom of Crowds James Surowiecki identifies a variety of contexts, including prediction markets, in which crowds perform better than their best members.1 The strongest case for a prediction market, however, is not that it simply aggregates knowledge by averaging beliefs. Prediction markets do tend to allow individuals within a crowd to learn about the beliefs of others, but this will not always be important. There are some contexts in which the errors of individual intuition tend to cancel each other out, but there are others in which errors are correlated and the masses are wrong. What a prediction market excels at is identifying the wisdom in crowds, identifying the individual challenger (or group) that is so committed to a particular view that no one remains willing to accept a challenge. This will not always be the individual who in fact has the most wisdom and knowledge. But the procedures of prediction markets generally do a better job of identifying that person than do the alternatives.

This book argues that organizations in the private and the public sectors ought to use this procedure, as well as more sophisticated procedures building on this basic design. Prediction markets can serve as general-purpose prediction mechanisms, and although one can debate how they best should be designed, they generally are superior to alternative procedures for generating predictions not involving direct financial incentives. A counterargument is that someone who learns of the results of a prediction market might still be able to improve on the prediction by scrutinizing the evidence directly, and so institutions, too, ought to make a prediction market forecast just one of many decision-making inputs. Someone who does independent research, however, could participate in a market and change its forecasts in that way. A refusal to do so suggests lack of sincerity, especially if the market is heavily subsidized to make predictions attractive.

Even if there is a plausible alternative explanation for such a refusal–perhaps someone with information has a low tolerance for risk–there are reasons to prefer market predictions to those of individuals. In general, individuals who are asked to make predictions as part of a decision-making process have many conflicting incentives. One incentive might be to preserve one’s reputation by announcing predictions that turn out to be correct, but other incentives might skew the announcement of a prediction. The predictor may alter a prediction in order to make a desired consequence more likely, modify a genuine prediction to please a third party such as a boss or a friend, or make a prediction on the basis of hopes and fears rather than hard analysis.

A prediction market forces someone making a prediction to bear the financial consequences of allowing extrinsic factors such as these to influence a prediction. What is more, it gives third parties incentives to assess the sincerity of the predictor, bringing to bear all their knowledge about the incentives that might be driving that individual. One might still argue that some individuals can be trusted to make accurate predictions yet somehow cannot be induced to participate in a prediction market, and so we might obtain better predictions by relying on these individuals. But developing reliable mechanisms for selecting them is not easy. And even if we could do so, we could induce these individuals to announce predictions and then allow third parties to take these predictions into account in their own prediction market bets. To prefer predictions that are untested by the competition of a market is to believe that only by granting power to designated decision makers can we achieve uncorrupted analysis.

A prediction market, then, is a decision-making device for excluding from consideration predictions that no one appears willing to back with money. Only with the development of the Internet has creation of prediction markets become economical, and regulatory obstacles have slowed their development. Nonetheless, as this book documents, prediction markets are becoming increasingly common. Preliminary empirical evidence confirms that they generally make accurate predictions, and if later research confirms this, there is every reason to believe that they will gradually replace alternative means of making predictions. Moreover, prediction markets are sufficiently cheap that they can be used to make predictions that are not now made. Many decisions are made without explicit quantification of various relevant projections. Prediction markets can improve decision making by creating a useful procedure for making those projections explicit.

It might appear that prediction (including the implicit kind) is but a fringe element of public and private decision making. This book argues, however, that prediction is central. This is perhaps more obvious in a private entity such as a corporation, where many decisions are aimed at achieving concrete goals, most notably (though not exclusively) making money. Public decision making, in contrast, involves a balancing of various private rights and interests, rather than straightforward maximization of specified variables. Even so, representative government itself can be viewed as an implicitly predictive exercise. We delegate power to legislatures, agencies, and other institutions largely because we believe that the decisions that they reach will accord more closely than alternatives with the decisions that the citizenry as a whole would reach, if only the citizenry had the time to become fully informed about particular issues.

We know, however, that the impossibility of all voters’ becoming fully informed on every issue means that the democratic process often fails to achieve this modest goal of aggregating hypothetically informed preferences. Sometimes it fails because relatively uninformed people are given too much of a voice. Policy becomes the voice of the mob. At other times it fails because the public’s lack of information means that appointed decision makers are able to advance their own interests at the expense of the public’s. Policy becomes the voice of special interests. Governmental design can be seen in large part as reflecting attempts to balance these and other problems, to steer decision making to the institutions and individuals who seem likely to come closest to reflecting a hypothetical majority sentiment.

More populist institutions, such as juries and referenda, decrease the risk that decision makers will act against the will of the people, but they are costly and increase the risk that decisions will not reflect the best available information and analysis. This helps explain and justify the use of legislative committees, administrative agencies, and courts. By channeling decisions to a relatively small number of individuals with a comparatively large amount of expertise, we ensure that decisions are made by those in the best position to assess the consequences of those decisions. The law of small numbers, however, means that their decisions are not always representative. Institutional design faces a fundamental trade-off between representativeness and expertise.

The project of government and political science has proceeded on the reasonable assumption that there is no objective way of ascertaining what the people would decide if only they were fully informed. The technology of prediction markets changes this assumption. Prediction markets can be used to predict, for example, the probability that a randomly selected person, after having an opportunity to consider an issue fully, will conclude that a proposed policy would be beneficial. If market participants do not know who will be selected, they will try to divine what an average decision maker would conclude. The market will help to identify the most sincere predictor, likely to be someone who has a great deal of information about both the underlying issue and the underlying population from which the randomly selected individual will come.

Prediction markets do not eliminate the tradeoff between representativeness and expertise altogether; we must still decide, for example, whether to pick a person at random from the population as a whole, or from a smaller body such as the legislature or the judiciary. But relative to alternatives, they enable decisions that are both relatively representative and relatively informed. We will also see that prediction markets can enable more principled decisions. Someone announcing how that person would resolve an issue may be more inclined to consider higher-order principles, such as fairness to minorities and consistency with precedent, when the decision will merely resolve a prediction market, in the sense of determining payouts, than when the decision will actually affect policy. Prediction market evaluations of policy alternatives can thus be faithful to a society’s highest ideals.

My aim, of course, is not to suggest that we should discard institutions developed over hundreds of years in favor of unproven alternatives. We should not. But nor should we close our minds to the possibilities of democratic institutions structured in radically different ways from the ones that we now have. If these alternative institutions have advantages, then they can provide useful contributions in particular contexts. This book seeks to describe some relatively modest possible reforms using prediction markets, but it also envisions the most radical possible reforms. Only by comparing the most central of our existing institutions to prediction market alternatives can we identify most clearly the strengths and limitations of prediction markets. For today, this can and should be only an academic exercise; many of the ideas developed are theoretical and untested, although the book also discusses empirical evidence of existing prediction market designs. Much more experimentation would be necessary to refine these designs.

We cannot identify alternatives to complex institutions merely by repeated use of playground challenges. This book will show, however, how bets can serve as a foundational alternative to votes in the construction of institutions. For example, a principal goal of many institutions is to encourage deliberation that may sway the votes of, for example, jurors or senators. We will see how prediction markets can similarly encourage deliberation by requiring individuals to bet on whether they will be able to persuade others of their views. Similarly, we will see how combined series of bets can be used to produce not merely numeric predictions but also consensus “texts,” much as governmental procedures lead to the creation of statutes, regulations, and judicial opinions. We will also see how prediction markets can be combined into a kind of “web” so that every assumption and calculation that underlies a prediction or policy evaluation can be explicitly identified and challenged.

The prediction market designs described in this book build on one another, with more complex designs generally introduced later in the book than more basic designs. Each chapter focuses primarily on a different institution or group of individuals that could benefit from prediction markets, with hypothetical prediction markets replacing the central institutions of government in the later chapters. The final chapter imagines predictocracy, a form of government in which prediction markets serve as the foundation for all decision making. My intent is not to endorse this form of government, and I occasionally note the dangers associated with abandoning the status quo. Prediction markets offer the possibility of a new way of thinking about structuring decision making, and this approach can be used for problems large and small.

Why might markets for ideas fail? A simple answer is that many members of the public will generally be ignorant about particular issues and underlying theoretical frameworks, often rationally so.4 Acquisition of information requires time and money. Just as economic markets can lead to suboptimal results as because of imperfect information,5 so, too, can limited background knowledge make accurate evaluation of new claims about the world extraordinarily difficult. So it is not surprising that surveys indicate that the public’s views about issues often differ from the views of those who might be considered experts. For example, survey evidence suggests that economists generally believe that trade agreements between the United States and other countries have helped create more jobs in the United States, whereas the public generally believes that such agreements have cost American jobs.6 The direction of this discrepancy should not be surprising. Stories about people losing jobs are more engaging than stories about people gaining jobs from trade, and so some doses of media may make people more ignorant than they would be in the absence of any media.7

The media, of course, cannot be expected to fix all cognitive errors among the public. When the media accurately report information, the public may pay little attention. For example, although the media presumably accurately report the identities of congressional candidates, in many elections the vast majority of voters cannot name a single congressional candidate in their district,8 let alone give information about voting patterns. And if the media could succeed in informing the public with facts of this type, they cannot hope to replace the educational system in providing background knowledge and skills necessary for making relatively informed voting decisions. All is not lost, and some political scientists argue that voters still are able to make rational decisions.9 Nonetheless, it should not be too idealistic to hope that the media, though constrained by the need to entertain, would generally seek to improve the knowledge of the viewing and reading public.

The media, however, may be hesitant to provide information about the consensus views of experts concerning particular issues because there may be controversy about what the consensus is. Ironically, a desire to promote objectivity sometimes may interfere with attempts to provide accurate information. When an issue is contested, a media outlet perhaps can best achieve the goal of objectivity by presenting both sides of the issue. The public, however, often cannot identify which argument is more true and compelling. Suppose, for example, that a news program about crime features one expert who contends that crime rates in a particular city will soon rise and another who contends that crime rates will soon fall. Even a criminologist or a statistician might need to spend hours evaluating the literature to assess the relative strength of the competing arguments. The public might benefit from a statement by the news program that three-quarters of the experts contacted subscribed to one view. Understandably, however, journalists are reluctant to report the results of such unscientific polls, yet lack the resources to conduct more sophisticated polls of experts, assuming that they could find an objective definition of who should count as an expert.

With regard to some issues, “just the facts” may be all the public needs, but in many cases an important fact may be what the consensus opinion about the issue is or whether such a consensus exists. This chapter suggests that prediction markets can provide objective gauges of expert consensus for the media to pass along to the public. My claim is not that prediction markets provide the best possible predictions, overcoming all irrationality. Indeed, individual experts sometimes might be able to beat markets. But experience suggests that prediction markets are generally fairly accurate, and at least they provide incentives for those who could beat the market to push predictions in a sensible direction.

I begin by describing a relatively simple structure for a prediction market that can be used to estimate the probability that a particular event will occur. After recounting the experiences of perhaps the most famous prediction market, the nonprofit Iowa Electronic Markets (IEM), I describe a more general type of numeric prediction market that the IEM also uses. Although it has focused primarily on predicting the outcomes of elections, prediction markets conceivably can be used to predict virtually any outcome, and the for-profit Web site Tradesports.com has used them to predict the results of athletic contests and a range of other events. A prediction market is not the only possible prediction technology–pari-mutuel wagering has long been shown to produce predictions about horse racing–but this chapter argues that prediction markets offer considerable advantages. They also present unique challenges, and I assess the danger that they might be manipulated. Finally, I suggest how the media might use prediction markets to provide better information to readers and viewers.

Consider, for example, a subject that recently was of great import: whether the actress Angelina Jolie and the actor Brad Pitt will marry. Suppose that Us Weekly wished to provide its readers with the most accurate information possible about the probability that a marriage would result by January 1, 2010. Of course, Us Weekly presumably wants to present its readers with a range of relevant information, such as reports of supposed friends and analyses by body language experts who scrutinize photographs of celebrities as they seek to escape from the paparazzi. But it might also want to provide a concrete number, an estimate of those in the know of the chance that the marriage really will happen. This would be useful both for readers in a hurry and for readers who are unsure of their own ability to weigh the different pieces of evidence. Of course, Us Weekly might simply make up a number, but such numbers might seem arbitrary or sensationalistic. If it wanted to produce a more credible estimate, it might launch a probability estimate prediction market.

Here is how such a prediction market might work, placing aside for now concerns about whether creation of such a market in the United States would be legal (see Chapter 2): Us Weekly would sell two different kinds of tradable contract to the public, a marriage tradable contract and a no-marriage tradable contract. For example, it might hold an on-line auction for one hundred of each type of share. It would promise to issue a fixed payoff, say, one dollar, on January 1, 2010, depending on whether Jolie and Pitt in fact become married by that date. So, should they become the Jolie-Pitts by that date, each marriage tradable contract would be redeemable for one dollar on that date, but the no-marriage tradable contract would be worthless. On the other hand, should Jolie and Pitt not be married to each other by that date, then the no-marriage tradable contract would be redeemable for one dollar on that date, but the marriage tradable contract would be worthless. In announcing the prediction market, Us Weekly presumably would seek to limit the possibility of ambiguities, such as whether a marriage followed by a divorce would count. But if there were ambiguities, then Us Weekly would resolve them either after or before the payoff date.

The initial auction of such tradable contracts itself would produce some information that could help assess the tradable contracts. For example, suppose that the marriage tradable contracts sold for an average of about sixty cents, and the no-marriage tradable contracts sold for an average of about thirty cents. That would seem to indicate that the marriage share is considerably more likely than the no-marriage share to be redeemed for one dollar. Hazarding a specific probability from such information would be difficult, however. Perhaps the purchasers of the marriage tradable contracts are sentimentalists, willing to indulge their romantic notions for a collective total of sixty dollars. This is not a fully satisfactory answer, for it does not explain why the no-marriage tradable contracts did not sell for more. After all, if marriage seemed unlikely, at thirty cents each, the no-marriage shares would have been a bargain, and someone should have had an incentive to bid more for them. But given the inherent uncertainty of celebrity love lives, playing in the market carries some risk and demands some transactions costs, which include at least the time it takes to purchase and redeem shares. These factors can explain why the combined auction prices might be less than one dollar. Meanwhile, the possibility that individuals might obtain pleasure from holding tradable contracts independent of their financial value will tend to push prices up.

Whether the initial tradable contracts are distributed by auction or by some other means, the key to the prediction market is that holders of tradable contracts can sell them. Typically, an on-line exchange serves to facilitate such transactions. Someone interested in purchasing the marriage tradable contract could submit a “bid price,” the price the potential purchaser would be willing to pay for that tradable contract. Someone interested in selling it could submit an “ask price,” the price at which an owner of the tradable contract would be willing to sell it. Whenever a bid or an ask is submitted, the exchange would seek to pair the offer with another. For example, if I offer to sell a marriage share for sixty cents, and then you offer to buy a marriage share for up to sixty-five cents, the on-line exchange could then complete a transaction, presumably at sixty cents, since the offer to sell came first. When an offer does not immediately find a match, it is placed in a queue, with the best offers placed at the front of the queue. At any given time, the on-line exchange maintains a “bid queue” and an “ask queue.” At the front of the bid queue is the most generous offer to buy, and at the front of the ask queue is the most generous offer to sell. The ask price will always be greater than the bid price, because when a new offer matches an existing one, a transaction is immediately completed. In economics, this arrangement is known as a continuous double auction.10

Just as the auction prices provide some indication of the probability of the Jolie-Pitt union, so, too, can the bid and ask prices provide some hint. For example, suppose that the bid price is fifty-five cents and the ask price is sixty cents. That would indicate that no one is willing to sell the tradable contract for less than sixty cents but that there are people willing to buy the tradable contract for fifty-five cents. If the event were viewed as having a significantly more than 60 percent chance of occurring, then one would think that someone would be eager to purchase the tradable contract for more than sixty cents. If the event were viewed as having a significantly less than 55 percent chance of occurring, then one would think that someone would be willing to sell for less than fifty-five cents. The midpoint of the bid and ask prices might thus provide at least a rough probability estimate, and this estimate may be averaged with the corresponding estimate from the no-marriage tradable contract. An alternative approach to deriving a probability estimate would be to base it on the most recent transaction or possibly on an average of several recent transactions. Either way, a significant advantage of a prediction market is that it will produce data that change over time as new information becomes available.

This reflects a key aspect of prediction markets. Prices do not simply reflect an average assessment by a group but also reflect the degree of confidence that different members of the group have in their estimates.11 Simply taking the average estimate of a group may work well for some problems but not for others. For example, Francis Galston studied a competition in which contestants guessed the weight of an ox; the average guess of the 787 contestants, 1,197 pounds, was only one pound short of the actual weight.12 But when Cass Sunstein, a law professor, asked his colleagues to estimate the weight of the fuel that powers space shuttles, they gave a median answer of two hundred thousand pounds, far short of the actual answer of four million pounds.13 People may on average systematically misestimate certain numbers, and even if a few people in a group know the correct answer, those people may have only a small impact on the average.

Because no one is forced to participate in a prediction market, those who participate tend to be those who have information relevant to the particular prediction or at least those who can obtain the information at low cost. Among participants, individuals who have the most information should be willing to place the most money at risk. It will not always work out this way, of course. Sometimes, someone who has relatively little information or erroneous information might nonetheless be bold and wager a great deal of money on a particular position. For example, someone might invest heavily after overhearing a conversation in which Pitt refers to Jolie as his wife, when others who heard the conversation realized that Pitt and Jolie were merely discussing their roles in the movie Mr. and Mrs. Smith. Other market participants might surmise wrongly that this trader has some valuable information and change their own initial probability assessments on the basis of this individual’s trading.

Meanwhile, someone with excellent information will face some constraints on liquidity. For example, suppose a friend learned that Pitt and Jolie had agreed that they would flip a coin to determine whether to marry. Assuming the reliability of this information and the fairness of the coin, the friend could be sure that the correct probability is 0.5. This friend would probably want to purchase some no-marriage shares for thirty cents, and gradually that would move the price toward fifty cents. But unless the friend can credibly reveal the information to the world and liquidate the resulting financial position, this strategy is not guaranteed to produce profit. A purchase of the no-marriage share for thirty cents produces a 50 percent chance of a seventy-cent profit and a 50 percent chance of a thirty-cent loss. A risk-averse individual might take this deal but decide to stop buying once the no-marriage share reached forty-five cents. Willingness to invest depends both on a trader’s assessment of the quality of the trader’s information and on the trader’s liquidity, so mistaken self-assessments and liquidity constraints can lead market prices astray.

The case for a probability estimate prediction market thus cannot be that it will somehow produce perfect information. Such a market cannot tell us for sure whether Pitt and Jolie will marry. And we cannot even be sure that the probability estimate that the market produces will be the best one possible based on existing information. Theory suggests, however, that a probability estimate prediction market can serve as a relatively simple technology for aggregating individual probability assessments. Self-assessments of information quality seem likely to be at least correlated with actual information quality, and so a prediction market in effect provides a mechanism for weighing the estimates of a group of individuals based on the information that members of the group possess. The financial incentives of prediction markets ensure that forecasts will reflect genuinely held beliefs. Prediction markets reflect the intuition that when someone puts his money where his mouth is, he has greater credibility than when he does not.

The question remains: How good a technology for producing probability estimates is the probability estimate prediction market? There are, after all, competing approaches. One could seek to identify a group of individuals who might have relevant information or who are experts in the field and survey them. Perhaps a few phone calls to friends of Pitt and Jolie would produce more accurate estimates. Of course, a prediction market participant might make such calls, but there is no guarantee that the market prediction will take this information into account to the optimal degree. Even if prediction markets are superior to alternatives such as surveying experts, another prediction market design might be preferable to the probability estimate prediction market. If Us Weekly ever were to accept such markets, they would likely need to become far more commonplace, but for this to occur, there must be empirical evidence that the probability estimates that they produce are accurate.

Generating reliable evidence about the accuracy of probability estimate prediction markets, however, is not easy. The occurrence of an event cannot show that a probability estimate of that event was correct or incorrect. Suppose, for example, that the Pitt-Jolie market predicts with 99 percent confidence that they will marry. If they do marry, that provides some reassurance, but perhaps the result is a mere coincidence. And if Pitt and Jolie do not get married, that might appear to cast substantial doubt on the market. But it could be that the 99 percent estimate was reasonable based on information available at the time, and the l 1 percent possibility has come to pass. Probability estimate prediction markets are not able to miraculously anticipate that events that would seem unlikely to anyone with the relevant information in fact will come to pass. The best hope for them is that their probability estimates will be better than alternatives’. We can only reliably gauge the accuracy of probability estimate prediction markets after much experience with them.

Figure 1.1 illustrates the prices at which contracts last traded in the market at the end of each trading day.16 It is clear that the prices of the Bush and the Kerry shares are virtual (though not perfect) mirror images of each other. This alone is not enough to validate the results, of course, but if the Bush and the Kerry shares appeared to move inconsistently, that might furnish an argument that prediction market predictions appear irrational or random. The price lines also appear rational when compared against actual events in the election campaign. Although they cover only four months of the election cycle, the price lines appear at least roughly to correspond to the candidates’ chances of victory as indicated by the polls. Figure 1.2 shows Bush’s polling share (excluding undecided or third-party voters) for all major national polls.17 The race appeared to be quite close until Bush began to take a lead at about the beginning of September, possibly in part as a result of aggressive advertising by a group called Swift Boat Veterans for Truth that questioned Kerry’s record in the Vietnam War, though this lead dwindled through mid-October. The spikes in figure 1.1 are greater than the spikes in figure 1.2 because a small difference in poll numbers can result in a large difference in probabilities. Had Bush been leading Kerry 60 percent to 40 percent in the polls, after all, his victory would have been virtually certain on Election Day. Electoral trends are somewhat easier to deduce with the prediction market forecasts of figure 1.1 than with the polls of figure 1.2, because trends are accentuated and because there is less noise from anomalous polling results.

A better assessment of overall predictive accuracy can be obtained by comparing predictions with results across a range of markets. Figure 1.3 reports results from twenty-five winner-take-all markets sponsored by the Iowa Electronic Markets. Wolfers and Zitzewitz collected data about the last trade each day for each tradable contract in each of these markets, producing a total of more than twenty-three thousand observations.18 For each one-percentage-point market price interval, they then counted the proportion of times the event being predicted in fact occurred. For example, if prediction markets are accurate, then one would expect contracts priced at about fifty cents to result in payoffs about 50 percent of the time and contracts priced at about eighty cents to result in payoffs about 80 percent of the time. The data appear to reflect this expectation. Higher contract prices appear to bear an approximately linear relationship to probabilities. For comparison purposes, figure 1.3 includes a forty-five-degree line that would represent perfect prediction accuracy.

Figure 1.3 does show a variety of anomalies. But this is because in a sample of only twenty-five elections, a few unexpected election outcomes or dramatic shifts mid-election can affect a relatively large number of data points. Figure 1.4 thus shifts from the Iowa Electronic Markets to TradeSports, a for-profit exchange that includes trading on a wide variety of events, including athletic and political contests. For example, for a particular baseball game, TradeSports might include contracts concerning whether a particular team will win, whether the total score will exceed a particular number, and whether a particular team will win by at least a specified number of points. It thus serves as a market alternative to more traditional forms of sports betting. Figure 1.4 reflects 145,388 trades on a total of 1,508 contracts determined by the outcome of Major League Baseball games that TradeSports decided to feature in 2005. With this larger sample, the data appear to reflect more consistently the forty-five-degree line that one would expect if prices can be interpreted as probabilities.19

Two significant caveats are worth making. First, we cannot guarantee that the regularities observed in this case will be perfectly replicated in other markets. In an analysis of 384,655 trades on National Football League games over the course three seasons, Richard Borghesi found some small but systematic anomalies.20 Borghesi found that on average, prices decreased from early values by approximately 3.42 ticks (34.2¢ on a $10 contract), indicating that bettors generally are overly optimistic about the probability that the contract will pay off, that is, that the named team will win. By combining TradeSports data with play-by-play football data, Borghesi shows that the market tends to underreact to information. In the minute following a touchdown, the market continues the rise caused by the touchdown itself, and, indeed, the trend continues for another nine minutes, with additional average increases in price for a touchdown for the named team of 0.14 ticks. Although these findings are statistically significant, they imply only small deviations from accurate probability estimates, so prices still can be said at least roughly to reflect probabilities. One possible explanation for the poor performance of these markets relative to the markets illustrated in figure 1.4 is that the figure includes only “featured games,” which may have higher liquidity and receive more careful attention from traders.

Second, that the prices of contracts reflect probabilities does not mean that the prediction market is particularly accurate. If for each of a large number of athletic contests I made a trade at fifty cents on a randomly selected team, then I would win about 50 percent of the time (and lose a great deal of money on a Web site that charges commissions). Although it could be said that my trades reflect accurate probabilities, they reflect guesses or no information at all. Given prices that correspond closely to probability estimates, the higher the proportion of trades on the extremes of a probability distribution, the more confidence a prediction market will inspire. For example, if the predictions implied by prices for a season’s baseball games were almost all below five cents or above ninety-five cents a day before the games, and those prices still turned out to reflect actual probabilities, that would either be truly impressive or reveal an unusually imbalanced level of baseball competition. It is possible to derive formulas that can be used to compare sets of probability predictions (see Chapter 4), but in the absence of a set of predictions from some other source to compare to those used by TradeSports, one cannot say whether these are particularly good or particularly bad.

What one can say with confidence is that if there are methodologies that can produce predictions considerably better than those used by TradeSports, the gap should narrow over time. The reason is that someone with a superior methodology has a profit incentive to trade on TradeSports. If the methodology is superior, then it should be profitable. The trading on the methodology might have at least some effect on prices of transactions not involving the party practicing the methodology, if third parties take the trades into account in formulating their own probability assessments. If trades have no effect on prices of transactions by other traders, then the individual with the better methodology will make more and more money over time and presumably will be willing to risk progressively higher amounts in the market. Eventually, then, this trader would be responsible for a high volume of trading. Betting pursuant to a successful methodology will nudge market predictions in the direction suggested by that methodology.

This does not mean that at any given time the predictions of TradeSports will be better than what any given person could produce. Because of commission charges and the costs of risk, one will only have an incentive to place bets when one’s methodology allows for a considerable improvement in the market price. It will not be worth paying a commission of four cents per tradable contract to improve a prediction by less than four cents per tradable contract. But given current commission levels, a risk-averse participant will generally trade on information when the bet will improve the probability by 0.8 percentage points or more.21 This suggests that, at least if trading on a prediction market actually occurs at regular intervals, the prices at which the trades occur should be close to the best that can be produced by any other known method, on average and in the long run. Sometimes, someone might be hesitant to trade because of uncertainty about whether the market price already incorporates insights from the methodology, but at least those with the best information will trade when they have reason to believe that market prices appear markedly wrong.

The strongest theoretical defense of prediction markets it that traders can profit from information suggesting that the market price is wrong, yet at any given time they have not done so. Current trading prices in prediction markets in which considerable trading occurs provide roughly accurate estimates of the probability that a designated event will occur. One caveat is that at any given time, some traders might be able to beat the market with careful analysis. But there may be no uncontroversial way of identifying the people who are most likely to do so, and a prediction market allows these market-beaters to push the market price in the right direction up to the limits of their risk tolerance. Another caveat is that prediction markets are superior only to known methodologies, and it is possible that other predictive institutions might produce better results. But prediction markets provide at least a modest financial incentive for individuals to identify better methodologies and apply them.22

It is interesting that prediction markets have proved successful even in the absence of a financial incentive. Some prediction markets use “play money” or virtual currencies rather than real cash, in part because of regulatory obstacles to real-money markets. And yet many of those play-money markets have been shown to have considerable success. For example, four economists compared the accuracy of NewsFutures, which uses play money, with TradeSports in predicting the outcome of National Football League games in 2003 and found that neither performed better than the other.23 Similarly, the Hollywood Stock Exchange, a play-money market that predicts the box office returns of various Hollywood movies, has been shown to fare well in comparison with expert predictions.24 Of course, this success might be attributable in part to the intrinsic interest of football and movies and might not be replicable for issues without substantial interested populations. For participants on these exchanges, play money had real value. Where it is not feasible to create a real-money prediction market, a play-money prediction market under certain circumstances might provide an equivalent means of aggregating public opinion. It seems doubtful that this will work, however, for boring or mildly interesting issues.

Pari-mutuel wagering provides incentives similar to those provided by the probability estimate prediction market. In the latter, one can profit by buying low or selling high–that is, by identifying tradable contracts that are mispriced. In pari-mutuel wagering, one may profit by betting on an outcome that too few people have bet on. For example, if a horse player expects that there is a 10 percent chance that a horse will win but less than 10 percent of the money in the pool has been placed on that horse, then the horse player might have an incentive to bet on the horse, depending on whether the difference is sufficiently large to outweigh the expected commission. If pari-mutuel wagering works sufficiently well, less money will be gambled on long shots than on favorites. Given any particular commission, it is straightforward to convert odds into probability estimates; they are two different ways of expressing the same information.

Economic studies have shown that pari-mutuel wagering, indeed, is relatively efficient.26 That is, just as probability estimate prediction markets provide numbers that, experience suggests, truly can be interpreted as probability estimates, so will odds tend to correspond to the expected probability that any particular horse will win. This is true also for more complicated bets involving relatively large numbers of outcomes, for example, the trifecta. In order to win the trifecta, a bettor must correctly predict the first-, second-, and third-place winners of the race. At least one anomaly, however, is apparent, even in bets on a single race. In general, favorites tend to produce better payoffs than long shots, controlling for the fact that more money will be wagered on favorites than on long shots.27

Some evidence suggests that this is the result of some type of human cognitive bias,28 but there is also a relatively simple possible partial explanation.29 Bettors with relatively good information, based on either superior analytical abilities or particular facts known only to a few, often bet at the last minute. Otherwise, other bettors might see heavier-than-expected betting on a particular horse and add their own bets on that horse to the pool, thus reducing the horse’s odds. The horses that remain long shots after such last-minute betting can then be expected to return less per dollar bet than horses that attracted last-minute betting. A related point is that some bettors may be far more sophisticated than others. If a certain percentage of people bet on horses based on caprice or on flawed methodologies, more sophisticated bettors will have incentives to bet against them. The activity of the more sophisticated bettors should move the odds closer to true probability representations. Because of commissions, however, sophisticated bettors will not have an incentive to bet enough to compensate entirely for the actions of unsophisticated bettors.

There is an additional incentive to place pari-mutuel bets at the last minute: information can change, so one might as well make a final decision at the last possible moment. A particularly important type of information is the amount that other bettors will place on a particular outcome, so even if one is confident that one’s probability assessment will not change before the start of a race, it may make sense to wait to determine which outcomes are relatively good bargains. This incentive to place bets at the last minute, however, suggests a significant weakness of pari-mutuel betting relative to prediction markets. Traders in a prediction market will generally have an incentive to trade on information at the earliest possible time, before the market has fully priced the information. As a result, probabilities in a prediction market should come close to reflecting all available information at any time. Moreover, the ability to buy and sell in a prediction market allows someone to profit on information without waiting for the conclusion of an event that might depend on many subsequent pieces of information as well. Someone who acquires information early can buy tradable contracts and then sell them once the information becomes generally known. With pari-mutuel betting, by contrast, some available information will not be priced efficiently until later, and bettors must wait for the event to conclude in order to profit on sound bets.

We will see that one could design a pari-mutuel betting scheme that, like a probability estimate prediction market, produces incentives for information aggregation over time (see Chapter 3). There is relatively little incentive, however, for race tracks to adopt such a scheme, because it does not matter to race tracks whether the odds at any given time are as accurate as they can be. The media, however, generally have an interest in reporting the best available information, even for topics such as horse racing. It is interesting that prediction markets for horse racing also exist (for example, on TradeSports), and so the information from such markets might well be valuable to bettors. I do not know of any study comparing early betting on TradeSports to early odds, but it seems possible that TradeSports might be a slightly more accurate predictor well before a race. Of course, because the total amount bet on such prediction markets is relatively small, bettors with private information still might choose not to reveal their information by participating in the prediction market, and so the final odds at the track might still better reflect the outcomes of the race than those offered by TradeSports.

At the least, newspapers that publish odds and betting lines for sporting events might consider publishing prediction market forecasts, especially when significant trading has occurred sufficiently in advance of the event to make the late edition. Newspapers sometimes claim that they publish odds for entertainment purposes. However dubious this claim, a sports fan might derive some value from obtaining predictions of upcoming events. Of course, prediction markets may suffer from a chicken-and-egg problem; newspapers will not publish prediction market results when readers do not understand what they mean or think them relevant, and readers will not learn when newspapers will not publish. Someday, though, it is quite possible that prediction market predictions will earn a place in the morning sports pages beside the scores and the standings.

Sports pages might also include prediction market trading on past events. TradeSports allows betting not only before but also during athletic contests; pari-mutuel wagering, by contrast, cannot occur during the event. In a graph representing prediction market trading, the price lines show the progress of a game. Consider, for example, figure 1.5, which reports the TradeSports prices from a basketball game in which the favorite initially fell behind and then staged a comeback, winning in overtime. In contrast, a game in which the favorite performs about as expected can be visualized as more or less a straight line rising slightly to the maximum payoff (which, in TradeSports, is ten dollars rather than one dollar per contract). The more ups and downs the line has, the more exciting the game. In the game represented in figure 1.5, the excitement began only shortly before the two-hour mark. It may be a long time before readers of the sports pages do what readers of the business pages often do: look at a price graph for a snapshot of the day’s news. But such graphical representation can leave the media to focus on what the graph cannot convey: the individual plays and performances that make sports more interesting to watch than stock tickers.

This provides at least informal support for the proposition that prediction markets might help overcome heuristics and biases that ordinarily might distort human probability estimates. When a team has a winning streak, each of the wins matters in the standings, and one’s estimate of the team’s quality should correspondingly increase. But some fans might believe that streaks are more significant than they are. Conventional wisdom in sports, after all, has long supported the notion that a player with a “hot hand” will be more likely to succeed than the player’s season and career numbers would suggest, but economists appear to have debunked this with statistical research.30 The belief in the hot hand, as well as the related overemphasis among fans of the significance of a streak, is an example of what is sometimes known as the “clustering illusion,”31 a tendency among some humans to see a pattern where none exists, where really there is nothing more than a sequence of independent events. The relative stability of prediction market prices over time provides informal evidence that in part the market overcomes the clustering illusion.

The clustering illusion is one of a number of biases that may affect individual probability estimates. Among the most famous is the hindsight bias, the tendency of individuals to see actual outcomes as more likely ex ante than they in fact were.32 Other biases include the primary and recency effects, the tendencies to weigh initial and recent events more than other events. Another fallacy to which sports fans might be particularly susceptible is the illusory correlation bias, the tendency to infer causation from a correlation, for example, giving a new coach credit for a team’s changing fortunes. Many of these biases may be related to what is perhaps the most important bias, the availability error,33 the tendency to place too much weight on events that are easily accessible in memory. A baseball fan, for example, is more likely to remember a player’s dramatic home run than the many instances in which the player grounded out to the shortstop. As a result, the fan might overestimate the probability of another home run. Although a fan knows that a player with a .350 batting average and a .400 on-base percentage is more likely than not to make an out in any given plate appearance, it does not seem that way.

Overcoming biases afflicting sports fans might not promote the welfare of news consumers. But overcoming biases can be of greater importance. Timur Kuran and Cass Sunstein have argued that the availability error leads to bad policy consequences and that the media play a major role.34 For example, Kuran and Sunstein claim that the publicity surrounding Love Canal, a neighborhood built on a landfill containing chemical waste, had no basis in science. Residents and other observers placed too much weight on individual instances of health problems, disregarding studies indicating that the overall level of health problems was not statistically anomalous. The media then reported the concerns, raising the availability of the health problems in public discourse and furthering mistaken risk perceptions by the public. Kuran and Sunstein label the effect an “availability cascade.” Whether or not they are right about Love Canal, the example demonstrates how the media can aggravate public misperceptions. Potentially aggravating availability cascades are reputational cascades, in which media and other experts conclude that agreeing with the consensus will promote their career prospects.35

Journalists generally admit to having at least one bias: a preference for stories with news to stories without news. The local evening news thus reports on several incidents of fire and crime but rarely on people who emerged from a given day comparatively unscathed. The evidence for the availability heuristic suggests that the public as a result will overestimate the probability and prevalence of the events that tend to lead to news reports. The media, of course, sometimes might counteract this tendency by reporting hard facts. Sometimes hard facts will make a less vivid impact than more sensational news stories, but they tend to move individual probability assessments in the right direction. There are not always hard facts to report, however, especially when the media report on the possibility of future risks. If prediction markets help overcome individual cognitive biases, consistent reporting on prediction market predictions might help.

The empirical question is the extent to which prediction markets do overcome cognitive biases. Early evidence suggests that prediction markets do not succeed altogether in vanquishing them. Paul Tetlock, for example, analyzed pricing in TradeSports markets to determine whether the market overcame the favorite–long shot bias. He also tested for what is known as the reverse favorite–long shot bias, an observed tendency of baseball and hockey wagering markets to underprice teams that are not favored to win.36 He found both. The favorite–long shot bias dominated for very unlikely events, and the reverse bias occurred in middle probability ranges.37 He also found evidence of “return reversals.”38 That is, after the market initially moves in a particular direction following the appearance of new information, it is likely to move partway in the opposite direction afterwards. This indicates overreaction to new information, which might be attributable to the availability heuristic.

Such evidence disproves any claims that prediction markets are completely free of cognitive biases. Nonetheless, it is possible that other methods of deriving consensus probability estimates–such as asking sports fans their opinions–might be more prone to error. There is at least a theoretical reason to believe that prediction markets should help alleviate biases. Individuals with more sophisticated models for assessing probabilities generally have greater self-confidence about their predictions and thus greater willingness to place bets on prediction markets when those models suggest that current prices are inaccurate. Indeed, some sophisticated market players might focus consciously on the specific types of cognitive errors that are likely to produce poor probability assessments and trade against them. Tetlock’s analysis suggests a particular betting strategy to take advantage of the reverse favorite–long shot bias that, he estimates, would produce phenomenal positive returns of 10 percent after commissions. The strategy requires buying contracts that have recently experienced bad news and selling those that have recently experienced good news. If Tetlock is correct, then someone who is persuaded by his analysis will implement this betting strategy, and prediction market pricing should improve. This may not happen overnight, because every trader must worry that others are already trading using the strategy, but it seems likely to occur eventually.

If prediction markets do help counteract biases, it will be a long time before media reporting of the results of prediction markets will significantly affect public policy. For starters, only a relatively small number of prediction markets currently exist, at least outside the area of sports. These may do little to correct important public misperceptions. The many markets predicting the results of elections, for example, may have little public value aside from testing prediction market mechanisms. Consumers of media might invest a great deal of time reading about election contests, but biases in predictions of the outcomes of such contests likely have relatively little consequence for the world. Ultimately, if the media are to play an important democratic function, it cannot be merely by telling the public who will win elections; it must be also by telling the public what it can expect in the future.

Prediction markets forecasting the probabilities of different possible news events might be considerably more useful if the public learned to understand both their abilities and their limitations. For example, figure 1.6 reports the trading prices in prediction markets used to predict whether weapons of mass destruction would be found in Iraq by at various points after the U.S. invasion of that country.39 On one hand, the graph shows that probability estimates can lead people astray. At the beginning of May, the market expected that such weapons probably would be found by the end of September, yet they were not. This should not, however, necessarily be seen as an indictment of prediction markets, for the general reason that prediction markets are not wrong merely because low-probability events end up occurring. Although the market would have been more impressive if it had bucked the conventional wisdom and proved correct, at least the market approximately conveyed the conventional wisdom of people who were relatively well informed, based on the information then publicly available.

More important, at any given time, both the graph of trading prices and the current trading price tells a concise story. A reader, to be sure, might wish to read all of the publicly available information in order to develop an independent prediction of the probability that weapons would be found, and a reader willing to devote a great deal of effort to the task might expect to be able systematically to earn profits by trading on the information on the market. But some readers, if convinced that prediction markets provide at least approximate probability estimates, might prefer to consider the prediction market summary rather than all of the underlying information. When partisans on all sides of contested issues can be expected to seek to sway public opinion by offering confident predictions about the future, the media often have trouble objectively indicating what the expert consensus on the issue is, leaving readers to sort through an awesome quantity of data. These readers will be subject to the usual assortment of heuristics and biases, and they may therefore make systematic errors. By seeking objective means of aggregating expert views, the media can help overcome such errors. Reporting of prediction market results in particular will leave readers not only with probability estimates that in general should be more accurate than their own deductions but also with more time to read analyses of appropriate policy responses.

These approaches are somewhat clunky, for they do not allow for an easy calculation of the average expected result. For example, on January 10, 2006, just after Alito’s confirmation hearings began, the bid-ask midpoint for the “more than sixty” contract was 60.8, and for the “more than seventy” contract the midpoint was 15.2. That suggests that the market estimated a 45.6 percent chance that Alito would receive between sixty and seventy votes but provides little information about what the best estimate actually would be. A partial solution would be to offer additional contracts–more than sixty-two votes, more than sixty-four, and so on–but this, too, becomes cumbersome, in part because there might be few people interested in trading on each of these contracts. The interval approach does have its uses–it is the most straightforward way of determining, for example, the probability that more than senators will vote to confirm–but it is less useful when one is seeking a point estimate of the number of votes that Alito is likely to receive.

There is, however, a simple alternative. The prediction markets we have seen so far are binary–for example, paying off ten dollars per contract if the event occurs or not paying at all. But payouts can range depending on a number that is verifiable ex post. For example, TradeSports might have used this approach with Alito, promising to pay nothing if Alito received no votes, ten cents if Alito received one vote, and so on. Of course, it is particularly convenient that the Senate happens to have one hundred members at this time, making each vote worth a dime. It will similarly be simple to design a prediction market that predicts a percentage. For example, in addition to sponsoring “winner-take-all” markets, the Iowa Electronic Markets sponsors vote share markets. Figure 1.7 shows the vote share market for the 2004 U.S. presidential election. Each penny in closing price corresponds to an anticipated 1 percent of the vote. Aside from a temporary spike in Kerry shares (to which I return below), the anticipated outcome was remarkably stable, providing the appearance of an almost boring election cycle. Of course, the election was not boring, because the anticipated Bush and Kerry shares happened to be quite close at some points in the election cycle. The graph, however, emphasizes that the real fight was about a relatively small number of swing votes.

A numeric estimate prediction market can also be used also to predict numbers that will fall on a closed interval between numbers other than zero and one hundred. One possibility, if the interval is a subset of that range, is simply to use the same approach as the vote share market. For example, TradeSports includes shares that predict the number of games each team in the National Basketball Association will win in a particular year’s regular season. The maximum possible is eighty-two according to the current NBA schedule, but this simply means that no one will rationally purchase a chance of a $10 payout for a price above $8.20. Another possibility is to map the range linearly from zero to one hundred. So, for example, TradeSports might have provided that a tradable contract would pay off 1000/82 cents per win.

More generally, a prediction market can be used to predict any number that will become apparent at some later time. A prediction market could be used for a number that might turn out to be positive or negative. Either the interval can be translated to a more familiar one, such as zero to one hundred, or trading at negative prices could be allowed. If I purchased a tradable contract at minus five cents I would receive five cents from the seller. A prediction market also might be used to predict a number on an unbounded interval, say, from zero to infinity. The only problem is that if very high numbers become plausible, then the prediction market sponsor might be unable to pay off the contract, and individual traders might not be able to afford to engage in transactions. Thus, as a general matter, the trick is to translate a plausible range of numbers into a manageable contract size. The rules can always provide that if the actual number exceeds the maximum, then the payoff will be the maximum. For example, a prediction market might be used to predict how many points an NBA team will score in a season, with every hundred points paying off one cent. The rules could stipulate that ten thousand points or more will pay off one dollar. If this outcome is particularly unlikely, it should have little effect on the pricing.

Robert Erikson and Christopher Wlezien have sought to debunk the claimed superiority of prediction markets to polls, at least well in advance of elections.44 Although on average the markets beat a naïve reading of polls, a more sophisticated methodology, discounting the favorite’s lead on the basis of research indicating that leads tend to shrink, makes polls better than the market. One danger of this conclusion is that it is often possible after the fact to construct a model that produces high levels of accuracy. We cannot be sure that Erikson and Wlezien would have chosen the same methodology if forced to announce a prediction algorithm before the elections they surveyed. Their approach, however, does not appear particularly complicated, and it at least suggests that we should not be confident that market prices are better than any that might be derived using alternative techniques, such as the regression analysis that Erikson and Wlezien provide.

It would be a mistake, however, to conclude from Erikson and Wlezien’s analysis that markets did a poor job of aggregating predictions. Their analysis of discounting of leads, after all, had not been performed prior to the election, and if it had been, reasonable observers might have differed about the extent to which past trends would continue into later elections. Erikson and Wlezien point out that given the volume of trading on the Iowa markets, their profit from trading on the strategy would have been quite low. Should they or others make available predictions based on their methodology in future elections, other traders making assessments might take these numbers into account in their own trading. Erikson and Wlezien say that polls “that are properly discounted for the favorite’s inflated lead outperform the market,” but the properly discounted numbers were not widely available at the time.

Prediction markets cannot necessarily be counted on to beat any other method of aggregating data and predictions. Laboratory experiments involving securities markets show that sometimes these markets do not reflect “rational expectations” by perfectly incorporating information.45 For instance, Joyce Berg and Thomas Rietz found small inefficiencies in the Iowa Electronic Markets that could serve as the basis of a profitable trading strategy.46 What makes prediction markets particularly useful for the media is that they are relatively objective while still accurate compared to other approaches to aggregation. Someone with an agenda, for example, promoting a particular candidate, might devise a statistical methodology that makes that candidate appear stronger than previously thought. And usually it is easy to find some basis for defending the methodology on the merits. A journalist is not in a good position to make decisions about whether the creators of a methodology were motivated by a desire to promote particular results or simply by a desire to develop the best predictions possible. And even if a journalist can be confident about objectivity, different objective people might reasonably take different approaches to arrive at different conclusions.

Of course, a prediction market can be seen as just another methodology, but at least it will not be easy to create a prediction market to promote a particular agenda. Robert Forsythe and coauthors have shown that although the participants in an early election market were predominantly Republicans, and though many participants only traded in favor of their preferred candidates, this did not produce a biased prediction.47 Rather, enough people traded both for and against their preferred candidates that the demographics of the trader community did not have much effect. The traders who were particularly influential were more likely than other traders to place limit orders at prices slightly different from the market price. This suggests that if one opens a prediction market to a modestly large group of traders, one will not be able to affect the market price much by choosing a demographically unrepresentative sample. It will be particularly difficult to skew a prediction market generally available to traders on the Internet.

The great virtue, then, of prediction markets is not that they will always be more accurate than alternative future methodologies or existing ones. Rather, it is that prediction markets are relatively objective. At the same time, there is a strong theoretical reason to believe that prediction market participants will seek to take into account at least easily accessible data such as opinion polls.48 By citing prediction market forecasts, a journalist can give a concise, consensus prediction, whether about an upcoming election or about any other event being forecast by a prediction market.

There might be prediction markets, however, that some individuals would like to manipulate.49 Suppose, for example, that some voters decide whom to vote for in part based on whom they expect to win.50 Further, suppose that some of these same voters base their predictions of who will win either directly or indirectly on prediction markets. It might then be in the financial or at least the ideological interest of some wealthy individuals to seek to manipulate prediction markets that forecast election returns. A significant investment in prediction markets might have the direct effect of causing financial loss, at least if prices immediately push back in the original direction. Such an investment, however, conceivably could affect the election if the results were readily available to the public. This scenario admittedly seems unlikely today given the lack of attention to prediction market forecasts, but the question remains whether prediction markets will remain effective if intense pressures exist to manipulate them.

There are theoretical and empirical reasons to believe that prediction markets can be manipulated in the short term. The theoretical point is simply that the number of orders on the bid and ask queues at any time is finite. Rational traders will limit their exposure to trading on bid and ask queues, since leaving offers on such queues opens traders to losses in the event that new developments make substantial changes in probability assessments. So, if I buy enough shares, I can buy not only all the shares available at the lowest quoted price but also some shares at the next higher price. At least at the moment the trade is completed, the last traded price will likely be different from the previous price, and the midpoint of the bid-ask spread will also rise. Indeed, for enough money, I should be able temporarily to drive the price to whatever level I please.

The empirical point is that there have been efforts to manipulate prediction markets. Such an attempt presumably explains the early October blip in Kerry shares visible in figure 1.7. Other attempts have been documented. For example, speculators briefly increased the value of Pat Buchanan shares in an Iowa Electronic Markets contest.51 In addition, Paul Rhode and Koleman Strumpf document and analyze two sustained attempts to manipulate the TradeSports 2004 election markets on September 13 and October 15, 2004, leading to large price changes that do not appear to have been based on any information.52 Rhode and Strumpf show that such attempts at manipulation were nothing new. Precursors to modern election markets existed between 1880 and 1940,53 and Rhode and Strumpf identify apparent attempts at manipulation in these early markets.54 In all cases, the authors show that the effects of manipulation attempts on prices decreased over time.

Rhode and Strumpf also ran an interesting field experiment on the Iowa Electronic Markets during the 2000 presidential campaign.55 At predetermined times, they made a total of eleven planned trades, determining at random the side in which to invest. On some occasions, Rhode and Strumpf bet only on one of the election markets (either vote share or winner take all), to determine whether the trade on one market contaminated the other. Such contamination indeed may have occurred in the minutes after the initial trade, though it was not statistically significant. On other occasions, Rhode and Strumpf attacked both election markets simultaneously. That approach might be expected to be more likely to convince other market participants that the trades were based on real information, because someone seeking to profit on new information is likely to bet on both markets. In all cases, Rhode and Strumpf demonstrate significant changes in the short term on the level of other participants’ trades, but as hours passed, there was no effect. Planning random trades would appear to have no effect at all on prices twenty-four hours later.

These results are not surprising. There is relatively little “private” information about presidential elections, or at least relatively little that one could keep secret for very long. Someone playing the markets, therefore, would place relatively little emphasis in deciding what to do about trades made twenty-four hours earlier. Over time, the effects of trades should be expected to dissipate. But this is not entirely reassuring for two reasons. The first is John Maynard Keynes’s observation that in the long run we will all be dead. The long run in this context is very short, but the short run is still long enough that manipulation conceivably could have some effect. This is particularly true if a market is expected to end at a particular time, so manipulation immediately prior to the market end might go unchecked. The second is that some prediction markets might depend on a relatively small number of traders who have a fair amount of private information or private analysis of public information. In such a market, a trader might put considerable emphasis on preceding trades.

Other experimental evidence, however, suggests that manipulation may be difficult even in a market in which private information exists, at least when market participants are aware of the incentives for manipulation. Robin Hanson, Ryan Oprea, and David Porter ran an experimental market in which subjects traded an asset and received different information relevant to determining the value of the asset. Some randomly chosen traders were given an incentive to manipulate the market: they were promised that the higher the final market price, the higher the payout. All traders were aware that some traders were given an incentive to manipulate the market and knew whether the incentive was to manipulate prices up or down. Those who were given manipulation incentives in fact submitted higher bids, but this ultimately had no effect on market prices. Individuals who were not given manipulation incentives bid against the manipulators and drove the market prices down.

More generally, to the extent that traders know in advance that certain parties will have incentives to manipulate, they will have incentives to counteract such manipulation by seeking to push prices in the opposite direction. Indeed, Hanson et al. show that attempts to manipulate the market might increase market accuracy.56 In securities markets generally, “noise traders,” that is, traders who do not trade according to fundamentals, may increase market accuracy because informed traders can profit by trading against them. (In Chapter 7 we will encounter arguments that noise traders may decrease market accuracy.) Adding unidentifiable noise traders for any particular security may decrease accuracy, but noise traders as a whole should increase accuracy. In a prediction market, the larger the number of people who can be expected to trade in a market for reasons unrelated to fundamentals, the greater the incentives for others to enter the market, and the additional insight provided by these entrants should increase market accuracy. From the perspective of market efficiency, Hanson et al. argue, manipulators can be seen simply as noise traders.

Manipulation is more likely to succeed when market participants are not informed of the incentive of some participants to manipulate a market. An experiment by Martin Strobel confirms this.57 The experiment used prediction markets to estimate the number of black balls and white balls in an urn. Different participants in the market were able to view different subsets of the balls. In some iterations of the experiment, a robot trader sought to manipulate the market in one direction. The results confirmed that the manipulation attempts did have a statistically significant effect in the expected direction. Market participants in this experiment cannot know whether someone with whom they are trading is trading on the basis of information or on the basis of a manipulation incentive. They thus ascribe some positive probability to the contingency that the trades reflect information, and this changes the assessment of market participants about the number of balls in the urn.

Manipulation seems likely to have a long-term effect on markets only to the extent that market participants misestimate the extent of attempts at such manipulation. For example, if market participants believe that there are probably individuals who are seeking to bid up market prices, these participants will respond by seeking to push prices in the opposite direction. But if in fact there are no manipulators, then this error will cause prices to be too low. On the other hand, if market participants underestimate the extent of market manipulation, that manipulation will be somewhat successful. In order to manipulate a market successfully, one has to attempt manipulation in higher volumes than other parties expect. This analysis suggests that manipulation might affect prediction market prices, but only to the extent that neutral market participants are genuinely fooled into misestimating others’ private information. Prediction markets should continue to reflect consensus predictions, but misinformation through trading can have some effect. When the general level of manipulation is known but it is not known which tradable contracts have been targeted and in which direction, manipulation should increase overall average price accuracy while nonetheless succeeding at biasing the particular targeted results.

Manipulation may be the most serious obstacle to widespread adoption of prediction markets, but it will not tend to bias prices when the degree of incentives of participants to manipulate is known. It seems especially unlikely to be a problem where there is little private information and thus little derivatively informed trading, that is, where individuals’ price assessments are not greatly affected by the trades of others. If manipulation is nonetheless deemed too serious a problem for some applications of prediction markets, there are two possible solutions. First, where there is a discrete group of potential manipulators, those individuals can be barred from participation. Of course, there is always a danger that these potential manipulators can pay off other market participants, but legal or contractual sanctions can reduce that possibility. Second, prediction markets might be limited to a group of authorized traders who are believed to have no incentive to manipulate the outcome.