Prediction Markets: the Collective Knowledge of Market Participants

Prediction Markets: The Collective Knowledge of Market Participants Justin Wolters Associate Professor ofBusiness and Public Policy Wharton School, University of Pennsylvania Philadelphia

Prediction markets provide an information-aggregation technology applicable to a variety of topics, including political and financial risk. Because of the human idiosyncrasiesidentifiedbybehavioralfinance, predictionmarkets canfail, buthistorical data show them to be as accurate as traditional polling methodologies and far less expensive to establish and maintain.

n this discussion, I will make three substantive Sports Prediction Markets. Interestingly, I claims about prediction markets. First, if mar sportsbetting markets manifest manyof the charac kets really are efficient, as the efficient market teristics expected of an efficient market, and they hypothesis asserts/then the prices that come out of now allow participants to trade stock during the any market contain valuable information. To make gameonthelikelihoodofa particularteamwinning. the case for this claim, I will use data from sports For example, Figure 1 shows real-time betting dur betting markets. Second, prediction markets can be ing Game 6 ofthe 2003 NationalLeague Champion used to trackpoliticalrisk, whichcanbe a keyfactor ship Series (NLCS). This prediction market from driving investment performance. Third, prediction Intrade pays $1 ifand only ifthe Chicago Cubs win . markets can fail, so I will conclude my discussion this particular game. At the beginning of the game, by describing why prediction markets work and itlooks as ifthe Cubshave abouta 75 percentchance what causes them to fail. of winning. In the top ofthe eighthinning, the Cubs go ahead of the Florida Marlins 3-D, and the chance Prediction Markets and the of the Cubs winning the game and going to the World Series reaches 95 percent. Unfortunately, in Efficient Market Hypothesis the bottom of that same inning, one of the great e efficient market hypothesis asserts that market collapsesinsportinghistoryoccurred. A fan reaches prices fully reflect available information. But the out from the stands and prevents a catch by Cubs degree to whichinformationis available determines outfielder Moises Alou. The Marlins proceed to theform inwhichthe hypothesis reveals itself. Inits score eight runs in that inning, and the market strong form, prices reflect all information; in its responds immediately to this new information. In a 'semi-strong form, prices reflect all publicly avail matter ofminutes, the marketprice declines from 95 able information; in its weak form, prices reflect cents to about 5 cents. Only past data. In anyone of these forms, however, The questionmoreimportantthanwhetherpre rices can reflect, summarize, and aggregate a huge diction markets respond quickly to news, though, is ount of information, but the unfortunate impli- whetherthese marketpricesareaccurate. To answer ation of the hypothesis is that excess returns are that question, I have charted data from alI 21,885 predictable. Thus, prices can be said to follow a major league baseball games played from 1991 andomwalk. through 2000, and the data show that the predic tions are quite close to the actual outcomes. For is presentation comes from the 2008 CFA Institute Annual Confer instance, if the market says a team has a 33 percent held in Vancouver, British Columbia, Canada, on 11-14May2008. chance ofwinning, it actualIywins 31 percent ofthe

2009 CFA Institute • cfapubs.org JUNE 2009 • 37 CFA Institute Conference Proceedings Quarterly

Figure 1. Real·Time Betting on the Chicago Cubs: 2003 NLCS, Game 6

Price ($) 1.00 ,------;:------, ~Io.-~--- Fan Spoils Alou's Catch

0.80

Top of the 8th Iruting 0.60

0.40

0.20

o 8:05p.m. 9:00p.m. !0:00p.m. 11:00 p.m.

Note: Time is given in central (Chicago) time. Source: Based on data from Intrade. time. When the market says a team has a 67 percent ingtake, whichare, again, quite accurate. Therefore, chance of winning, it wins 65 percent of the time, an analyst covering the movie industry may find and when the market says a team has a 90 percent these forecasts quite useful, which means that such chance of winning, it wins 88 percent of the time. an exchange moves us toward the realm of socially The difference is how Las Vegas makes money. But useful prediction markets. the real point is that if a person wants accurate For instance, Deutsche Bank and Goldman information on the likely outcomes of baseball Sachs briefly started a new market (now folded, games, the betting odds do an excellent job of pro viding that information. unfortunately) that they called the "economic The same point holds true for other sports. For derivatives market." It allowed investors to trade example, based on market expectations and game on thefuture outcomes ofparticulareconomicindi outcomes for the National Football League from cators, such as retail sales and business confidence. 1984 to 2000 (Le., for 3,791 games), a regression The economic derivatives market provided fore analysis shows a slope of almost 1.0 (0.997 to be casts for nonfarm payrolls, initial unemployment exact). Thus, when the market says that a team is claims, retail sales, and business confidence, and likely to win and gives a margin of a prespecified based on data thatRefet Giirkaynak andI gathered, number ofpoints (the "pointspread"), thatteam, on these forecasts were more accurate than an average average, wins by the point spread. or consensus forecast of 35 economists. Donot, however,thinkofthesefindings as mere My conclusion, therefore, is that rather than conversation topics. Rather, think of prediction averaging theforecasts of35economists, thosesame markets as a technology that can be used to aggre economists should be put in a room and told to bet gate opinions about anything of importance. or trade against each other. Taking the rnarket Socially Useful Prediction Markets. Consider aggregated price from this tradingwouldprovide a the firm called the Hollywood Stock Exchange. It more efficient way of aggregating forecasts. The runs prediction markets on the opening weekend task, then, is to harness this technology to more box office take of various movies. Essentially, it challengingissues. One suchissueis politics. Infact, allows people to trade, or bet, on how well each prediction markets have been used quite exten movie will perform. Data from that market can sively to predict political risk, which is a key factor show the aggregated forecast and the actual open- driving investment performance.

Prediction Markets and Political therefore, represent a valuable gain in accuracy and a significant reduction in costs. Risk il Results from recent cycles. In the 2004 U.S. In2003, a recall electionwas held to remove Califor election cycle, investors cou1d trade not only which nia Governor Gray Davis from office and replace candidate wou1d become president but also who him with one of several contenders, one of whom wou1d win the Electoral College votes ofeach state. was Arnold Schwarzenegger. Prediction markets Based on data from Intrade, the prediction market allowed people to trade on the question of whether price accurately predicted every state that went for Schwarzenegger would win the governorship. A George W. Bush and every state that went against contract paying $100 if Schwarzenegger won could him. For the 2004 U.S. Senate races, theIntrade price be traded on two exchanges-Intrade, which is was correct for every state but Alaska. In the 2006 based in Ireland, or the World Sports Exchange, election cycle, most knowledgeable observers and which is based in the Caribbean. Attempts to recall many Republicans did not expect the Democrats to other governors had occurred in California, but win enough seats to gain the majority in the Senate. none had been successful. Nevertheless, I set my But every Single Senate race went exactly the way computer to look for arbitrage opportunities the election-eve prediction market said it wou1d. between Intrade and the World Sports Exchange. ii1 History ofpolitical prediction markets. Such Every four hours, my computer checked these two predictionmarkets seemlike a marvelous new tech websites and searched for potential profits. By the nologyfor forecasting electionresu1ts, butas itturns end of the exercise, I had made no money, but I had out,betting onpolitics is as oldas politicsitself. Two learned an important research lesson. When I fol enterprising economic historians-Rhode and lowed the chart of the two exchange prices, the bid Strumpf-have gathered data on the amount ofbet price in one market was never above the asking ting that has occurred on U.S. presidential elections price inthe other, whichmeans thatthe twomarkets from 1884 through 1928. Opinion polling did not are closely linked. The price pattern that we are so exist in any systematic or nationwide manner at the familiar with in other financial markets appears to time, butbetting on elections occurred regu1arly on hold true for poli.tical prediction markets. the curb where the American Stock Exchange now stands. In fact, large sums of money were traded. Accuracy of Political Prediction Markets. For the 1884-1928 period, an average of $37 million These markets have a tremendous historical record. (in 2002 dollars) was bet per election, which turns In 1988, facu1ty from the University ofIowa created out to be a $2.28 bet per vote and the equivalent of a trading market now called the Iowa Electronic 54 percent of total campaign spending. By using Markets. This market has followed hundreds of archival data, one can follow the performance of elections in the United States and elsewhere, so an political prediction markets for well over 100 years. extensive record exists to chart its predictive accu In fact, in a comparison of the election-eve pre racy. Except for a few misses, the predicted outcome dictionmarketsfor 1880 through2004 withtheactual and the actual outcome of most elections converge results, once again, the prediction markets accu along the same line. rately indicated-with two exceptions-whether a Considerinparticu1ar theU.5. presidentialelec Republican wou1d win the election. In2000, the pre tions. When I imagine trying to predict the winner diction markets picked George Bush to win, which ofthenextelection, I typically thinkoftwomethods. he did, but he did not win the popu1ar vote as The first is to create a traditional polling operation. expected. The other exception occurred in the 1948 I would hire hundreds of staff members, set up election andis, I think, the most famous exceptionin massive phone banks, call thousands of Americans U.S. political history-the election in which Harry around the country, and then have top-notch statis Truman unexpectedly defeated Thomas Dewey. ticians analyze the resu1ting data to come up with a Those two resu1ts aside, however, the prediction forecast. The second method is to set up a website markets have been decidedly accurate. and let people log on and trade to their heart's content. I wou1d then use the resu1ting prices to Utility of Prediction Markets. Based on the establish a prediction. The first method is quite historical evidence, the prediction market seems to accurate. It tends to predict U.S. presidential elec be a wonderful tool, if for no other reason than it tions by rougWy ± 2.0 percentage points, but it is makes all of us political forecasters. The efficient also expensivetooperate. Thesecondmethod-that markethypothesissaysthatallavailableinformation is, a political prediction market-is far less expen is embodied inthe marketprice. Predictionmarkets, sive, and it forecasts U.S. presidential elections to such as Intrade or the Iowa Electronic Markets, offer within ±1.4 percentage points. Prediction markets, access to that very marketprice. Therefore, all I have

to do is look up one price and I become a political chances of winning the nomination did not change. forecaster. I donotneedto readthe Wall StreetJournal, TheprimariesinMayledtoa splitdecision-Indiana the Washington Post, or the New York Times. In fact, I for Clinton, North Carolina for Obama---so both have begun acting on this very idea. I am now writ could claim a victory. But at this point, the political ing political commentary on the current election prediction markets decided that splitting victories cycle for the Wall Street Journal, and the only data I was better news for Obama than for Clinton, so his am allowed to use are prediction market data. lf the stock price climbed dramatically. political prediction market actually aggregates all available information, that is the only information The key point is that prediction markets pro source any of us >;hould be watching. We can ignore vide an excellent way to track political risk. But the all the other commentary. ability to track political risk is not an end in itself; Figure 2 illustrates the sort of narrative I have its ultimatepurpose, inthis case atleast, is its utility beenpresenting sofar. Itshows thefluctuatingprice as a valuation factor. of a security during the 2008 Democratic nomina Political Risks and Financial Markets. As tion race that pays $100 if Barack Obama becomes a social scientist, if I wanted to find out whether thenominee. Untilhis surprisingvictoryintheIowa Democrats or Republicans are good for stocks, I caucuses, theprice onthe securitywas only$20, but would like to do the same thing a physical scientist it quicklymoved upward, to around $70 at the time would do: Run an experiment in which I randomly ofthe New Hampshireprimary. Atthatpoint, how ever, Hillary Clinton's dramatic comeback caused assign the presidency half the time to Republicans Obama's stock to fall butnot topre-Iowalevels. The and half the time to Democrats and then measure Michigan primary turned out to be no news for how stocks performed under each party. Fortu anyone, butSuper Tuesdayproved tobe goodnews natelyfor me, suchanexperiment actually occurred for Obama, and he got more good news with the on Election Day 2004. Potomacprimary. After Obama lostOhio andTexas By following the changing probability of who to Clinton, Obama's stock declined again slightly. George Bush or John Kerry-would become presi Even after·Obama lost the PennsylVania pri dent and comparing that probability with the mary, his stock value remained high. According to simultaneous rise and fall of the S&P 500 Index, I the New York Times, Clinton's victory in Pennsylva was able to perceive the likely effect of either pres nia was clearly important; however, seen through idency on the stock market. For example, on 2 the lens of political prediction markets, she won November 2004, from noon (when the probability PennsylVania by the margin expected in a state with of Bush retaining the presidency was 60 percent) to Pennsylvania's characteristics. As a result, her 6:00 p.m.-when exit polls put Bush's fortunes at

Figure 2. Probability of Obama Winning the 2008 Democratic Nomination, 1 January 2008 to 12 May 2008

Probability (%) 100 ,, , IA INH MI Super !Potomac 'OH,TX PAl, IN,INC Tuesday ,[Primary 80 i I

OL.L--'--L__~'-----,---__,-,-- ~__----,-_~'------.J I I/lan/08 I/Feb/08 I/Mar/08 1/Apr/08 I/May/08 r f

40 • JUNE 2009 ©2009 CFA Institute • cfapubs.org Prediction Markets their lowest point (i.e., a 30 percent probability that The challenge is to find creative ways to deploy he would retain the presidency)-the S&P 500 fell this technology to better forecast our future. By 0.7 percent. Thus, if reducing the likelihood of a usingprediction markets, qualitative issues that are Bush presidency by 30 percentage points causes hardto valuecannowbe quantifiedandaggregated stocks to decline 0.7 percent, I can infer that remov in a useful way. ing George Bush entirely would cause stocks to decline 2.3 percent (-0.7 percent/-30 percentage Why Prediction Markets Work and points). As the actual vote counting began, how ever, the probability of Bush retaining the presi What Causes Them to Fail dency rose 65 percentage points injusta few hours. "When it comes to large social issues-such as war Everyone had already voted, and the S&P 500 rose or global warming-media commentators may err 1.3 percent, thus suggesting that stocks would be by taking popular rather than accurate positions. worth perhaps 2 percent more (1.3 percent/65 per After all, they are not taking positions with their centage points) under Bush than under Kerry. own assets. But that is exactly what people in the Republicans, I am sure, will be reassured to learn prediction markets, like those in the financial mar that their candidate is good for the stock market, kets, do every day. So, prediction markets are quite and Democrats will be reassured to see that the effective at removing incentives to be popular difference between the two is actually rather small. instead of accurate. Similarly, prediction markets, Equally interesting is the issue ofwhich compa like financial markets, encourage participants to nies are expected to do well under which candi seek outbetter information and act on thatinforma dates. According to a paper by Knight (2006), a tion, thus increasingtheir chances ofoutperforming George Bush stock index includes such businesses other participants. as big oil, big tobacco, and big pharmaceuticals. By Ifno usefulinformation is available to be aggre contrast, a John Kerry index includes environmen gated, however, prediction markets are not espe tallyfriendly manufacturingcompanies andgeneric cially helpful. After all, they are nothing more than pharmaceuticals. From March to October 2004, as anefficientmeans ofaggregating available informa each nominee's fortunes rose and fell, so did their tion. For example, when the prediction markets indices. As Bush's fortunes rose, so did his index, were asked whether weapons of mass destruction and when his fortunes fell, his stock index also fell. wouldbe discoveredinIraqbyMay2003, June2003, July 2003, or September 2003, they were no more Prediction Markets and Broad Political accurate than the rest of us. Issues. Predictionmarkets canbe used to measure other forms ofpolitical risk. For example,before the Behavioral Finance and Long-Shot Bias. United States went to war in Iraq, Intrade had a Besides the question of available information, we security that was worth $100 if Saddam Hussein mustalso keep inmindrecent researchinbehavioral was no longer the leader of Iraq by June 2003. finance. Such research demonstrates thatpeople do Because the United States appeared less likely from not always act in their apparent best interests. SeptemberthroughNovember2002 to go to war, the Therefore, human behavior can prevent prediction Saddam Hussein security dropped from a value of marketsfrombeingasefficientaspossible. Consider about $80 to about $40. Then, as the likelihood of one of the best-documented mistakes that individ war rose for the next several months, so did the uals make in markets. The difference between a SaddamHussein security, reaching a value ofabout small probability and a tiny probability can be dif $75 in February 2003, a couple ofmonths before the ficult to perceive, even in situations like horse rac war actually began. ing, where reasonably reliable odds are known in In contrast, the S&P 500 moved largely in the advance. Despite this knowledge, individuals tend opposite direction of the Saddam Hussein to bet too much on an extremely low-probability security-aclearly negative correlation. One could eventrelative to whattheyare prepared to wager on interpret this movement as a warning from the an outcome with better odds. Anyone who follows markets, incontrastto theviews expressedbymost the markets should be aware of this reality. political and economic commentators, that going to war in Iraq would be bad for the U.S. economy. Behavioral Finance and Human Pre The data suggestthatthe difference betweenpeace conceptions. Another problem in markets is that and war was worth as many as 15 percentage we all bring our preconceived notions to the mar points for U.S. stocks. Based on these and similar kets. To help people understand this problem, I results, it seems clear that political risk is a key sometimes conduct a little thought experiment. factor in investment valuations. When I ask people to imagine a great CEO, most

people will choose a tall, white male wearing a prices often provide efficient forecasts, including pinstripe suit. Thus, female CEOs do not fit the forecasts of political risks. Making money in the typical representation ofa great CEO, and investors markets may be difficult, but prices offer a useful may be likely to undervalue companies directed by summarystatistic of available information. Second, a female CEO. To investigate this hypothesis, I gath prediction market prices can be useful inputs for ered all the earnings announcements in the United decision making in finance, management, and pub States duringthelastdecadeand organized themby lic policy. Third, the success ofprediction markets is the size of the earnings surprise that occurred. context specific, and the idiosyncrasies of human Because women represent about 1 percent of all behavior found in behavioral finance may provide CEOs, they also represent about 1 percent of all hints of profit opportunities. Finally, prediction earnings announcements. But when I look at the markets offer two particularly interesting uses to largest positive earnings surprises, women are dra investment professionals: (1) providing market matically overrepresented, and when I look at the aggregated forecasts of political risk as an input to largest negative earnings surprises, womenare dra the investment process and (2) providing a tool for matically underrepresented. This finding implies to information aggregation within investment firms methatmarketparticipants systematicallyunderes themselves. Most companies have an information timate the value of companies directed by women. aggregation mechanism that includes neither Notonlythat,butfurther analysis ofthe data shows accountability nor a reward system. It is called a that female CEOs have a tremendous propensity to "staff meeting." A prediction market within a com surprise male analysts. pany could be used to aggregate information from staff in a far more efficient manner.

Conclusion Editor's Note: Justin Wolters proVides consulting services to sev A few things need to be kept in mind when dealing eral prediction markets. with prediction markets. First, if one assumes that the efficient market hypothesis is correct, market This article qualifies for 0.5 CE credits.

REFERENCES

Giirkaynak, Refet, and Justin Wolfers. 2006. "Macroeconomic Election." Journal ofPublic Economics, vol. 90, no. 4-5 (May): Derivatives: An Initial Analysis ofMarket-Based Macro Forecasts, 751-773. Uncertainty, and Risk." NBER Working Paper 11929 Ganuary). Rhode, Paul, and Koleman Strumpf. 2004. "HistoricalPresiden Knight, Brian. 2006. "Are Policy Platforms Capitalized into tial Betting Markets." Journal of Economic Perspectives, vol. 18, Equity Prices? Evidence from the Bush/Gore 2000 Presidential no. 2 (Spring),127-141.

Question: Are prediction mar Wolters: Because I am an econ prediction market, a play money kets subject to manipulation? omist,Iwouldlose myunioncard market tends to be more efficient if I did not say that real money and thus may approach the effi Wolters: Many attempts have alwaysmakesa difference. Oneof ciency of a real money market. been made to manipulate predic the important things about pre tion market prices. A good exam Question: Why are some of diction markets is that they ask ple occurred when Pat Buchanan the prediction markets used for participantsto take positionswith was running for the Republican economic analysis no longer their own money. presidential nomination. He available? HaVingsaidthat,lancespoke encouragedhis supporters to buy with someone who had created a Wolters: Which markets sur Pat Buchanan stock. Their behav prediction market company that viveandwhichdonotis a function ior, however, did not change the uses play money, and he bet me of which markets are profitable to reality thatPatBuchananhadlittle thathis play moneymarketcould run for those who runthe market. chance of winning the nomina do as wellas a realmoneymarket. For example, Goldman Sachs and tion. Therefore, if someone says I So, we set up a competition in Deutsche Bank were charging am willing to buy stock in Pat which his market predicted the what I considered unbelievably Buchanan, someoneinthemarket winner of Saturday football high fees, which proved to be a willsell. Buttheverynatureoffree games while I used Intrade, a real disincentive to trade.So, itwasnot entry and the supply response money market, to predict win a profitable business for them to meant that Pat Buchanan's sup ners. By the end of the season, the stay in. Nevertheless, if such a porters could not affect the price two markets had amassed compa marketexisted, manyofus would for more than a COli-pIe of hours. rable prediction records. In that find it valuable. So, manipulation attempts instance, therefore,moneydidnot Many governments around sometimes occur, but I am aware make a big difference. the world are thinking about run of no attempt that has succeeded I can offer two explanations ning such markets within their for a sustainedperiod. Therewere for this result. First, prediction bureaucracies so that, as I men manipulation attempts again in markets-whether using play tioned earlier, their economists the2008 presidentialrnarkets,but will have a more efficient aggre these too ultimately were unsuc money or real money-do some gation tool than staff meetings. I cessful, as prices reverted to fun thing useful that polls do not do. foresee such markets being used damentals over the course of a Prediction markets ask the right question. Theydonotask, as polls within organizations more often couple of weeks. do, "Whowouldyoulike to win?" than between organizations. Question: Can widespread use or"Who do youplanto votefor?" Question: How far in advance of prediction markets influence Prediction markets ask, "Who do do prediction markets work? subsequent behavior in the sense you think will win?" Second, of persistence momentum? there is no way to get rich in play Wolters: Let me make two moneypredictionmarkets. There observations. First, the closer we Wolters: At this point in the fore, therearenobigplayers influ come to any event, the easier it politicalcycle, predictionmarkets encing suchmarkets,whichis not is to predictthe outcome. Second, arestilla fringe phenomenon.The the case in real money markets. at any given horizon, prediction number of words devoted by the For example, Warren Buffett markets seem to do better than press to prediction markets com has a big influence on equity the alternative. paredwiththose devoted to polls u.s. markets. That is probably a good is sufficiently small that we prob Question: What impact do thing, and it makes the market ably have little need to worry tail events have onprediction more efficient. Butthere may also about persistence momentum. markets? be large pools of uninformed Question: Is there a difference money in real money markets, Wolters: Every time a tail in the behavior ofprediction mar which may make the real market event happens, I get calls asking kets thatuse real moneyandthose lessefficient.Withoutunin whether the prediction markets that use play money? formed money in a play money are wrong. This happened after

©2009 CFA Institute • cfapubs.arg JUNE 2009 • 43 CFA Institute Conference Proceedings Quarterly the 2008 New Hampshire pri the difference between a 5 per Wolters: Remember how well 1 mary. The New Hampshire cent chance, a 2 percent chance, play money prediction markets primary markets said there was and a 1 percent chance. do with no money? Then you I only a 7 percent chance Clinton should remember, too, that high would win; Clinton won. When Question: Whatare some ofthe better prediction markets? volume times zero price equals such things happen, we tend to zero dollar volume. Yet, play forget a truism about 7 percent Wolters: Letmesayupfrontthat money markets do quite well probabilities, which is that they I have a conflict of interest here despite their zero dollar volume. will happen 7 times out of 100. because I advise many of these In fact, my observation-and lf they happen less than that or companies. That said, Intrade has this is just an intuitive answer, more than that, then the markets a greatdeal ofgooddataandoffers are in trouble. not a scientific one--is that pre perhaps the best U.S. political cov diction markets do surprisingly But also keep in mind that erage. Anotherexcellentoneis Bet prediction markets are not well well with only a small number fair, although ithas more British of participants. calibrated over very small prob than U.S. political coverage. The I have seen within corpora abilities. When markets are pre Iowa Electronic Markets is pre dicting between a 10 and 90 tions perhaps a dozen people dominantly political, so it has a percentchance ofan eventoccur trade in a market and do quite narrower range than the others. ring, I take that as a fairly good well. And remember, the right assessment of probability. Nev Question: Howmuchvolumeis metric is the accuracy of the fore ertheless, prediction markets required for the prediction mar cast. Is it more accurate than you seem to be fairly bad at telling kets to be seen as accurate? would have had otherwise?