There are hundreds of academics studying prediction markets around the world, and thousands of academic papers have been published. The compilation below is far from being exhaustive and is presented for informational and educational purposes only.
Rules limiting insider trading may encourage investment, but they may also discourage exploration of new less-decentralized corporate information processes, such as prediction and decision markets. I review standard corporate information processes and insider trading rules, outline possible improvements that prediction markets might offer, and consider ways we might change insider trading rules to allow both more flexible innovation of information processes and better-encouraged investment.
This article compares the forecast accuracy of different methods, namely pre- diction markets, tipsters and betting odds, and assesses the ability of prediction markets and tipsters to generate profits systematically in a betting market. We present the results of an empirical study that uses data from 678–837 games of three seasons of the German premier soccer league. Prediction markets and betting odds perform equally well in terms of forecasting accuracy, but both methods strongly outperform tipsters. A weighting-based combination of the forecasts of these methods leads to a slightly higher forecast accuracy, whereas a rule-based combination improves forecast accuracy substantially. However, none of the forecasts leads to systematic monetary gains in betting markets because of the high fees (25%) charged by the state-owned bookmaker in Germany. Lower fees (e.g., approximately 12% or 0%) would provide systematic profits if punters exploited the information from prediction markets and bet only on a selected number of games.
Daniel E. O’Leary
Internal prediction markets draw on the wisdom of crowds, gathering knowledge from a broad range of information sources and embedding that knowledge in the stock price. This chapter examines the use of internal prediction markets as a forecasting tool, including as a stand-alone, and as a supplement to forecasting tools. In addition, this chapter examines internal prediction market applications used in real-world settings and issues associated with the accuracy of internal prediction markets.
Paul W. Rhode (University of Michigan)
Koleman Strumpf (University of Kansas School of Business)
Political future markets, in which investors bet on election outcomes, are often thought a recent invention. Such markets in fact have a long history in many Western countries. This paper traces the operation of political futures markets back to 16th Century Italy, 18th Century Britain and Ireland, 19th Century Canada, and 20th Century Australia and Singapore. In the United States, election betting was a common part of political campaigns in the pre-1860 period, but became increasingly concentrated in the organized futures markets in New York City over the post-1860 period.
Despite the popularity of prediction, markets among economists, businesses, and policymakers have been slow to adopt them in decision-making. Most studies of prediction markets outside the lab are from public markets with large trading populations. Corporate prediction markets face additional issues, such as thinness, weak incentives, limited entry, and the potential for traders with biases or ulterior motives — raising questions about how well these markets will perform. We examine data from prediction markets run by Google, Ford Motor Company, and an anonymous basic materials conglomerate (Firm X). Despite theoretically adverse conditions, we find these markets are relatively efficient, and improve upon the forecasts of experts at all three firms by as much as a 25% reduction in mean-squared error. The most notable inefficiency is an optimism bias in the markets at Google. The inefficiencies that do exist generally become smaller over time. More experienced traders and those with higher past performance trade against the identified inefficiencies, suggesting that the markets' efficiency improves because traders gain experience and less skilled traders exit the market.
2010 Information Systems: A Manager’s Guide to Harnessing Technology
Many Web 2.0 efforts allow firms to tap the wisdom of crowds, identifying collective intelligence. Prediction markets tap crowd opinion with results that are often more accurate than the most accurate expert forecasts and estimates. Prediction markets are most accurate when tapping the wisdom of a diverse and variously skilled and experienced group, and are least accurate when participants are highly similar.
Leighton Vaughan Williams
In summary, the papers presented in this issue demonstrate that there are several factors, such as the natural risk aversion of well-informed traders and the trading behaviour observed immediately after the release of new information that can induce biases in prediction market forecasts. Equally, that certain predicted classes (such as extreme predictions) can be under- priced or over-priced. In addition, the papers demonstrate that the accuracy and precision of forecasts derived from prediction markets can be affected by a range of factors, including environmental factors, the size of the market, the set-up of the market, the external incentives provided and when in the market cycle predictions are observed.
However, evidence is presented that some aspects of prediction markets, despite a priori expectations, do not influence their accuracy, such as features of the event and the
composition of the participants. In addition, provided markets are given sufficient time, accuracy of prediction market forecasts increases and forecasts derived from these markets outperform, on key metrics, other forecasting mechanisms such as opinion polls and financial markets.
Consequently, the evidence presented here should provide comfort to the potential users of prediction market forecasts. Taken together, the papers suggest that by thinking carefully about the design of prediction markets, by setting up procedures to adjust forecasts to account for known biases and by planning carefully when forecasts from prediction markets are used, they can be a powerful source of forecasting ability.
These findings reflect a notable increase in interest in the applications to which prediction markets can be put, as well as the broader body of evidence suggesting that these markets might produce better forecasts than alternative forecasting mechanisms. The insights gained have also been shown to have potentially valuable applications for policy, not least when accurate forecasts are required in relation to quantifiable targets. The findings also reflect a wider interest in the best way to design and implement prediction markets.
Even so, some have questioned how far prediction markets can outperform other methods of forecasting, or whether they are at best a supplement to more traditional forecasting methodologies. These doubts have attracted added focus in very recent years following some high-profile forecasting failures in the context of major political event outcomes. In particular, the 2016 EU referendum in the UK and the 2016 US presidential election produced results that were a surprise not only to the great majority of pollsters abut also to followers of the prediction markets. There are various theories to explain why the markets failed in these big votes. One theory looks to the basic laws of probability. An 80 per cent favourite can be expected to lose one time in five, if the odds are correct. In the long run, according to this explanation, things should even out. A second theory to explain the surprise results is that something fundamental has changed in the way that information contained in political prediction markets is perceived and processed. One interpretation is that the hitherto widespread success of the markets in forecasting election outcomes, and the publicity that was given to this, turned them into something of an accepted measure of the state of a race, creating a perception which was difficult to shift in response to new information. This is a form of ‘anchoring’. Linked to this is a herding hypothesis, which is that because the prediction markets had by 2016 become so firmly entrenched in conventional wisdom as an accurate forecasting tool, people herded around the forecasts, propelling the implied probabilities of favoured outcomes upwards. A third theory is that conventional patterns of voting broke down in 2016, primarily due to unprecedented differential voter turnout patterns across key demographics, which were not correctly modelled in most of the polling and which were missed by political pundits, political scientists, politicians and those trading the prediction markets. Finally, there has been widespread discussion of the impact of manipulation, not only of the markets themselves but also of the distribution of information and misinformation.
Future research into the uses and applications of prediction markets might focus on what can potentially be learned from recent forecasting failures to improve the efficiency of forecasts derived from prediction markets, as well as to identify when markets are likely to be most informative. The ways in which market forecasts can be adjusted to allow for systematic
biases is also likely to attract the continued interest of researchers. Besides comparing the efficiency of prediction markets with other forecasting methodologies, there is in addition a growing interest in the idea of combining forecasts derived from the different methodologies to generate improved forecasts. Increasing attention is also being paid to the design and incentive structure of markets and the contribution they can make as a corporate and policy decision support tool.
The papers selected for this special issue of the International Journal of Forecasting reflect this diverse research agenda, and offer an exceptionally strong contribution to the existing literature on prediction markets. Importantly, they provide a valuable framework upon which future work can build. It is with no little anticipation that those interested in theory, evidence and applications relating to prediction markets look forward with great anticipation to the outputs of further research in this area.
Joyce E. Berg
Thomas A. Rietz
Berg and Rietz analyse the forecasting efficiency of binary prediction markets. Such markets, which predict probabilities of discrete outcomes, have become quite popular. However, researchers generally cite measures of forecasting efficiency based on linear prediction markets. Linear markets forecast the level of an outcome. To measure efficiency, forecasts can be compared to actual outcomes directly. This differs from binary markets where one must compare the forecast outcome probabilities to actual outcome frequencies under repeated, essentially identical, conditions. The authors use a unique set of repeated, binary prediction markets to study efficiency using frequency and logistic analysis. While they document a pricing bias at intermediate horizons, the bias disappears before the forecasted events occur. The observed bias conflicts with static models of bias based on the longshot bias, overweighting of low probability events, prospect theory and other models that predict over-pricing of low-probability events. The bias observed is transitory, affects only very high and low prices, and follows a pattern predicted by the information-based, over-reaction story of Daniel, Hirshleifer and Subrahmanyam (1998). Early in a market, when there is little information, prices appear unbiased. As traders observe information, prices respond, leading to transitory overpricing of high probability events (and underpricing of low probability events). As the event approaches and more information is reflected in prices, this bias fades and, ultimately, disappears. The paper makes two observations that are important in interpreting binary market prices: (1) Prices are forecast probabilities and cannot be compared directly to outcomes. (2) Prediction markets may appear overconfident (i.e., over-price very high probability events and under-price very low probability events) at intermediate horizons.
Brown and Yang analyse the role that crowd size plays in the accuracy of prediction/betting market prices. They ask whether large markets with high participation rates provide more accurate forecasts than smaller markets. There are conflicting theoretical predictions in this area. On the one hand a large market populated with noise traders (De Long et al, 1990) may produce biased prices. On the other hand, large markets create the incentives for information acquisition as returns from informed trading are higher. Indeed, information may actually be dispersed amongst the crowd (Galton, 1907) and therefore as long as forecast errors are not correlated a larger crowd will produce more accurate predictions. The authors examine repeated natural experiments in tennis betting, where every other year the Queen's Club tennis tournament clashed with a major soccer tournament. Importantly, these clashing tournaments exogenously reduced the participation rate in tennis betting. They found that larger markets (without the clashing soccer tournament) produced more accurate predictions. This a result of practical importance, as many prediction market designers can set rules or incentives which determine whether the prediction market will be large or small. In this case, bigger is better.
Tai, Lin, Chie, and Tung provide a decision-support framework to help policy or decision makers who count on the forecasts of prediction markets. Regardless of the accuracy of prediction markets in general, sometimes the failures in their predictions will put decision- makers’ stake at risk. Instead of blindly relying on markets’ predictions, therefore, this paper offers a systematic way of assessing the credibility of the predictions. On the basis of previous studies about factors influencing the efficiency of prediction markets, the authors incorporated a list of variables into statistical models as well as machine learning algorithms to explore the underlying (and possibly nonlinear) relationships between these factors and the prediction results. More specifically, four classification models were used to classify the credibility of prediction market forecasts. They then used a combined forecasting technique to integrate the classification results from these four models. By doing so, they are able to provide a judgment about a specific prediction market’s forecast prior to the predicted event. A large dataset involving 650 markets was used to train and test their combined forecasting framework. The results indicate that the method was able to discover models capable of predicting the failures of prediction markets according to different criteria. Most interesting is the flexibility of the proposed framework which allows decision-makers to build their own models with different sets of variables, classification models, and decision objectives.
Oh Kang Kwon
Grant, Johnstone and Kwon treat the contracts in prediction markets as financial assets. Contracts are "priced" from the traders' subjective perspective using utility theory or the finance expression of utility theory known as the capital asset pricing model (CAPM). The contract's price implies its discount rate or required rate of return. By exploring the price- implied discount rates of binary contracts, the authors clarify how a trader's rational required expected return reacts to the trader’s subjective probability of winning. The general finding is that a trader who is more confident of the contract expiring in the money requires a lower expected return on that contract. The expected return required of a typical prediction market contract that pays zero or one is found to increase linearly in its ratio of ex ante perceived subjective payoff variance to payoff mean. A surprising but immediate conclusion is that the natural risk-aversion of well-informed traders induces a favourite-longshot bias in prediction markets, where longshot contracts cost more than they should relative to their low chance of winning and favourites are relatively too cheap. This bias is explained not by errors in traders' probability assessments, but by the posted bids and asks of traders whose probability assessments are accurate but who are risk-averse.
Luis Felipe Costa Sperb
Costa Sperb, Sung, Johnson and Ma investigate the influence of environmental factors on the calibration of probabilistic estimates derived from prediction market prices. The literature suggests that environmental factors, such as weather and atmospheric conditions, can affect the information processing and cognitive ability of decision-makers, leading to sub-optimal estimations about uncertain events. This, in turn, has the potential to impair the effectiveness of prediction markets as a means of appropriately aggregating and weighting relevant information dispersed in the public. In the context of horserace betting markets, the authors found that even after the effects of these conditions on the performance of contestants (horses and jockeys) have been discounted, the accuracy of the probabilities derived from market prices is affected systematically by the prevailing weather and atmospheric conditions. By correcting for this phenomenon, they showed that significantly better forecasts can be derived from prediction market prices, and that these have substantial economic value. Importantly, the results of this paper suggest that when the purpose of a prediction market is to derive accurate probabilistic estimates from final contract prices, forecast accuracy can be improved greatly by identifying and correcting for conditions where prediction markets systematically under-perform.
J. James Reade
Leighton Vaughan Williams
Reade and Vaughan Williams consider the forecast performance of prediction markets alongside that of opinion polls in the context of US election outcomes. The first contribution of this paper is to evaluate a much wider range of prediction markets than previous studies have been able to, most notably considering the large commercial prediction markets, Intrade and Betfair. Prediction market forecasts are probabilistic in terms of which candidate will win, whereas opinion polling output tends to be reported in terms of vote shares. This makes it difficult to compare how close each type of forecast was to providing the most efficient forecast, where efficiency is measured in terms of information incorporated at the time the forecast was made. Although these two variables are related, their correspondence requires a set of assumptions, be they theoretical or statistical. Page (2008) provides theoretical assumptions, and the second contribution of this paper is to use the related statistical assumptions to make an empirical conversion between probabilities and vote shares. As both opinion polls and prediction market forecasts can be corrected for bias, it is more informative to think about the precision of forecasts. On this metric, prediction market forecasts of election outcomes perform better than opinion polls.
Di, Dimitrov, and He consider the role of external incentives on the ability of prediction markets to aggregate information. Most papers on external incentives show that when external incentives exist, it is likely that prediction markets will not always accurately capture all agents’ private information. However, papers that consider external incentives tend to assume agents’ costs external of the prediction market are symmetric. The authors show that even when agents may take actions external to the prediction market that influence the outcome of the traded event, if the actions external of the prediction market are present, prediction markets may indeed capture all agent’s private information. In particular, so long as the desired external action is rewarded more than the undesired action, prediction markets do not incentivise undesirable actions. This insight hopefully addresses concerns some managers may have in using prediction markets in the workplace due to potentially incentivising undesirable actions.
Oliver B. Linton
Auld and Linton identify the night of the UK EU Referendum of 2016 as providing a unique natural experiment to study the degree to which information was discounted in two parallel forms of prediction market: financial and betting markets. In particular, they explore the behaviour of the sterling dollar exchange rate as well as binary contracts listed on Betfair that pay out according to the outcome of the Referendum. The authors argue that for those few hours overnight the sole determinant of prices in these assets was the flow of information provided by the results of the vote. Using public information they construct an ex-ante joint prior distribution for the vote-share of every constituency that announced that night. Using a Bayesian methodology they update this prior as results arrive and compute a dynamic probability that the UK leaves the EU. They find that although both markets are slow in pricing the information contained in the results, by around 2 to 3 hours, the betting market is less inefficient than the financial market. Further, by constructing a theoretical model that links the prices of these contracts, independent of the outcome of the vote, they show that there were violations of weak market efficiency. There were apparently very profitable arbitrage opportunities involving selling the pound and placing hedging bets on Betfair that would pay out whether or not the UK voted to leave the EU.
Lohrmann and Luukka link stock markets to prediction markets and analyse the prediction of S&P500 intraday returns. In contrast to much of the existing literature, this is set up as a four- class problem and not as a regression or binary classification problem. Premised on these four return classes, four trading strategies are tested against a simple buy-and-hold strategy. All the suggested trading strategies conduct buy- and sell-decisions only based on certain predicted classes. The research indicates that the four classes differ in their contribution to the return on the strategies. In particular, the two extreme classes with ‘Strong Positive’ and ‘Strong Negative’ predicted returns lead overall to higher mean returns than the ‘Slightly Positive’ and ‘Slightly Negative’ return classes. This result holds true even when misclassified returns are included in the averages. Therefore, using strategies that act only on a subset of the predicted classes (such as the extreme predictions) can work to generate higher profits than following all the predictions.
Strijbis and Arnesen go beyond the current focus of the prediction market literature by combining observational and experimental analyses of prediction market errors. They investigate the prediction error of a real money prediction market using a logarithmic market scoring rule for 65 direct democratic votes in Switzerland. The authors distinguish between prediction market error due to the set-up of the market, features of the event to be predicted, and the participants involved. They find that the prediction market accuracy varies primarily according to the set-up of the market, while the features of the event and especially the composition of the participant sample hardly matter. Hence, those applying prediction markets should consider carefully the specific configuration of their market while they can remain more relaxed about the composition of their sample of traders.
J. James Reade
Leighton Vaughan Williams
Brown, Reade and Vaughan Williams consider the informativeness of prediction market prices. In the aftermath of information events, how do market prices react? This naturally matters, as prediction markets are extensively used to elicit forecasts about uncertain future events. The paper uses the Intrade exchange and the release of opinion polling data. They investigate, in particular, what happens in predictions markets in the immediate aftermath of opinion polling releases. Such releases are significant news events, and in the case of Gallup polls, occur on a regular schedule, at 1pm Eastern Time during most days during the 2012 presidential campaigns in the US. The authors find that in the immediate aftermath of a poll release, there is an increase in trading activity. However, much of this activity involves relatively inexperienced traders, and as a result market efficiency declines in those moments. Once more experienced traders return in the following hours, price efficiency recovers. These findings are of practical importance, as they give a sense of the extent to which prediction market prices might be relied upon for accurate forecasts in the aftermath of significant news events.