The above picture includes prediction market trading from the PredictIt prediction market as of July 18, 2023.
Prediction markets are virtual markets designed and conducted to forecast future events. They differ from typical markets in that their purpose is to provide forecasts rather than a means with which to buy and sell products such as stocks.
The way prediction markets work is that people buy and sell shares based on how they think a candidate will do in an election or how a movie or TV show would do in the marketplace. The buying and selling activity generate market prices that are highly predictive of what will happen.
advertisement
advertisement
Prediction markets rely on a research methodology based on a theory called the wisdom of crowds. This is different from polls and sampling methods that rely on probability theory.
Whereas probability theory says that if you randomly choose a subset of people or sampling units, the information you glean from them will be representative of a larger population.
The wisdom of crowds, on the other hand, says that if you total people's insights, the combined wisdom of the crowd will render forecasts that are close to actual outcomes.
Although not widely used, the wisdom of crowds has been around a long time.
According to a book written by James Surowiecki entitled The Wisdom of Crowds, the theory was developed in 1906 by an 85-year-old British scientist by the name of Francis Galton.
Galton was at a regional fair where local farmers and townspeople gathered to appraise the quality of farm animals. This included placing wagers on the weight of an ox. The best guesses received prizes.
After the contest was over, Galton asked for the tickets and ran statistical tests on the answers. There were 787 in total. However, he had to discard thirteen because they were illegible.
One of the tests was to calculate the mean of the contestants' guesses. The average Galton calculated was 1,197 pounds. The actual number was 1,198 pounds. To Galton's great surprise, the wisdom of the crowd was highly accurate.
Galton's intent was to prove that people's breeding would provide the best answers. What he learned was that democratic judgement was a lot better than expected.
Examples of modern-day uses of wisdom of crowds include the Iowa Electronic Market (IEM), the Hollywood Stock Exchange (HSX), PredictIt and Media Predict. These are all what are known as prediction markets.
IEM and PredictIt predict outcomes of political elections, HSX predicts movie box-office results and Media Predict was created to forecast audience sizes for TV show concepts.Most prediction markets use play money. However, IEM is a real money market and Media Predict gives participants money to use so they cannot lose their own.
The fact that there is money on the line that respondents can lose makes people more truthful in their assessments. In a poll or survey, there is not the same incentive for people to be honest.
According to a University of California Berkley Haas Business School professor by the name of Don Moore, most election polls report a 95% confidence level.
Yet an analysis of 1,400 polls from 11 election cycles found that the outcome lands within the poll's result just 60% of the time. And that's for polls just one week before an election — accuracy drops even more further out.
The confidence level is only a gauge of sample error. It doesn't account for all the other types of error that can creep in while conducting polls and surveys.
In 2007, three researchers from the Henry B. Tippie College of Business at the University of Iowa conducted a study called Prediction Market Accuracy in the Long Run to determine whether the IEM prediction market outperforms polls a long way in advance -- not just on the eve of an election.
They aggregated over 965 polls over five Presidential elections and the IEM market was closer to the eventual outcome 74% of the time. Furthermore, the market significantly outperformed the polls in every election when forecasting more than 100 days in advance.
There is also a lot of academic research done for HSX. In a study about HSX prediction accuracy conducted by Benjamin Olsho at Pennsylvania State University, HSX had a Pearson Correlation Coefficient of 0.96 and a mean absolute percentage error (MAPE) of 23.
The Pearson Correlation Coefficient measures the strength of the relationship between two variables, in this case the HSX prediction and the actual box-office revenue. A 1 is strongest and a 0 is least strong. A 0.96 score is very high.
The MAPE means that on average, the HSX prediction the prior day before the opening weekend of a film was within plus or minus 23 percent of the actual opening weekend box office figure.
It's more difficult for any prediction methodology to forecast a precise number like box-office revenue or average audience size from more than 100 days in advance. HSX's MAPE increases to 44% at 80 days before release.
When I was working there, the Media Predict prediction market produced a white paper based on its first year of operation. At that point, it had generated 53 predictions for shows that averaged over 1M viewers. For eve-of-premiere predictions, MP averaged 85% predictive accuracy (+/- 15% of actual figures). Long-range accuracy was similarly high even as far out as 70 days.>
From experience I can tell you that prediction markets are not a crystal ball. Perfect prediction is not available to us mere mortals.
They are, however, a powerful prediction tool for people looking to make decisions about donations to political campaigns and for investments related to film and TV show production.
Ignore them at your own risk.
Very interesting, Ed. But I'm not sure that I buy the idea that random probability surveys, when properly executed, give you a less reliable reading than a "crowd" survey as, in theory, the random probability method is meant to represent the total population's views or activities.
The basic reason for random probability research is cost. You can't afford to do anything approaching census level research for most projects---so as a matter of affordability, much smaller but hopefully representative samples are employed. If, instead, you go "the crowd" route does this mean that you get your information from a much larger but readily available sample of folks who may or may not represent a true crossection of the population. In other words, is a large sample---even if it's membership may not represent all segments of the population better than a much smaller sample which is designed to do just that?
Generally speaking, most of the random probability surveys have been proven reasonably accurate over the years---meaning that while they are off by a few percentage points or even less from the truth, they come close enough to provide useful data relative to their cost. The political polls are frequently cited as exceptions to this but when their tiny samples are accounted for---allowing a variability of several points in the findings---they are usually predictive when used in a sensible manner. Inother words a fairly conducted poll of 1000 likely voters which has POTUS candidate A beating B by two percentage points is usually cited as an example of a hopelessly wrong poll, implying that a sample of 50,000---or a "crowd" survey----would have done better. Maybe so, maybe not, but as the parties involved usually can't afford "crowd" surveys we seem stuck with smaller sample research. The question being whether the design is fair and the study is properly executed----wherein lie most of the problems.
Hi Ed -- First off, I want to say that you're very smart and everyone benefits from your commentary.
Regarding your feedback here, you have to remember that prediction markets aren't based on probability theory. They're based on the theory of the wisdom of crowds. As a result, sample size and representation aren't necessarily key factors when it comes to harvesting good predictions.
One of the really interesting findings from my time at Media Predict was that woman were just as good at predicting the success of male-oriented shows as men were at predicting the success of female-oriented shows. The demographics didn't matter. Moreover, the people who were bad at making predictions would predict poorly in both directions and cancel each other out.
If you were going to use a prediction market for research, you don't have to pull a sample. You would use the existing panel and pay to have a question added to the market. It doesn't have to be more expensive to use a crowd methodology.
This op-ed isn't to bash polls or surveys. I still look at them. However, when available, I look at prediction market data as well.
Thanks, Ed, appreciate your comment.
About the wisdom of"crowds" though---if I understand you---it doesn't necessarily matter what kinds of people you get opinions/predictins from so long as there are lots of them----hence your example of men correctly evaluating something of primary interest to women or the reverse. While I, too have seen examples such as you describe and there are plenty of built-in biases in many "probability sample" studies---especially about how the questions are posed---I remain somewhat skeptical about the "crowd" appraoch---though that doesn't mean that it wont work---especially when dealing with personally relevant or highly emotional subjects like politics.