Guess the outcome of June 14's votes!

How heavy is the cow grazing in your garden? Your guess may be very far from reality, but if you manage to combine the estimates of a large number of people you will approach the real value. When Francis Galton performed this experiment in 1907, he demonstrated the relevance of what was later dubbed as "the wisdom of crowds." Basically, to obtain a reliable prediction it is better to collect the opinion of a large number of people rather than reading tea leaves or checking the position of the planets!

A team at EPFL’s Artificial Intelligence Laboratory repeated Galton’s experiment during the most recent Innovation and Technology Exhibition (STIL). The quantity of M & M’s in a jar -easier to transport than cattle- was submitted to the visitors’ discernment." By doing a simple harmonic mean of the 180 estimates collected, we obtained a result with 97% accuracy on the number of such treats, reaching even 98.9% precision when using a geometric mean," reported the postdoc student Florent Garcin, in charge of the project, now developed within the Umanytics start-up.

A full-scale test is being performed this week: we invite you to give your opinion about the outcome of June 14’s votes. You can participate even if you do not have the right to vote in Switzerland. The more participants, the more accurate the prediction will be!

What algorithm to use?
The choice of the mathematical model to combine all the different opinions is crucial for Swissnoise.ch. "People who enroll are divided into two groups. Some bet using a common mathematical model derived from research conducted at Harvard and at George Mason University. The others use an algorithm developed by EPFL. This allows us to compare the two theoretical models."

The first group of users has the option of betting π, a virtual currency to buy "shares" corresponding to a specific answer. "As with a stock system, some actions go up in value depending on the popularity of their response. At the end, users earn points based on the actual result of the event on which they were betting, "says Florent Garcin." In the case of EPFL’s model, each opinion can change its subsequent compensation. A minority opinion will be paid a bit more, but not too much, because we must encourage users to indicate what they really think and not push them to take risks with excessive payments."

Cheaper and more flexible surveys
One of the main advantages of EPFL’s model is that it does not need to be compared to a real result. "For example, a car manufacturer might contact us to find out what prototype to build because, unlike his competitors, we do not need to compare our results with actual sales figures at the end of the study. “But it is important to formulate the questions in a neutral way to get relevant results. "Even on political subjects we do not ask people what they are going to vote for, because on the one hand we cannot be sure that our user group will be representative of the population and secondly such questions are easily biased. We always ask users about the results they expect. For example, one can think more to the left, but expect the right to win. "

A new tool for the media
The research carried out in the Artificial Intelligence Laboratory by Florent Garcin and his students Jérémy Gotteland, Jonathan Link and Michael Hobbs is already attracting media interest. "Conventional opinion polls are very expensive and only catch a picture of a given time. Conversely, our model could, for instance, trace the trends throughout a political campaign for a much lower cost", suggests Florent Garcin.