Cantor Loss shows Crowdsourcing, not Polling, is the Future of Prediction
Eric Cantor, the 2nd most powerful person in the House of Representatives, lost in the Republican Primary today to the relatively unknown Dave Brat. While others have focused on the historical nature of the loss, given Cantor’s position in his party, or on the political ramifications, I was most intrigued by the fact that polls conducted recently predicted Cantor would win by 34 points or 12 points. In the end, Cantor lost by more than 10 points.
How did the polls get it so wrong? In an age where people are used to blocking out web ads, fast forwarding through commercials, and screening their calls, using automated phone technology to ask people who they will vote for and assuming that you’ll get an unbiased sample (e.g. people who answer such polls don’t differ from those who do not answer automated polls) seems unwise. The first banner ad got 44% clickthrough rates, but now banner ads are only clicked on by a small minority of internet users. As response rates fall, bias is inevitable.
Pollsters may try to weight their polls and use new techniques to produce more perfect polls, but non-response bias will only get worse as consumers learn to block out more and more solicitations using technology. On the other hand, a good crowdsourcing algorithm, such as the algorithm we use to produce Ranker lists, does not require the absence of bias. Rather, such an algorithm will combine multiple sources of information, with the goal being to find sources of uncorrelated error. In this case, polling data could have been combined with the GOP convention straw poll, the loss of one of his lieutenants in an earlier election, and the lack of support from Republican thought leaders, to form a better picture of the election possibilities as the non-response bias in regular polling is a different kind of bias than these other measurements likely have, and so aggregating these methods should produce a better answer.
This is easy to say in hindsight and it is doubtful that any crowdsourcing technique could have predicted Cantor’s loss, given the available data. But more and more data is being produced and more and more bias is being introduced into traditional polling, such that this won’t always be the case, and I would predict that we will increasingly see less accurate polls and more accurate use of alternative methods to predict the future. The arc of history is bending toward a world where intelligently combining lots of imperfect non-polling measurements are likely to yield a better answer about the future than one attempt to find the perfect poll.
- Ravi Iyer