Are prediction markets more accurate than polls?

Sometimes, but the honest summary is that the record is mixed. Some studies found prediction markets beating polls, others found them matched or beaten, and the answer depends on the question, the timing and the comparison.

The best-known win for markets comes from US presidential elections. A study of the five contests from 1988 to 2004, by Berg and colleagues, found that prediction markets gave a more accurate estimate of the voting result than 74 percent of the opinion polls studied. Yet Erikson and Wlezien (2008) found that "damped" polls, adjusted before being turned into a forecast, outperformed all markets or models. But a randomized experiment from 2016 found prediction markets 12 percent less accurate than prediction polls, another way of combining a crowd's probability judgments. And polls weighted by the confidence people expressed in their forecast performed as well as market-based forecasts.

Markets have also had famous misses. They leaned heavily toward the UK staying in the EU in 2016 and failed to predict the outcome, and they also failed to predict the winner of the 2016 US presidential election. One explanation offered is that traders treated the market odds as correct probabilities and did not update enough on outside information.

Timing matters too. Predictions are better for events close in time, and for elections more than a year away, prices are biased toward 50 percent. Polls have their own weakness: early in a campaign they are poor measures of voters' future choices. The sensible reading is that markets are another useful input, not an oracle.